Narrative-first, evidence-driven

Back to case studies [pdf]

Receipts: tracked in Methods and Sources by type: Primary documents | Official data | Independent analysis

• At a glance
• 1. Five scenes
• 2. What’s happening
• 3. Why it’s happening — the mechanisms
• 4. Five-scene comparison
• 5. Where it broke — evidence against the good story
• 6. Where it held — evidence the window is real
• 7. Human Command test and Exit test
• 8. Shared Gains test
• 9. Minimum floor
• 10. What to do
• Loop Effect
• North Star verdict
• Research gaps

At a glance

• What changed: AI tools deflated the cost of building software products and knowledge-work deliverables — consulting analyses, equity research, legal documents, creative campaigns, compliance reviews. Inference costs fell 99.7% from 2023 to 2025. Tasks that required full teams can now be completed by small teams or individuals at comparable speed. [confirmed for speed; quality parity varies by domain and task type]
• Where power moved: Not to the new entrant. To whoever controls the channel between building and getting paid: enterprise procurement bundles, cloud platform defaults, regulatory certification floors, client trust relationships, liability coverage requirements, and professional credentials.
• The bottleneck shift: More competitors reach the gate. The gate has not gotten easier to pass. Production cost deflation without distribution contestability means more competition at the build layer and no more competition at the market layer.
• Two gate types: In software and enterprise tech, the gates are structural and one-time — procurement bundling, certification costs, platform API dependency, VC exit pressure. In professional services — consulting, financial analysis, legal, and agency work — the gates are relational: trust-based distribution, malpractice coverage requirements, proprietary data infrastructure, and regulatory credentials. Both types resist compression.
• Who gets squeezed: Displaced workers whose new ventures reach the market but can’t survive the gate period. Junior workers whose training pipeline is narrowing as the tasks they would have learned on are automated away. Customers who briefly see more choice before consolidation narrows it.
• What the minimum floor asks for: Procurement bundling enforcement, FedRAMP reform on schedule, documented alternatives evaluation as a buyer discipline, insurance access for small professional services firms, published AI governance standards in regulated services, and antitrust scrutiny of acqui-hire consolidation.
• The window of competition is real across all five scenes described here. It is also fragile. The question is whether buyers and policy can keep it open long enough to matter.

1. Five scenes

The following scenarios are illustrative composites built from documented patterns in this case. They are not sourced events.

Scene 1: Dev/software

A software engineer at a mid-size SaaS company gets laid off in January. The company cited “AI-driven restructuring.” She has fifteen years of experience. She also has access to Cursor, Claude, and GitHub Copilot.

Within three months, she has a working product. It took her alone what used to take a team of six and eight months. The unit economics look reasonable. Two enterprise prospects are interested.

Then the procurement process starts. Prospect A renewed their Salesforce contract last quarter. Salesforce just released a native AI feature that covers most of what her tool does. It is already in the contract, already approved, already integrated. The product champion loves her demo but cannot get the decision in front of anyone who can act on it. Prospect B is a federal agency subcontractor. They need FedRAMP Moderate authorization. She gets the quote: $500,000 to $1,000,000 and twelve to eighteen months of compliance work. She has neither.

She has built a better product than what the incumbents are shipping. She cannot get it in front of the customers who would pay for it. The wall to build got lower. The wall to sell did not move.

Scene 2: Agencies / creative shops

A creative director with twelve years at a mid-size agency gets laid off when the agency loses two anchor accounts and cuts headcount by a third. Within six months, he has a two-person shop. With AI tools he can produce brand strategy, campaign concepts, and production-quality design work that would have required a team of eight a few years ago. His work is good. He has three client references from his agency years who will take his calls.

The references go well. Then the clients explain their process. One is required to run any new vendor through a procurement review that takes four months and requires insurance coverage at a level that costs more than his first year of projected revenue. Another is happy with the work but reluctant to move the account — their CMO has worked with the incumbent agency for nine years and does not want to manage the transition. The third makes the switch.

One in three. Not because the work wasn’t there. Because the evaluation never really happened for the other two.

Scene 3: Management consulting

A management consultant with a decade at a Big 4 firm leaves during a restructuring round and starts a boutique with two colleagues. They use AI tools to compress the research, analysis, and deck-production work that used to require a full engagement team. Their first three proposals are competitive on price, faster on delivery, and draw on the same frameworks their former employer charges ten times as much to apply.

They win one of the three. The other two go back to the incumbent. The stated reasons: “We need the liability coverage.” “Our board expects a recognized name.” “The procurement committee wasn’t comfortable with a new vendor for a project this size.”

None of those objections are about the quality of the work. The wall to build fell. The wall to be trusted did not.

Scene 4: Financial analysis

An equity research analyst at a mid-size sell-side firm gets a notice that her role is being restructured. The firm is consolidating its research platform. Within a year, she builds an independent research product: AI-assisted earnings analysis, sector coverage, and client-ready reports. Her work is timely and she can cover more companies at lower cost than her former employer.

The problem is distribution. Her former firm’s research reached clients through a Bloomberg Terminal connection, institutional sales relationships, and FINRA-registered publishing infrastructure. To publish regulated research under her own name, she needs to pass the Series 86 and 87 exams and register under FINRA. Institutional clients with compliance mandates will not subscribe to unregistered research. The clients who valued her judgment are now inside firms that require registered counterparties.

The AI compressed her production cost. It did not compress the registration requirement or the institutional distribution infrastructure. She can produce the analysis. The clients who would pay for it cannot access it without a structure she doesn’t yet have.

Scene 5: Legal and compliance

A commercial litigation associate leaves a large firm after a round of reduced summer hiring and slower promotion timelines. She starts a boutique with a focus on contract review and regulatory compliance for mid-market clients. AI platforms cut her document review time dramatically. Her turnaround on complex contract analysis is faster than the firms she left. On the work itself, she is competitive.

Her first potential corporate client runs an RFP through procurement. The procurement checklist includes: minimum professional liability coverage of $5 million per claim, named insured status, and minimum firm revenue thresholds. Her malpractice coverage — standard for a solo practice — is $1 million per claim. The gap is not about her legal judgment. It is a procurement form that was designed around firms with fifty lawyers, not one.

The wall to produce fell. The wall to be insured at the required level did not.

2. What’s happening

AI tools have made it materially cheaper and faster to build software products and knowledge-work deliverables — consulting analyses, equity research, legal documents, creative campaigns, compliance reviews. That is real. Displaced workers and small teams are reaching the market faster and better-tooled than any prior generation of entrants.

But production cost was never the primary gating factor in enterprise software or professional services markets. The gates vary by market:

Software and enterprise tech:

1. Procurement channel access — who is already in the contract
2. Regulatory certification floors — who has already paid the compliance cost
3. Platform API dependency — whose inference costs make unit economics work

Professional services (consulting, finance, legal, agency):

1. Trust-based distribution — client relationships, institutional credibility, and the ability to have a named decision-maker vouch for the work
2. Liability coverage — malpractice insurance, indemnification structures, and the institutional scale to absorb error
3. Proprietary data — real-time market feeds, client benchmark databases, decades of engagement IP
4. Regulatory relationships — FINRA registration, bar membership, UPL enforcement, and procurement qualification floors

These gates existed before AI. AI did not install them. What AI changed is the ratio: the production gap between a large incumbent and a small entrant has narrowed significantly, but the gate gap has not.

