Make your company readable to agents
A company context layer for AI agents is one permissioned map of your systems, metrics, processes and decisions that every agent and every person reads from.
This playbook belongs to Data and ontology.
Why now
Atlassian wrote in its August 2026 shareholder letter: “Organizations can hire intelligence by the token. Context is much harder for organizations to build. And it cannot be hired.” Most companies use the same frontier models, so a company’s own context is what sets its agents apart.
The companies reporting results built this layer first. L3Harris unified 33 ERPs on Palantir Foundry. Scoping which programmes a glass shortage affects now takes about 10 minutes, against about three months and 30 people before (company statement, September 2026).
McKinsey’s June 2026 study of 15 AI-savvy companies found the same constraint. It found that the main constraint was knowledge held only in people’s heads.
10 min
to scope which programmes a shortage affects, from about 3 months (L3Harris)
27 days
shorter average time-to-sign in hiring (Uniper)
Pick three decisions the layer must serve
Begin with the decisions the layer will serve. A data catalogue built without a specific decision in mind keeps growing and is rarely used. Name three recurring decisions where people today chase information across systems: which orders are at risk this week, which accounts need attention, which programmes a supplier problem touches.
Each decision tells you which entities, metrics and events the layer needs first. Write down the questions an agent must answer for each one, in the words the decision owner uses.
For example, experian consolidated a division built from four acquisitions onto one revenue-intelligence platform, starting from three decisions to improve: coaching, continuity and renewal risk (vendor-reported, September 2026) (source).
The step produces three named decisions, each with an owner and a list of questions an agent must answer.
Inventory where the truth lives
For each decision, list the systems that hold the entities involved: customer, order, product, plant, supplier contract. Record who owns each system, how fresh its data is, and what an agent may read or write.
Most companies find the same entity defined in several ERPs, a CRM and a set of spreadsheets. L3Harris, built through more than 150 acquisitions, found 33 ERPs behind its programmes. The problem is largest in companies that grew through acquisitions.
The step produces a filled-in data and ontology inventory for the three decisions.
Agree one definition per metric
Agents fail quietly when “active customer” means three different things in three systems. A semantic layer gives every person and every agent the same meaning for each metric, with an owner and a version.
Snowflake runs its own curated semantic views, versioned and tested in CI through dbt (August 2026).
Start with the metrics in your board pack and your weekly operating review. Those are the metrics whose definitions people most often dispute.
The step produces a versioned metric glossary with one owner per metric.
Add how work actually flows
Tables show the current state of records, but agents also need the process: the steps, the handoffs and the places where work waits. Event logs from your ERP, CRM and ticketing systems already contain this.
Uniper grounded a hiring agent in Celonis process-mining data on how hiring actually flows. Average time-to-sign fell by 27 days and screening by 13 days (Celonis press release, September 2026). In a side-by-side test before production, the same prompts on raw HR-system data returned generic answers, while the process-grounded agent named Uniper-specific bottlenecks.
The step produces process maps for the three decisions, generated from event logs where they exist.
Expose your systems as permissioned tools
Give agents access through the interfaces you already run. Uber built an MCP gateway that crawls internal service definitions and exposes existing APIs as agent tools, with tool-level authorisation and redaction. Hundreds of servers and thousands of tools went live with zero changes to downstream services (October 2026).
Ethan Mollick made the same point in October 2026: pipelines that hand-feed AI the right context keep losing to agents that find information themselves. Agents need permissioned access to the systems where the work happens. The media company Every reached a similar conclusion when it replaced personal agents with one shared company agent and put its permission rules in code (October 2026).
The step produces a tool catalogue with read and write scope per agent role.
Record decisions as well as data
The most valuable context is why people decided what they did. NVIDIA, working with Palantir, records each planner’s final supply allocation, why they accepted or overrode the optimiser, what they expected and what happened.
A model trained on those records reached 86.7% accuracy on a back-tested development set, against 55.5% for a large general model (September 2026). Planners keep final authority. Start logging decision, reason, expectation and outcome for your three decisions now. The log serves as an audit record now, and it can be used as training data once there are enough entries.
The step produces a decision-log schema running inside the systems where the decisions are made.
Put one agent on each decision and measure it
Ship one agent per decision to the people who make it. Measure time to a sourced answer and the quality of answers, graded by the decision owner.
CVS Health put a research agent on its governed Databricks catalogue to answer finance and store-operations questions about its labour-demand forecast. Forecast error improved by one point within two months, against about a year for the previous one-point gain; the speaker values that at roughly $5–10M a year (speaker estimate, August 2026). Kirkland & Ellis says fund-formation work that took a lawyer days now takes minutes on its Palantir-based engine (August 2026).
The step produces one agent in use per decision, with a baseline and a current answer time.
Do this week
- Name three recurring decisions and the one person who owns each.
- Ask each owner to list every system they open to make that decision, and count them.
- Pick the five metrics that cause the most argument in operating reviews and write one definition for each.
Measure
| Metric | How to measure | Target direction |
|---|---|---|
| Time to a sourced answer | Median minutes from question to answer with source, sampled weekly per decision | Down |
| Metrics with one definition | Share of board-pack metrics with one owner and one versioned definition | Up |
| Agent answer accuracy | Owner-graded weekly sample of agent answers | Up |
| Systems reachable as tools | Share of inventoried systems agents can query with scoped permissions | Up |
| Decisions logged | Share of target decisions recorded with reason, expectation and outcome | Up |
Questions
What is a company context layer for AI agents?
It is the shared, permissioned description of your company that agents read before they act: the entities and how they connect, one definition per metric, how processes actually flow, which systems agents may call, and a log of past decisions with reasons. People use the same layer, so a person and an agent asking the same question get the same answer.
Do we need an ontology platform such as Palantir?
Several cases here run on Palantir, including L3Harris, Hexion and Kirkland & Ellis. Others use a semantic layer in Snowflake or Databricks, process mining from Celonis, or an MCP gateway over existing APIs, as Uber did. Pick the tool after the inventory in step two shows where your truth lives and which decisions you are serving first.
How is a context layer different from a data warehouse?
A warehouse stores data for analysts. A context layer adds what an agent needs to act correctly: agreed meanings for metrics, the process around the data, the tools it may call with their permissions, and the record of why past decisions were made. Most companies build the context layer on top of the warehouse they already run.
Who should own the context layer?
Give it one executive owner, usually the COO, with the head of data running the build. Each metric and each system in the inventory gets a named owner who approves changes. Without named owners, definitions drift within months and agents start giving different answers to the same question.
