Economy for Everyone
Most of us picture AI as a box where someone types a question and gets an answer. That’s the part we can see. When AI is used to evaluate an insurance claim, score a job application, or check benefits eligibility, the important choices usually happened earlier: what the system was allowed to know, which records it pulled in, which rules it trusted, what it ignored, and whether it kept a trail.
The interface might be a chatbot, a worker clicking through a portal, or nothing at all - the affected person never touches the system and just receives a result. Either way, the pattern is the same: you see the answer. The institution controls what the system saw.
Anthropic draws the line precisely. Prompt engineering covers “methods for writing and organizing LLM instructions for optimal outcomes.” Context engineering covers “strategies for curating and maintaining the optimal set of tokens during LLM inference”: the retrieved documents, source rankings, policies, memory, and exclusions the institution built in before anyone arrived.
The prompt is the UI. Context is the architecture.
What “context” means
When a company deploys an AI agent for customer service, claims processing, benefits eligibility, or hiring review, the context is everything that system is allowed to know. It includes the documents pulled at query time, the rules for which documents get pulled, the memory the system carries from prior interactions, and which source wins when documents conflict. Plus things the system has been told to ignore.
None of that is the prompt. The prompt is the part someone can touch. The context is everything the institution loaded behind the scenes.
In a productivity tool, this gap barely matters. The cost of a poorly curated context is a worse answer or a missed source. For a system evaluating your insurance coverage or determining your benefits eligibility, the context architecture is where the decision gets made. It determines which criteria the system treats as evidence, which sources it consults, and what record, if any, exists afterward.
Curation discipline runs both directions. Missing the right source can break the answer. So can stuffing the system with too much material: long, noisy context can bury the signal even when the right information is technically present. The institution that loads every relevant policy rather than the right ones can make determinations less reliable than one with a well-curated smaller set. Neither failure is visible to the person receiving the result.
The institution chose before you arrived
Whoever controls the context controls the decision. The prompt may accept questions phrased a hundred different ways, but the control sits upstream. Before the system ever returns an answer, the institution decided:
- which criteria count as evidence for the determination
- which documents get retrieved when your case is evaluated
- which source wins when the criteria conflict
- what the system is forbidden to consider
- what record of the decision the system produces
You see the interface. The institution built the engine. This isn’t a criticism of AI specifically. Every automated decision system works this way. Rules-based systems can be easier to audit when criteria and logs are preserved, but even transparent logic can harm people at scale when no individual review happens before enforcement. AI context pipelines add another layer: retrieval, ranking, memory, exclusions, and outputs that can vary even on the same input. Often none of that is visible to the institution’s own reviewers after the fact.
What follows for accountability
If you can’t see the context, you can’t evaluate the decision. A reviewer looking at an AI-generated determination sees the output. They don’t see which documents were retrieved, which sources ranked highest, or what the system was and wasn’t allowed to consider. The review is present on paper. Without access to the context, it’s absent in practice.
A VentureBeat report on Pinecone’s Nexus launch described the problem this way: “Run the same task twice against the same data, and an agent may return different answers with no record of which sources drove either result. For enterprises where auditability is a compliance requirement, that is a structural disqualifier, not a tuning problem.” When the pipeline is non-deterministic and decisions aren’t logged, no audit trail exists to produce - regardless of how many humans are nominally in the loop.
The problem compounds across sessions. A system tested in short controlled runs may pass review comfortably, with answers looking good at iteration three. The same system in production can get messier as a case stretches across handoffs, notes, appeals, and old records. The institution’s confidence was calibrated on clean tests. The affected person late in a long-running case may be working with a noisier context than anyone in the test environment saw.
Colorado SB 26-189 (effective January 1, 2027) and the EU AI Act (key provisions for high-risk systems now targeted for December 2027, following a provisional May 2026 agreement) are moving toward requiring explanation, the right to appeal, and logging for AI decisions that affect people. Systems that can’t reconstruct which inputs and criteria drove a specific decision will struggle to meet that floor. The compliance gap sits in the context architecture.
What the floor looks like
The floor is two things: an explanation of which inputs and criteria drove the conclusion, and a way to challenge it with enough information to act. Both require the system to have produced a decision record at the time of the determination. A system not designed to log which retrieved documents drove which decision can’t produce that record after the fact, no matter how it’s prompted or which model runs underneath. The prompt is the surface the user touches. Context determines whether the system’s decisions can be accounted for at all.
One steady action to take this week
Ask this about any AI system that shapes consequential decisions: what record does it produce at the time of the determination, and who can see it?
Action ladder
Short term
If you build AI systems:Take one workflow where decisions affect people and write down four things: what does the system know and where does it come from? What’s logged at decision time? Can the affected person see which inputs and criteria drove the result? How do they challenge it? If you can’t answer those, the system isn’t ready.If you use services shaped by AI:When an automated system issues a denial or adverse determination, ask for the written explanation. Note the contestation window immediately. Those windows are often shorter than the time needed to gather documentation.
Medium term
Builders and deployers:Build toward traceable context: source-linked records and a decision trail a reviewer can read. Treat unlogged non-determinism as a design constraint, not a known limitation.Procurement and compliance teams:Add one question to every AI vendor review: can this system produce, at the time of decision, a source-linked record of which inputs and criteria drove the result? If the answer is no, the system may struggle to meet the accountability standards that are coming.
Long term
Institutions and policymakers:Colorado SB 26-189 (January 1, 2027) and the EU AI Act (key provisions for high-risk systems now targeted for December 2027) are building toward explanation, appeal, and audit requirements. Support legislation that makes decision-record production a condition of deployment, not a reporting layer added after the fact.Anyone paying attention:The regulatory floor is being built state by state. The gaps are where accountability failures accumulate. Track enforcement, not just passage.
How to talk about it
Any interface the person can see is the prompt layer. Underneath, someone decided what the system knows, which sources it trusts, what criteria it applies, and whether any of that is logged. That’s the context layer. When a system denies your claim or flags your application, the context is doing the work.
A common misframe: “We have a human reviewer in the loop.” Reviewers can only evaluate what they can see. If those decisions aren’t logged, the reviewer sees an answer with no explanation of where it came from: which documents were retrieved, which sources ranked highest, which criteria applied. The review is nominally present and functionally absent.
A useful framing: the interface is what people can see. The context architecture is what the institution controls. In most consequential AI decisions, the affected person didn’t configure the interface and may not have touched it at all. They received an output. Whether that output can be explained, challenged, or corrected depends entirely on decisions the institution made before the determination ran.