New entrants who clear the production wall hit these gates at speed. In software, most do not clear them without VC subsidy — which introduces exit pressure that usually ends in acquisition. In professional services, the trust cycle accumulates through engagements, referrals, and demonstrated track record, and cannot be compressed quickly regardless of capital. More people can now reach the gate at production-quality. The gate is processing the same volume with the same criteria.

3. Why it’s happening — the mechanisms

Mechanism: Production cost deflation (real, but the ceiling shifted for everyone)

What it is: AI tools lower the cost of building software and producing knowledge-work deliverables. They do not lower the cost of surviving the gate period, or the cost of being trusted. And they lower costs for incumbents too — not just new entrants.

What is confirmed:

• Inference cost for GPT-4-class performance fell 99.7% from March 2023 to August 2025. [confirmed, Epoch AI / NavyaAI]
• The deflation rate accelerated post-January 2024, reaching ~200x per year. [confirmed, Epoch AI]
• Enterprise AI cloud spend grew 3x (2024-2025) despite price collapse, because cheaper inference drives volume consumption upward. [confirmed, CloudZero]
• McKinsey’s internal AI platform Lilli is used by over 75% of the firm’s 43,000 employees, averages 17 uses per week, and saves workers a reported 30% of time. [confirmed, McKinsey.com; self-reported]
• BCG Deckster, trained on 800-900 slide templates, is used by approximately 40% of associates weekly. [confirmed, Business Insider, April 2025]
• Harvey AI is used by 50% of Am Law 100 firms as of December 2025. [confirmed, TechCrunch, December 2025]
• AI platforms complete core legal research tasks 6 to 80 times faster than human lawyers in some task categories. [confirmed for speed; task scope varies; source is vendor and analyst benchmarks]

What is contested:

• A 2025 METR randomized controlled trial of 16 experienced open-source developers found a 19% slowdown in task completion time when using AI tools (Cursor Pro + Claude 3.5/3.7 Sonnet), against the developers’ own prediction of 20-24% speedup. [confirmed finding, contested generalizability]

  The METR study is rigorous and its finding should not be dismissed. But its scope matters: the tasks involved complex, mature legacy codebases; the developers were early in learning how to work with AI assistants; and coding standards constrained what AI could generate. METR itself launched a larger follow-up in August 2025 and found the signal unreliable in part because developers refused to participate in the AI-free control arm — suggesting strong practitioner conviction that the tools help. [confirmed, METR blog Feb 2026]

  A separate 12-month study across 300 engineers using an internal AI code platform found 33.8% cycle time reduction and 29.8% review time reduction (overall ~32% efficiency gain). [confirmed finding, arxiv 2509.19708]

  The honest picture: productivity gains from AI dev tools are real but context-dependent. Greenfield projects and simple tasks see larger gains. Complex legacy codebases see smaller gains or, in some conditions, slowdowns. [plausible, practitioner reporting; no controlled RCT of small startup teams]

• No independent study has compared deliverable quality between AI-enabled boutique consulting firms and Big 4 incumbents on equivalent client assignments. Quality parity is asserted in industry commentary but not demonstrated by controlled evidence. [unknown]

What matters most here: The production wall is falling for everyone. Incumbents are adopting the same tools at scale — McKinsey’s Lilli, BCG’s Deckster, Harvey at BigLaw. AI is making incumbents faster, not slower relative to boutiques. The production cost deflation story does not straightforwardly favor new entrants; it is as likely to strengthen incumbents by reducing their costs while keeping their gates intact.

What is unknown:

• How much of the cost deflation benefit accrues to small entrants versus large incumbents who can run inference at volume discounts. [unknown]
• Net headcount compression per unit output for small software teams. The claim circulates in VC narratives but has no empirical backing. [unknown]

Mechanism: Bundling gate (incumbents embedding AI into existing contracts)

What it is: The procurement decision gets made before the product comparison occurs. AI is bundled into existing platform agreements so the channel is pre-sold.

The clearest current example — Microsoft M365 / Copilot:

• Microsoft introduced Copilot as an M365 add-on in late 2023 at $30/user/month. [confirmed]
• In early 2025, Microsoft bundled Copilot into the consumer M365 subscription with a $3/month price increase and no consumer opt-out. [confirmed]
• A new enterprise E7 SKU ($99/user/month) bundles Copilot and Agent 365 together with M365 E5 security features in a single package. [confirmed, Microsoft 365 Blog, March 9, 2026]
• GitHub Copilot holds approximately 42% AI coding tool market share and is used by 90% of Fortune 100 companies. [confirmed, Gartner Magic Quadrant]

What the evidence also shows:

As of early 2026, only 3% of the M365 customer base has subscribed to paid Copilot — approximately 15 million paid licenses. [confirmed, Computerworld 2025-2026]

This matters. If the bundling gate had already closed, adoption would be higher and third-party displacement would be visible. The 3% figure suggests the bundling threat is being actively built, not yet fully installed. The window is still open. The question is how long it stays open as bundle normalization proceeds.

The FTC opened a wide-ranging antitrust investigation into Microsoft in November 2024, including Copilot and AI bundling, cloud licensing, and compatibility barriers. The investigation escalated under FTC Chair Andrew Ferguson (Trump appointee) as of February 2026. No enforcement action as of March 2026. The Trump administration AI Action Plan directs FTC to review whether prior investigations burden AI innovation — creating policy headwind for aggressive bundling enforcement. [confirmed facts; policy direction plausible, outcome unknown]

AWS / data gravity:

AWS Bedrock and SageMaker are native AWS services. Enterprises already on AWS can access foundation models without a separate contract. Bedrock Provisioned Throughput pricing starts at approximately $15,000/month — a floor that favors large enterprise buyers. [confirmed]

The mechanism here is not contractual exclusivity. It is integration convenience and data gravity: existing workloads on AWS make Bedrock the path of least resistance even without a formal bundle. [confirmed mechanism; market effect plausible]

IBM:

IBM Enterprise License Agreements span 3-5 years with substantial financial commitments. IBM routinely bundles additional products into ELAs at renewal. [confirmed, third-party IBM licensing consultancies]

No confirmed recent example of IBM using mainframe captive assets to specifically bundle AI offerings into multi-year ELAs. The general bundle-expansion pattern is documented; the AI-specific application is plausible but not confirmed by a named case. [plausible]

Mechanism: Certification moat (regulatory floors that favor established vendors)

FedRAMP:

• Legacy FedRAMP authorization takes 12-18 months and costs $500,000 to $1,000,000+, inclusive of third-party assessment, infrastructure changes, and internal labor. [confirmed, Secureframe / Convox 2025-2026]
• This is prohibitive for pre-revenue or early-stage companies. It is not prohibitive for incumbents who paid it years ago. [plausible; the competitive asymmetry is structural, not conspiratorial]
• FedRAMP 20x is a reform initiative targeting 3-6 month authorization for qualifying services. Phase 1 pilot ran April-September 2025 (26 submission packages). Phase 2 launched December 2025. Wide-scale availability targeted for FY26. [confirmed, FedRAMP.gov]
• FedRAMP 20x is not yet generally available as of March 2026. Whether it delivers on its timeline and cost-reduction promise is an open question. [confirmed status; outcome unknown]

SOC 2:

• SOC 2 Type I: $30,000-$80,000+ total first-year cost, including readiness, tooling, and internal labor. [confirmed, multiple compliance vendor guides]
• SOC 2 Type II: requires 6-12 months observation period; $25,000-$70,000+ in audit fees plus 50-100% of one person’s time for 4-6 months. [confirmed]
• SOC 2 is high table stakes, not a structural moat. A well-resourced funded startup can achieve it in under a year. The barrier is real but not prohibitive at this level. [plausible]

Mechanism: VC time bomb (subsidy creates exit pressure)

What it is: The gate period — the time between building a product and generating sustainable revenue — requires capital to survive. That capital comes from VC. VC creates return expectations that pressure toward exit. Exit in a consolidating market usually means acquisition by an incumbent.

