The 2027 context layer: how Fortune 500 companies teach AI agents how the business actually works

An AI agent can know IFRS 15 by heart and still post the wrong journal entry. It knows the standard. It does not know your revenue recognition configuration, your exceptions, or the value above which a human must sign.

That gap, between how a process is supposed to work and how this company actually runs it, is where most agent programmes stall. The fix taking shape in the largest enterprises is not a smarter model. It is a context layer: a governed description of the business that sits between raw data and the agent.

This is Part 1 of our 2027 agent stack series. It builds on our 2027 enterprise AI readiness checklist, which ranked agent authority and orchestration as priorities, and on the 2027 AI budget guide, which named integration and data readiness as the largest one-off cost. The context layer is where much of that money goes.

Questions this article answers

Why isn't giving AI agents access to enterprise data enough?

Because the same entity has different names across systems, and the same term means different things in different companies. More data without shared meaning makes agents confidently wrong. See why data access stalls.

What is a context layer for AI agents?

A governed layer of five parts: reconciled entities, relationships, policies and thresholds, real process variants, and controls with audit. It tells an agent how this particular business works. See the five parts.

How are Fortune 500 companies preparing agents to act, not just answer?

JPMorgan built the platform first and then added proprietary data. Walmart organised its agents around the people they serve. Both depend on shared business context underneath. See what they are building.

1. Why data access is not enough

The first wave of agent work assumed access was the hard problem. Connect the agent to the ERP and the CRM, and it would work out the rest. It doesn't, for two reasons.

The same thing has many names. One system records a supplier as IBM, another as International Business Machines, a third as IBM Inc. An agent with access to all three sees three suppliers.

The same word has different meanings. "Critical supplier" in one company is not "critical supplier" in another. Even inside one firm, "active customer" can mean a purchase in the last 90 days to the sales team and a live contract to finance. An agent that reads the label without the definition acts confidently on the wrong meaning.

2. The five parts of a context layer

Think of it as the briefing a new senior hire gets in their first month, written down so a machine can use it.

Reconciled entities. One trusted record for each customer, supplier, product, contract and account, with rules for which source wins when systems disagree.

Relationships. Which invoice belongs to which contract, and which account rolls up to which parent company. This is where knowledge graphs earn their place.

Policies and thresholds. What the agent may do, up to what value, and what needs approval, written as rules rather than buried in a prompt.

Real process variants. How work actually runs, measured rather than assumed. SAP's product chief describes customers who believe they run one procure-to-pay process being shown by process mining that they run 400.

Controls and audit. Approvals, access rights and a record of every action, attached to the work rather than added afterwards.

None of this is new. Master data management and process mining have existed for years. What is new is who relies on them. People made up for messy context with judgement. Agents cannot, and they act at machine speed.

3. What Fortune 500 firms are building

JPMorgan: platform first, then proprietary context. When JPMorgan Chase rolled out its internal LLM Suite in 2024, chief data and analytics officer Teresa Heitsenrether described two stages: making models available to staff, then adding the bank's own data. In August 2024 she said the second stage had only just begun.

By its February 2026 company update, the bank reported about 150,000 weekly users, who estimate they save around four hours a week. Those are self-reported savings, not measured P&L, which is the realisation-rate point from our budget guide. Chief analytics officer Derek Waldron has described the remaining "value gap" as largely an integration problem across thousands of internal systems. That is the context problem, described by the bank that has spent the most on solving it.

Walmart: organise agents around who they serve. In July 2025 Walmart announced four "super agents" for shoppers, employees, suppliers and sellers, and developers. Global CTO Suresh Kumar explained that specialised agents sit beneath each one, such as an agent that helps employees with benefits and one that helps merchants analyse sales trends. The front door is organised around users and their tasks, not around which system holds the data. That works only if every agent underneath shares one understanding of products, stores, suppliers and orders.

SAP describes the same shift for its own agents, from recommending a payment to making and reconciling it under strict policy with a person in the loop, which it calls governed execution.

4. Keeping context current and portable

A context layer that was right in March can be wrong by June. Three habits keep it honest:

• Reconcile records continuously, with explicit rules on source trust and recency.
• Watch how processes run, so new variants surface before an agent trips on one.
• Version policies like code, so you can show which rule an agent followed on a given date.

Context also has to travel. Agents increasingly work across systems through protocols such as MCP, but a protocol moves messages, not meaning. Keep your definitions and policies in a layer you can export and audit, not only inside one vendor's product. That is the strategic point behind the 2027 harness wars.

5. What this means across industries

Every established business already holds the raw material for a context layer. It is just rarely written for a machine. A manufacturer has bills of materials and quality tolerances. A retailer has pricing, promotion and returns rules. A hospital has clinical protocols and consent requirements. A bank has credit policy and delegated authority limits.

Regulated sectors have an early advantage because supervisors already make them document policy, authority and audit trails. Less regulated businesses often hold the same knowledge in the heads of experienced staff, which is harder to capture and easier to lose.

The trap is the same everywhere. These assets live in PDFs, SOPs, committee minutes and inboxes, and an agent cannot read a delegation of authority from a 2019 board pack. The 2027 task is turning documented policy into machine-readable context, with a named owner for each piece. In our CLEAR framework, that ownership is the accountability test.

Your context test on one page

Score one production workflow against these five tests.

If the agent's understanding of your business lives in a system prompt nobody owns, the workflow is not ready to scale.

Part 2 of this series covers the control layer, the deterministic checks that sit above agents and models. For an outside view of your own workflows, see our AI Production Readiness & Governance Review, or request a strategy conversation.

The 2027 Agent Stack series: Part 1: the context layer · Part 2: the control layer · Part 3: the mid-market agent stack