The Cursor case:

• Cursor (Anysphere) reached $2B ARR by March 2026 — among the fastest B2B scaling trajectories on record, from $0 to $2B ARR in approximately 36 months. [confirmed, TechCrunch March 2, 2026]
• Cursor raised at a $29.3B post-money valuation (November 2025). [confirmed, CNBC]
• Cursor is not profitable. At the $1B ARR stage, it was spending approximately 100% of revenue on AI model costs (Anthropic and OpenAI APIs), running a ~$150M annual loss. [confirmed, multiple sources]
• In October 2025, Cursor launched its own proprietary coding LLM to reduce API cost dependency from ~100% of revenue to a projected 30-40%. Projected gross margins if this succeeds: 74% improving to 85% by 2027. [plausible; post-launch unit economics not yet confirmed]

Cursor’s $2B ARR is evidence the window of competition opened. A new entrant broke through against a dominant incumbent with 90% Fortune 100 penetration. The window is real. Cursor’s VC dependency and current losses mean its survival still depends on either the proprietary model strategy working, continued VC funding, or acquisition. The test is not whether it launched — it is whether it survives. [confirmed facts; survival outcome unknown]

The Windsurf case:

• OpenAI announced plans to acquire Windsurf for approximately $3B (May 2025). Talks fizzled due to OpenAI’s existing Microsoft deal. [confirmed, The Verge]
• Google hired Windsurf co-founder and CEO plus other senior employees in a $2.4B licensing/acqui-hire deal (July 2025). [confirmed, CNBC]
• Cognition acquired Windsurf’s remaining entity — IP, product, trademark, brand, remaining talent — approximately 72 hours later for an estimated $250M. [confirmed, TechCrunch]
• 100% of Windsurf employees received financial participation and fully accelerated vesting. [confirmed, TechCrunch / CNBC]

Be precise about who bears the cost in this outcome. Windsurf’s founders and employees received strong financial outcomes. The VC investors who funded the gap period are the clearer losers — this pattern is confirmed for acqui-hires generally; the Windsurf-specific VC outcome is plausible but not confirmed in detail. The competitive market is also a loser: a product that was competing with OpenAI and Google is now owned by one of them plus a startup backed by the other. [confirmed facts; VC loss pattern: plausible for acqui-hires generally; Windsurf-specific VC outcome unclear; characterization of competitive loss is plausible]

The broader pattern:

• Big Tech spent over $40B on acqui-hire deals in 2024-2025 combined. In 2025, 281 AI startup exits were recorded, mostly acqui-hires. [plausible; source is Futurum Group / Clera Insights; historical baseline unverified]
• The license-and-acqui-hire structure lets incumbents acquire technology and talent without triggering HSR merger review thresholds. [confirmed mechanism; whether this constitutes regulatory arbitrage is a legal question, not yet adjudicated]

Mechanism: Trust-based distribution gate (the moat AI cannot compress)

What it is: In professional services, the client relationship is the distribution channel. Institutional clients — large corporations, regulated entities, government agencies — do not evaluate new vendors on merit alone. They evaluate on reputation, prior relationship, referral network, and brand credibility as a proxy for accountability. These factors cannot be manufactured quickly.

What is confirmed:

• McKinsey’s Lilli synthesizes 100 years of IP across 100,000+ documents and client interviews, available to consultants and, in customized form, to clients. [confirmed, McKinsey.com] This is proprietary and not reproducible by a new entrant.
• BCG built GENE (internal AI chatbot) trained on internal knowledge, and BCG GAMMA as a specialized AI/data science unit. Both McKinsey and BCG have shifted toward “consultant + AI engineer” hybrid team models for client engagements. [confirmed, BCG.com; Business Insider, April 2025]
• BCG generated approximately $2.7 billion (20% of $13.5B total revenue) from AI-related advisory services in 2024. McKinsey’s AI and tech advisory represents approximately 40% of its estimated $16 billion 2024 revenue. [plausible; secondary sources; no primary revenue breakdown confirmed] Incumbents are growing AI revenue fast. They are not losing overall ground; any boutique gains are in specific niches, not the full market.
• Specialized AI consulting boutiques are described as gaining market share in high-growth micro-domains including sustainability and cloud optimization. [plausible; DNYUZ/Business Insider, November 2025; no named firm, named client, or verifiable case study located]

The Abridge/Epic pattern as a reference case: Abridge (AI medical scribing) gained market position not by displacing the dominant EHR provider but by attaching to it. It became Epic’s first “PAL” (Partner and Pal) and co-developed capabilities inside Epic’s workflows. That strategy worked until Epic launched a competing AI product in 2025, creating direct tension with Abridge’s position. [confirmed, STAT News, August 2025]

The generalization is limited — Abridge is a product company, not a consulting firm — but the pattern is instructive: in knowledge services, clearing the distribution gate often means becoming dependent on the entity that controls the channel. That dependency has a built-in fragility. The incumbent can absorb the product.

Mechanism: Liability coverage gate (procurement forms as structural barriers)

What it is: Large-client procurement requirements include minimum indemnification thresholds, named insured requirements, and sometimes minimum firm revenue floors. These requirements were designed around large incumbent firms. They function as structural gates against small boutiques independent of the quality of the work.

What is confirmed:

• Professional malpractice insurance for solo consultants costs approximately $400-$5,000/year, median around $662/year. [confirmed, NEXT Insurance, Hartford, Insureon, 2025-2026]
• Large corporate clients increasingly require outside counsel to carry minimum malpractice coverage of $5 million per claim or more as a procurement condition. [plausible; legal billing commentary and procurement guides; no primary survey data]
• Procurement-driven indemnification clauses can require coverage levels that exceed standard small-firm policy limits. [confirmed mechanism; ABA, Peabody & Arnold client alert]

What is unknown: The cost of professional liability coverage for mid-size boutique firms (10-50 employees) compared to large incumbents. The absolute gap is structural but its magnitude is not documented from a primary source. Whether pooled professional liability structures exist that allow small firms to access coverage levels required by large-client procurement has not been documented in available research. [unknown]

Mechanism: Proprietary data moat (real-time beats foundation models)

What it is: In financial analysis, the productive advantage of incumbents is not AI capability — it is data infrastructure. Bloomberg and FactSet hold real-time market data, decades of structured financial databases, and terminal distribution networks that cannot be reproduced by training a foundation model on public data.

What is confirmed:

• Bloomberg Terminal costs $20,000-$31,980 per user per year. FactSet and S&P Capital IQ run $10,000-$25,000 per user annually. [confirmed, vendor pricing pages, 2025]
• Bloomberg trained BloombergGPT on 363 billion financial tokens; its primary data moat is real-time market data and structured historical databases, not AI model capability alone. [confirmed, Bloomberg research blog, 2023]
• Bloomberg’s market mindshare is declining marginally (33.2% vs. 34.5% year-over-year). FactSet is growing marginally (21.7% vs. 20.2%). The trend is slow, not accelerating. [confirmed, PeerSpot, March 2026]
• AI-native research platforms (Marvin Labs, Captide, AlphaSense) offer workflows at materially lower cost than Bloomberg. [plausible; no primary market share data against Bloomberg/FactSet found]

The durable moat: Real-time data delivery and structured institutional databases are not reproducible by foundation model training on public data alone. The moat on document analysis and research synthesis is being eroded. The moat on raw data access and real-time feeds is not.

Mechanism: Regulatory gate (FINRA, UPL, and bar membership)

What it is: In financial research and legal services, regulatory requirements create structural gates that apply to human practitioners, not AI tools. The AI can do the analytical work, but the named human still needs the credential to publish or advise.

What is confirmed on FINRA:

• Research analysts whose name appears on an equity research report must pass the Series 86 and 87 exams and register under NASD Rule 1050 (now FINRA). [confirmed, FINRA.org]
• In June 2024, FINRA issued Regulatory Notice 24-09 confirming that existing regulatory obligations apply to generative AI use, with no new AI-specific exemptions. [confirmed, Debevoise & Plimpton analysis of FINRA 2025 Regulatory Oversight Report]
• FINRA registration applies to the named human analyst, not the AI tool. [plausible; no enforcement action against an AI-enabled research firm under FINRA registration requirements located]

What is confirmed on UPL:

• Unauthorized Practice of Law is enforced at the state level. Enforcement is described by the ABA and NCSC as “infrequent and ad hoc,” often triggered by lawyer complaints rather than consumer harm. [confirmed, ABA Law Practice Magazine, March-April 2025; NCSC White Paper, August 2025]
• There is no uniform national definition of “practice of law” for UPL purposes. [confirmed, ABA, NCSC]
• Utah created a regulatory sandbox allowing innovative legal services to operate under relaxed UPL rules. Colorado’s Access to Justice Commission formed a subcommittee to consider AI-related UPL amendments. [confirmed, NCSC White Paper, 2025]
• No major UPL enforcement action against an AI legal tool provider has been located in the research period. [plausible; enforcement risk exists but has not materialized as a systematic gate as of March 2026]

The interaction with AI: An independent analyst or attorney can use AI to do 80% of the work faster and at lower cost. The regulatory gate applies to the output — who can publish it, who can advise on it, who carries liability for it. AI compresses the production; the credential still gates the delivery.

Mechanism: IT Ladder echo in professional services (the incumbents’ internal risk)

What it is: When AI automates the work that entry-level employees learn on, it eventually degrades the institution’s ability to oversee the systems it depends on. This risk is documented in IT and software organizations. The same dynamic is now operating in professional services — on the incumbent side.

What is confirmed:

• McKinsey is considering cuts of approximately 10% of its workforce (3,600 people at roughly 36,000 headcount, already 25% below historical peak). [confirmed, Bloomberg, December 2025]
• UK Big 4 graduate job listings in accounting fell 44% year-on-year by 2025. AI and offshoring cited as primary drivers. [confirmed, The Finance Story, 2025]
• EY delayed graduate start dates three years running (2023, 2024, 2025). 2025 hires not starting until March 2026. [confirmed, The Finance Story, 2025]
• BigLaw summer associate class sizes (2L) fell to their lowest average since 2021, averaging 12 students in summer 2024. [confirmed, BigLaw recruiting trackers]
• 44% of law firms had no formal AI governance policies in place as of 2025, despite 79% of legal professionals using AI tools. [confirmed, Bloomberg Law, 2025]
• Overall law graduate employment reached a record high in 2024 (93.4% employed within 10 months). AI displacement of junior lawyer jobs is not yet visible in aggregate employment data. [confirmed, NALP Class of 2024]

What is unknown: No documented case of an oversight failure or client-facing error directly caused by hollowing the associate bench in consulting, legal, or financial services. The capability loss risk is structurally plausible and named in industry commentary, but no primary incident has been documented. [unknown]

If junior consultants are not doing synthesis — because AI does it — the question of who supervises AI-generated synthesis with contextual judgment becomes load-bearing. That supervision function requires people who have done the work. Institutions contracting that function without building the bench are creating the same brittleness that AI-augmented IT departments have already demonstrated: human oversight that is present on paper and absent in practice.

4. Five-scene comparison

Scene 1: Dev/software shops

The tension that must be held: AI displacement of software workers is feeding a new-entrant pipeline that is real. Business applications reached 5.48M in 2023 (highest on record) and continued at elevated levels in 2024-2025. About one in four tech layoffs in 2025 was linked to AI-driven restructuring — this reflects employer-attributed characterizations in layoff announcements, not independently verified causal findings. [confirmed as reported characterization, TechCrunch 2025 layoffs compilation; causal attribution unknown]

High-earning independent workers (above $100,000/year) grew from 3M (2020) to 5.6M (2025). Full-time independent workers more than doubled from 13.6M to 27.7M over the same period. This covers independent workers broadly, not tech specifically. [confirmed, Talent Intelligence Collective; scope caveat applies]

The squeeze is real and the seed is real. Displaced workers are starting companies. Some of those companies build products better than the incumbents. Cursor is the sharpest evidence: $2B ARR against GitHub Copilot’s 42% market share and Fortune 100 entrenchment. [Cursor $2B ARR: confirmed, TechCrunch March 2026; Copilot 42% share: confirmed, Gartner Magic Quadrant; Cursor’s own share is not confirmed from a named source and is omitted here]

But the survival test is not the launch test. The bottleneck moves from building to selling, from selling to sustaining, from sustaining to surviving long enough that VC patience runs out. Most entrants in this sequence do not have Cursor’s growth rate. Those who don’t are more likely to fail or to be absorbed at a price that ends competition rather than extends it.

Do not resolve this tension into pure pessimism or pure optimism. Both are wrong. The accurate picture is: the window opened, and it is fragile, and what happens to it depends on policy choices that are currently contested.

Scene 2: Agencies / creative shops

The production gate is down for the right reasons here — small creative shops can now produce work that is genuinely competitive with established agencies. The wall fell; the question is whether the evaluation ever happens.

The incumbent defenses are not certification or platform bundling. They are:

• Relationship lock-in: long-standing client relationships built on trust, embedded brand knowledge, and familiarity with the client’s internal politics and preferences
• Institutional credibility: the ability to say “we are [Agency Name]” as a shorthand for accountability and risk absorption — especially relevant for large clients who need someone to blame if a campaign fails
• Team-composition theater: large agency pitches present account managers, strategists, creative directors, and production teams. A two-person shop with AI tools cannot replicate the presentation even if it can replicate the output

The good-loop case is stronger here than in the dev/software scene. A client who genuinely evaluates creative work on merit — and not on agency size or relationship inertia — can choose the smaller shop and receive equivalent or better work at lower cost. This happens. The constraint is not that the window is closed; it is that most clients do not run merit-based evaluations. They renew on relationship inertia.

The bottleneck here is buyer behavior and relationship lock-in, not platform ranking power. That shapes which levers apply: procurement reform and reference network access matter more than provenance labeling or anti-spam friction. A buyer who requires documented alternatives evaluation before renewing an agency relationship is applying the same discipline a procurement department should apply to enterprise software — and the same action is repeatable. [plausible mechanism; no controlled study of creative agency procurement patterns found]

Scene 3: Management consulting

The more documented story here is the opposite of disruption: BCG generated approximately $2.7 billion (20% of $13.5B total revenue) from AI-related advisory services in 2024. McKinsey’s AI and tech advisory represents approximately 40% of its estimated $16 billion 2024 revenue. [plausible; secondary sources; no primary revenue breakdown confirmed] Incumbents are growing AI revenue fast. They are not losing overall ground.

The defensive strategy is not platform lock-in or procurement bundling but workflow integration depth. McKinsey and BCG embed AI tools inside existing client relationships. A new entrant can produce analysis as good as a Big Three deliverable. Getting that analysis into a client’s decision workflow requires the same trust and procurement relationships that have always existed. AI lowered the production wall; the distribution wall — embedded tools, long relationships, and procurement that rarely compares — did not move.

Boutiques are described as gaining ground in high-growth micro-domains including AI advisory, sustainability, and cloud optimization. [plausible; DNYUZ/Business Insider, November 2025; no named firm, named client, or verifiable case study located] The gains are real at the margins. The mainstream enterprise market shows no documented shift toward AI-enabled boutiques winning at the expense of MBB.

Scene 4: Financial analysis

The FINRA gate is structural but not impenetrable. A displaced analyst can register, build the infrastructure, and operate independently — it takes time and capital, not impossibility. What AI compression changed is that the underlying analytical work now takes less time, which potentially makes the economics of a small independent research operation more viable.

The harder gate is institutional distribution. Equity research reaches clients through terminal connections, sales relationships, and compliance infrastructure. An independent analyst producing better research than a sell-side shop cannot distribute it to institutional clients who require registered counterparties and data compliance. The production wall has fallen. The distribution and registration walls have not moved in proportion.

The MiFID II experience in the EU is a reference case: when European regulations required unbundling of research payments, independent boutique and niche research firms gained market share relative to middle-market sell-side firms. [plausible; Integrity Research; Tandfonline, 2024] US research payments remain largely bundled. Whether MiFID II-style unbundling dynamics would replicate in the US is unknown. The regulatory mechanism that created the EU opening does not exist in the US market.

Scene 5: Legal and compliance

Harvey’s 50% Am Law 100 penetration tells the more accurate story about how AI is changing legal services: incumbents are adopting the tools and getting faster, not being displaced by boutiques with better tools. The production wall is falling for BigLaw. This may strengthen incumbents — their costs fall while their relationship and insurance gates hold.

The boutique entrant gains are real at the margins: mid-market clients, less complex matters, clients who will evaluate on merit and price. But the large-client market — the market with the highest fees and the most structural gates — shows no documented shift toward AI-enabled boutiques winning at the expense of BigLaw. [unknown; no documented case located]

UPL remains a gate with uncertain enforcement. It exists but is not being systematically enforced against AI tools as of March 2026. If enforcement tightens — plausible but not yet happening — the gate could harden for AI-enabled independent practitioners. The current uncertainty is a feature for incumbents: it creates legal ambiguity that risk-averse clients resolve by staying with licensed firms.

5. Where it broke — evidence against the good story

1. AI is making incumbents faster, not slower. McKinsey’s Lilli is in use by 75% of the firm. Harvey’s 50% Am Law 100 penetration means BigLaw is adopting AI tools at scale. BCG Deckster at 40% weekly associate use. These are figures for incumbents deploying AI at scale inside existing client relationships, with existing trust, liability coverage, and data infrastructure already in place. The production wall is falling for incumbents and boutiques simultaneously. Incumbents may be getting faster and more profitable while boutiques are still trying to reach quality parity. [confirmed adoption figures; competitive implication is plausible; no direct boutique adoption comparison data found]

2. Billing rates and revenues are rising, not falling. BigLaw billing rates climbed 9.2% in the first half of 2025. Law firm revenues rose 11.3%. Law firm net income rose 17% in 2024. AI consultant billing rates rose from approximately $550/hour in 2022 to $895/hour in 2024. [plausible; source is industry rate surveys (secondary sources), not primary data; Legal.io; Axiom Law] If production costs are falling, incumbents are capturing the gains as margin, not passing them to clients.

3. Slow incumbent adoption cuts both ways. Microsoft Copilot at 3% paid M365 adoption after two years means the bundling gate is not yet installed. But it also means the incumbent is absorbing the cost of building the gate infrastructure at a pace a new entrant cannot match — the slow close may be the slow structural lock-in. [confirmed fact; interpretation plausible]

4. The METR productivity slowdown. If AI tools do not reliably improve experienced developer productivity in complex legacy environments, the cost of building with a small team does not drop as dramatically as the production-cost-deflation story implies. The “team of one ships what used to need ten” narrative is plausible for greenfield projects, not confirmed as a general baseline. [confirmed finding with scope constraints]

5. No VC-independent US entrant has won. No confirmed example was found of a US-based, VC-independent AI software entrant gaining durable enterprise market share against an incumbent holding platform bundling position. Cursor leads the coding tools market — but Cursor requires continued VC funding or successful proprietary model execution to survive. Abridge gained healthcare AI market position through Epic distribution dependency. [unknown / no example found]

6. Boutique market share gains in professional services are reported but not documented. The November 2025 reporting on boutique consulting gains names no firm, no client, and no RFP. No primary market share data quantifying boutique gains versus MBB or BigLaw incumbents has been located. [plausible; no primary data]

7. The acqui-hire pattern is structural. 281 AI startup exits in 2025, mostly acqui-hires. The mechanism by which talent and technology reach incumbents without triggering antitrust review is operating at scale. [plausible; Futurum Group / Clera Insights; historical baseline unverified]

8. Open-weights deflation is real but not a market-structure fix. The Chatbot Arena quality gap between open and closed models shrank from 8% to 1.7% between January 2024 and February 2025. [confirmed, CB Insights / Red Hat Developer] This reduces the inference cost dependency that makes startups like Cursor unprofitable. But it does not address the procurement bundling gate, the FedRAMP gate, or the trust and liability gates in professional services. Better and cheaper models help entrants reach the gate faster; they do not change what the gate requires. [confirmed facts; competitive implication is plausible]

6. Where it held — evidence the window is real

1. Cursor reached $2B ARR against a dominant incumbent. GitHub Copilot had 42% market share and 90% Fortune 100 penetration when Cursor launched. Cursor reached $2B ARR in approximately 36 months — one of the fastest B2B scaling trajectories on record. A new entrant gained meaningful traction against an incumbent with near-total Fortune 100 penetration. The window opened. [$2B ARR: confirmed, TechCrunch March 2026; Copilot 42% share and 90% Fortune 100 penetration: confirmed, Gartner Magic Quadrant]

2. Market fragmentation exists in software. GitHub Copilot (42%), Google Gemini (used by 47% of developers), Anthropic Claude (used by 41% of developers), and Cursor at meaningful but unconfirmed scale — multiple entrants are operating simultaneously. This is not winner-take-all consolidation as of March 2026. [Copilot share: confirmed, Gartner; developer usage figures: confirmed, Menlo Ventures 2025; Cursor’s specific share not confirmed from a named source]

3. Niche specialization is a real opening in professional services. In high-growth domains where incumbents lack deep expertise — AI advisory, sustainability, specific regulatory domains — boutiques with specialized knowledge can win on substance. AI may expand the number of niches where a small team can compete, even if the mainstream enterprise market remains gated. [plausible; directionally supported by industry commentary]

4. Mid-market clients are less gate-dependent. The liability coverage, institutional credibility, and regulatory gates are most severe in large-client procurement. Mid-market clients are more likely to evaluate on merit and price, more likely to try a new vendor, and less likely to have procurement checklists designed around large incumbents. The window is more open in these markets. [plausible; no controlled market share data]

5. Open weights reduced the foundational model moat. If the quality gap between open and closed models continues to narrow, the inference cost dependency that currently makes new entrants unprofitable decreases. This creates a path to sustainability that does not require VC or acquisition — if the path is not foreclosed by bundling before it can be taken. [plausible]

6. FedRAMP 20x is a real reform. If it delivers on the 3-6 month, lower-cost authorization promise at scale in FY26, it materially lowers one of the primary certification gates. This is not confirmed — it is in Phase 2 pilot — but it is a concrete structural lever that is moving. [confirmed status; outcome pending]

7. UPL is not currently a systematic enforcement threat. The absence of major UPL enforcement actions against AI tools means the legal services window is currently more open than the formal gate structure implies. An AI-enabled boutique operating in the drafting and information domain operates in an enforcement gray zone that has not been systematically closed. That openness could tighten; it is currently real. [plausible; ABA; NCSC, 2025]

8. The IT Ladder risk creates a real opening for oversight-oriented boutiques. Firms that have hollowed their junior bench face genuine oversight risk as AI-generated work volume increases. A boutique that explicitly staffs for human review, quality control, and escalation judgment has a differentiable value proposition in risk-sensitive engagements. [plausible; no documented client preference data found]

9. Regulated industries have structural reasons to avoid lock-in. Banking, telecom, and healthcare sectors have data residency requirements that make on-premise deployment mandatory. These sectors have structural incentives to adopt open models and avoid proprietary API dependency. [confirmed, IBM Think / Red Hat Developer]

7. Human Command test and Exit test

Human Command test

This case is not primarily about AI making life-outcome decisions for individuals. The Human Command framework applies in two narrower places:

Workers displaced through AI restructuring. Workers laid off through “AI-driven restructuring” rarely receive specific, auditable reasons for their displacement. The decision is made by a system or a process, and the reason given is typically a category label, not a mechanism. Junior analysts, associates, and consultants whose training pipeline has been compressed face a specific version of this: they received delayed start dates, smaller summer classes, and restructuring notices — not notice that AI automation was reducing their developmental path. The affected workers have no meaningful appeal path. [plausible; no systematic disclosure study found]

Clients in procurement relationships. Enterprise clients running professional services through procurement checklists designed around large incumbents cannot easily evaluate alternatives. The checklist forecloses the comparison before it happens. The decision to use the incumbent was made before the boutique proposal arrived. This is a soft version of the Human Command failure: the evaluation was nominal rather than real.

Exit test

Exit is constrained differently across scenes:

• Software / enterprise tech: Once AI is embedded in an existing M365 or AWS contract, switching requires a separate procurement process that most IT departments will not initiate without a compelling reason. FedRAMP lock-in means federal market incumbents with existing authorization face no equivalent new compliance cost when expanding. Exit is not impossible — coding tool market fragmentation shows real switching behavior — but it is meaningfully harder for enterprise buyers and functionally impossible for federal market buyers without FedRAMP reform.
• Professional services: Switching from a large incumbent requires running a procurement process that most organizations will not initiate without a compelling reason. Relationship inertia is the primary constraint, not contractual lock-in. Exit is theoretically available; it is not routinely taken.
• Junior workers: Exit from a hollowed training pipeline is not reversible. A junior analyst whose core learning tasks have been automated away does not recover that developmental time when they leave. The capability loss is individual and structural.

Where exit is not realistic — especially for clients in regulated procurement contexts — governance requirements need to be stronger. The current enforcement posture is investigation without action in software and norm without requirement in professional services. [confirmed facts; governance gap is confirmed; adequacy is contested]

8. Shared Gains test

If AI creates productivity gains across software development, creative production, and professional services, who gets them?

In professional services, the evidence is unusually direct:

• BigLaw billing rates rose 9.2% in the first half of 2025. Law firm revenues rose 11.3%. Net income rose 17% in 2024. [confirmed, Legal.io; Axiom Law]
• Among legal professionals using AI widely, 25% increased prices, 11% reduced prices, and 8% added AI-specific fees. [confirmed, Clio Legal Trends Report, 2025]
• AI consultant billing rates rose from approximately $550/hour (2022) to $895/hour (2024). [plausible; source is industry rate surveys, not primary data]
• EY research found that AI-driven productivity gains are primarily fueling organizational reinvestment rather than workforce reductions or fee reductions. [confirmed, EY newsroom, December 2025]

If production costs are falling, incumbents are capturing the gains as margin. The shared gains test fails on the evidence in professional services.

Across all five scenes:

• Workers whose labor is displaced bear the cost — engineers, creative directors, analysts, associates, and consultants. Some receive severance; most receive no transition support tied to the AI productivity savings. [confirmed displacement; transition support unknown]
• New entrants receive lower production costs. Whether they can capture economic value from those savings depends on whether they can survive the gate period — bundling in software, relationship moat in agencies, trust-based distribution in consulting, FINRA registration in financial analysis, liability coverage in legal. Most cannot without VC or an existing client base. [plausible]
• VC investors fund the gate period and absorb loss when acqui-hires pay below valuation. In acqui-hire structures, VC investors are the clearest losers among the parties. [confirmed, Windsurf deal structures]
• Founding teams and employees in acqui-hires have received strong financial outcomes. [confirmed, Windsurf; Inflection; Character.AI]
• Incumbent platforms and large firms acquire technology and talent at below-market cost (relative to valuation) and eliminate competitive products and services from the market. [confirmed as pattern; individual deal economics vary]
• Buyers — enterprise customers, agency clients, and professional services clients: in a contestable market, they get more choice and lower prices. In a consolidated market, they get the incumbent’s bundle or retainer at the incumbent’s price. Each scene is currently somewhere between those poles. [plausible; no longitudinal pricing data found across scenes]
• Households: no direct mechanism for AI productivity gains in software, creative, or professional services to reach household costs found in the research. The long-run argument is that cheaper services lower prices across the economy. That argument is plausible and not confirmed at measurable timescale. [unknown]

The distribution question: the default is that gains flow to whoever controls the platform or the client relationship. Competition is the mechanism that changes that — customers win through pricing pressure when real alternatives exist, workers win through mobility when they can leave for a better option. The problem here is not that rules are absent. It is that the competitive window is at risk of closing before competition can do that work.

9. Minimum floor

For the competition window to stay open long enough to matter across all five scenes, five conditions must hold:

1. Procurement neutrality — software and enterprise tech. Enterprise AI products must be evaluable on a standalone basis before bundling locks the decision. This requires: antitrust enforcement that treats bundling-as-foreclosure as actionable; interoperability requirements that allow products to function in mixed-vendor environments; and FTC follow-through on the Microsoft investigation. [FTC investigation active; enforcement outcome unknown]

2. Documented alternatives evaluation — agencies and professional services. In creative and professional services markets, the equivalent ask is a buyer practice: require documented alternatives evaluation before renewing any agency retainer or professional services contract. This is not a regulatory requirement; it is a buyer discipline that procurement teams at large organizations can institutionalize without legislation. For knowledge services specifically, liability and E&O coverage thresholds that structurally exclude small firms warrant review — the current bar reflects risk management norms established before AI-capable small teams existed. [plausible; no reform active as of March 2026]

3. Certification access — software and federal market. FedRAMP 20x must deliver wide-scale availability for Low and Moderate impact services in FY26 at the 3-6 month, lower-cost promise. If it does, the federal market gate opens meaningfully for small entrants. If it slips or fails to deliver on cost, the structural advantage of pre-certified incumbents persists. [reform active; outcome pending; timeline is FY26]

4. Insurance access and AI governance standards — regulated professional services. If minimum indemnification coverage requirements in client contracts function as a structural gate, the remedy is either pooled professional liability structures that allow small firms to access required coverage levels, or procurement reforms that require justification when thresholds exceed actual engagement risk. Published AI governance standards from bar associations, FINRA, and procurement standards bodies would simultaneously raise the floor for incumbents and create a legible standard small firms can meet to demonstrate accountability. Neither exists at scale. [unknown; no pooled structure or published standard documented as of March 2026]

5. Acqui-hire as acquisition. If the license-and-acqui-hire structure is treated as a substantive acquisition for antitrust purposes — triggering HSR review — the mechanism by which talent and technology reach incumbents without scrutiny becomes more costly. The same logic applies when large consulting or advisory firms acquire boutique competitors primarily to absorb client relationships rather than integrate capabilities. This is currently an open legal question; no enforcement precedent has been set. [unknown; no action as of March 2026]

These are not abstract policy wishes. Each has a specific institution or practice responsible for it, a specific timeline or trigger, and a specific measurable outcome. They can be tracked.

10. What to do

Personal / interpersonal

If you are an enterprise or professional services buyer: ask your procurement team what the actual comparison process is for any vendor relationship — software, creative, consulting, legal, financial — before the next contract is signed or renewed. “It’s in the bundle” and “we’ve worked with them for years” are process descriptions, not decisions. Requiring a documented alternatives evaluation before contract signature is a repeatable action that procurement departments can institutionalize across all five scenes in this case.

If you are a displaced worker building with AI tools: the gate period is the survival problem, not the build problem. The specific gate depends on your market:

• In software and enterprise tech, the gate is certification and procurement: FedRAMP Moderate for the federal market is a 12-18 month, $500K-$1M commitment that does not compress because your product is good. SOC 2 Type II is achievable in under a year with $50,000-$100,000 and is the right first milestone for commercial enterprise sales.
• In creative and agency work, the gate is relationship inertia: target buyers who run documented competitive evaluations, not buyers who renew on autopilot. One client who chose you in a real comparison is worth more than ten warm introductions that never reach a decision.
• In consulting and professional services, the production problem is not your hardest problem. Know the specific insurance and regulatory requirements for your target client segment before you price the engagement. A mid-market client with a simpler procurement process and a lower coverage floor is a more realistic first client than a large enterprise with a procurement committee. The trust cycle takes time; capital can support you through it, but it cannot buy the track record.
• In financial analysis, the FINRA registration requirement is time and capital, not impossibility. Build the registration infrastructure early — it is the distribution gate, and distribution is the hard problem, not the analysis.
• Across all five: identify your gate before you build. The production problem is solvable with AI tools. The gate problem requires a strategy that matches the specific wall you are facing.

If you are a junior analyst, associate, or consultant entering a hollowed pipeline: the tasks AI is doing are the tasks that build judgment. Find the environments — boutiques, smaller firms, roles with oversight responsibility — where the learning work still exists. The capability gap between people who did the work and people who supervised AI doing the work will matter when oversight failures become visible.

Community / policy lever

Short-term: Advocate for FedRAMP 20x on schedule. The reform is real, its timeline is FY26, and it will slip without pressure. Contact your congressional representative’s technology policy staff and ask specifically about FY26 FedRAMP 20x implementation. This is narrow, specific, and actionable.

Also short-term: ask professional services procurement teams — in your organization, your employer, your industry group — to adopt a documented alternatives evaluation standard for services contracts above a materiality threshold. This is a norm change, not a legislative fight, and it is achievable at the organizational level.

Medium-term: Support FTC investigation follow-through on Microsoft AI bundling. The investigation exists; enforcement is the question. FTC investigations do not run public notice-and-comment periods, but industry associations and companies with standing can submit written statements through counsel, and congressional pressure on the FTC’s enforcement priorities is a documented lever. If you work in an organization affected by AI bundling, flag it to your industry association and legal counsel — their input can reach the record through proper channels. [confirmed investigation; policy headwind confirmed]

Support bar association and FINRA rulemaking that establishes published AI governance standards for legal and financial services practitioners. The ABA’s active work on UPL modernization and FINRA’s 2025 AI guidance are the right venues. Standards set in the next two years will define the floor for both incumbents and boutiques.

Long-term: Push for interoperability requirements in AI procurement — rules requiring that enterprise AI products function in mixed-vendor environments and that data is portable between platforms. No US equivalent of the EU’s Digital Markets Act is moving toward enactment in the 119th Congress as of March 2026. [confirmed gap; legislative action unknown]

Advocate for procurement neutrality requirements in professional services — rules that require large clients to demonstrate why their coverage and qualification floors are calibrated to actual engagement risk rather than to incumbent firm scale. No equivalent exists in the US market. [confirmed gap; no legislative movement identified]

Failure mode of these levers: Antitrust investigations that do not reach enforcement are regulatory theater. FedRAMP 20x that slips its timeline has not changed the structural problem. Procurement evaluation standards that exist on paper but are not enforced by budget owners are compliance theater. AI governance standards drafted without sunset and review provisions will be obsolete before they are enforced. Each lever has a real version and a nominal version. The test is the measurable outcome, not the announcement.

Loop Effect

Effect on the bad loop

• Monthly squeeze: Displaced workers — engineers, creative directors, analysts, associates, consultants, attorneys — face income disruption when their new ventures reach the market but cannot survive the gate period. Junior workers whose training pipeline is compressing face a subtler version: they are receiving the signal that their entry path is narrowing without being told why or for how long. [plausible]
• Insecurity: The speed and opacity of “AI-driven restructuring” layoffs — where the mechanism is rarely disclosed — creates insecurity without accountability. EY’s three consecutive years of delayed graduate starts, with no official explanation tied to AI automation, is a case of institutional opacity creating insecurity for workers who cannot contest what they cannot see. [confirmed facts; insecurity framing is plausible]
• Manipulation / scapegoats: The productivity-gains-from-AI narrative makes it easier to frame displacement as inevitable technology progress rather than as a distribution choice. Rising billing rates at the same time AI adoption is publicly promoted as a cost reducer creates a narrative gap. The gains are real; who receives them depends on whether competition is alive to distribute them. A consolidated market routes gains upward by default. [plausible; confirmed EY research]
• No fixes / more squeeze: Acqui-hire consolidation, unchecked bundling, and delayed FedRAMP reform each weaken contestability in software. Without procurement evaluation discipline or insurance access reform, the structural advantage of large incumbents in professional services strengthens as AI makes them more productive while keeping their gates intact. [plausible]

Effect on the good loop

• Security: Lower gates across all five scenes mean more companies can reach buyers — FedRAMP 20x for software entrants, open reference networks and evaluation norms for creative and consulting, reduced certification barriers for financial and legal services. More companies means more hiring, more wage pressure, and more product options. Boutiques that survive the gate period create real career alternatives for displaced professionals and real pricing alternatives for clients. [plausible, contingent on reform delivery in each scene]
• Choice: Real alternatives exist in all five scenes as of early 2026 — fragmented dev tools, independent creative shops with genuine capability parity, boutique advisory firms that can deliver against larger incumbents, and AI-enabled independent analysts and attorneys operating in enforcement gray zones. Sustaining those alternatives requires preventing the bundling gate, relationship moat, and trust-based distribution lock from closing before the entrant class can reach profitability without VC dependency.
• Competition: The minimum floor is the same ask applied to different gate types: procurement evaluation (don’t renew on inertia in software, agencies, or professional services), certification access (reduce structural cost of entry), insurance and governance standards (create legible floors boutiques can meet), and acqui-hire scrutiny (don’t let consolidation absorb the competitive window before it can function). Any one of these, delivered with enforcement or habit, makes rigging harder.
• Shared gains: When the competitive window stays open, gains flow through normal market behavior. Customers win through pricing pressure — more competitors means real choice on quality, price, and values. Workers win through mobility — a market with real alternatives means they can leave for a better employer when one exists. Competition does the distributional work. The problem here is not that rules are needed to force gain-sharing. It is that the competitive window is at risk of closing before competition can do that work. Foreclosed markets distribute gains upward by default. Keep the window open and the market follows.

Case verdict

• Net effect right now: Mixed, with variation by scene. The software window is real and currently open — Cursor’s $2B ARR against GitHub Copilot’s incumbent position is genuine evidence. Three or four meaningful competitors operating at scale in coding tools is not consolidation. In professional services, the window is more mixed: AI is making incumbents faster and more profitable, not slower. Billing rates and revenues are rising while production costs fall. The productivity gains are flowing to incumbents as margin.
• Forward projection (contingent): The trajectory tilts toward bad loop if specific closing mechanisms proceed unchecked: bundling normalization (M365 Copilot adoption accelerates past 3%), VC exit pressure resolves into acqui-hire absorption rather than profitable independence, certification gate reform slips, and professional services procurement norms don’t shift toward documented alternatives evaluation. If those mechanisms close before policy catches up, the current fragmentation resolves into consolidation. That outcome is plausible and worth preventing — but it is a projection, not the current state.
• What would change the projection: FTC enforcement action on Microsoft bundling that establishes precedent; FedRAMP 20x at scale in FY26 with confirmed cost reductions for small vendors; documented alternatives evaluation becoming standard in large-client procurement; published AI governance standards from ABA and FINRA; acqui-hire transactions reviewed under HSR merger standards; Cursor achieving profitability through proprietary model execution. Any two of these would be enough to hold the window open meaningfully. [Each is an open question as of March 2026]

One steady action

When your organization renews a vendor relationship — software, creative agency, consulting firm, law firm, or research provider — ask what alternatives were evaluated before the contract was signed. One written question: “what alternatives were considered and why was this selected?” That question applies equally to an enterprise software bundle, an agency retainer, a consulting engagement, and a legal or financial services contract. Records make renewal-by-default visible. Visible patterns are the precondition for changing them.

North Star verdict

The E4E thesis is: security enables choice, choice enables competition, competition produces shared gains. The lower-walls-harder-gates pattern is a case where the gains from AI productivity deflation are real but are not reaching the people most exposed to its costs.

In software and enterprise tech: displaced workers fund the new entrant pipeline. New entrants deliver innovation. That innovation is absorbed into incumbents through acqui-hire structures that protect founding teams and employees but eliminate the competitive product. The cycle repeats. The market fragments briefly and then consolidates. The gains flow to platforms.

In professional services: AI makes incumbents faster, not slower. McKinsey’s Lilli, Harvey at BigLaw, BCG’s Deckster — the production wall is falling inside existing client relationships, with existing trust, liability coverage, and data infrastructure already in place. Billing rates and revenues are rising. The boutique entrant is not displacing the incumbent; the incumbent is getting better, cheaper, and more entrenched.

This is not a verdict that “AI is bad for competition.” The Cursor case is genuine evidence that competition is alive in software. The open-weights deflation is genuine evidence that the foundational model moat is weakening. Boutiques are winning in niches and mid-market segments. The window is real.

The verdict is narrower: the window of competition opened, it is currently real, and it will not stay open by default. The minimum-floor conditions — procurement neutrality, certification access, documented alternatives evaluation, insurance access for small firms, and acqui-hire scrutiny — are not being met by current market behavior or policy. The current policy posture is investigation without enforcement and reform without completion.

A strong middle class stabilizes markets. Displaced workers — engineers, creative directors, analysts, associates, consultants, attorneys — who cannot survive the gate period are not stabilizers. They are a churn mechanism that funds consolidation and then returns to the labor market without the capital gain their work created. The good outcome here is not “a startup wins.” It is the window of competition staying open long enough that customers have real choice, workers have real alternatives, and no single platform or incumbent can foreclose the market before the next cycle begins.

The system lesson for professional services: When AI compresses production costs in trust-based markets, the gains accrue to the entity that holds the trust. Making those markets more contestable requires addressing the trust and liability gates directly — through procurement evaluation requirements, insurance access reforms, and published governance standards — not assuming that production cost parity translates to competitive parity.

Research gaps

• [RESEARCH GAP: What fraction of the 2023-2025 business formation surge represents tech-displaced workers starting software or AI companies versus other business types? Census BFS does not break out by industry at this resolution.]
• [RESEARCH GAP: Whether any US-based, VC-independent AI software entrant has achieved durable enterprise market share (>5% of addressable market) against an incumbent holding platform bundling position. No confirmed example found.]
• [RESEARCH GAP: What fraction of enterprise AI procurement decisions are driven by incumbent bundle versus independent tool evaluation. No survey data found measuring this directly.]
• [RESEARCH GAP: No independent, controlled study comparing deliverable quality between AI-enabled boutique consulting, legal, or financial services firms and major incumbents on equivalent assignments. Quality parity is asserted in industry commentary but not demonstrated with evidence.]
• [RESEARCH GAP: No primary market share data (revenue, RFP win rates) quantifying boutique gains versus MBB, BigLaw, or institutional financial incumbents. The directional commentary is plausible but not verified by named firms, clients, or transactions.]
• [RESEARCH GAP: No documentation of pooled professional liability structures that allow small consulting, legal, or financial services firms to access the coverage levels required by large-client procurement. Whether such structures exist as an active market is unknown.]
• [RESEARCH GAP: Professional liability insurance cost for mid-size boutique firms (10-50 employees) versus large incumbents. The absolute cost gap is structural but its magnitude is not confirmed from a primary source.]
• [RESEARCH GAP: Whether any procurement reform specifically requiring documented alternatives evaluation in professional services contracting exists in the US market, beyond general federal small-business set-asides. None located in research.]
• [RESEARCH GAP: Whether MiFID II-style research payment unbundling would produce analogous independent research firm growth in the US market. The EU precedent is documented; the US regulatory mechanism that would create the same opening does not currently exist.]
• [RESEARCH GAP: Whether large consulting, legal, or financial services firms are training or fine-tuning AI models on client engagement data beyond retrieval-augmented generation on internal documents. McKinsey’s Lilli and BCG’s GENE are documented for RAG on internal IP; AI-specific data moat creation via fine-tuning on client data is not publicly confirmed.]
• [RESEARCH GAP: No documented case of an oversight failure or client-facing error directly caused by hollowing the associate or analyst bench in consulting, legal, or financial services. The IT Ladder echo risk is structurally plausible and named in commentary; no primary incident documented.]

Back to case studies