New playbook: when your own AI model pays. Read
An enterprise ontology for AI is a governed map of your business objects, metrics and decisions that people and agents read the same way.
It lets agents answer questions that need data from several company systems. This page covers the CEO’s decisions at this layer, the options, cost drivers, and what L3Harris, Hexion, Uniper, CVS Health and NVIDIA built in 2026.
An ontology names the things a business runs on — customers, products, orders, plants, contracts — and links them across systems. A semantic layer gives each metric one definition. Together they let an agent answer a question that crosses five systems.
Atlassian’s August 2026 shareholder letter put the stakes in one line: organisations can hire intelligence by the token, and context cannot be hired. Snowflake’s internal team wrote the same month that the bottleneck was never the data; it was context.
The CEO decides which objects and metrics get one company-wide definition and who owns each, what is recorded by default (sales calls, approvals, overrides), and what agents may read and write.
A decision log records each choice, its reason and its outcome. The NVIDIA and Palantir pattern, presented in September 2026, logs each planner’s choice, the reason for overriding the optimiser, the expectation at the time, and the outcome. The log is used for audits and later as model training data.
Most companies combine two or three of these.
| Option | What it gives agents | Fits when |
|---|---|---|
| Semantic layer over the warehouse | One definition per metric for analytics questions | Agents mainly answer numeric questions |
| Operational ontology | Linked objects with actions across ERPs and plants | Decisions span many systems and sites |
| Process-mining context | How work actually flows, step by step | Agents must find bottlenecks in a process |
| Decision and conversation records | Why people chose what they chose | You want to train or audit judgment |
The number of source systems and the disagreement between them drive most of the cost. Managers have to agree each definition; software only stores the result. Versioning and testing the semantic layer, as Snowflake does with dbt, keeps it correct as systems change.
Mollick argued in October 2026 that agents finding information themselves keep beating hand-built context pipelines. Give agents permissioned access to existing systems before building custom context pipelines.
L3Harris unified 33 ERPs on one data layer in Palantir Foundry. The company says scoping which programmes a glass-fibre shortage affects now takes about 10 minutes, against about three months and 30 people before (September 2026). The figure covers impact scoping only.
Hexion built one ontology linking customers, products, inventory, plants, raw materials, supplier contracts and cost. Hexion says that when a customer’s consumption spikes, agents expedite orders and revise schedules, and a sourcing agent produces scored purchase options for people to compare (September 2026).
Uniper grounded a hiring agent in process-mining data on how hiring actually flows. Celonis reports average time-to-sign fell by 27 days and screening by 13 days (September 2026).
CVS Health put a research agent on a governed semantic layer for its labour-demand forecast. Forecast error improved one point within two months, after the previous one-point gain took about a year; a CVS Health speaker estimated its value at $5–10M a year (August 2026).
NVIDIA trained a model on planner decision records. NVIDIA reported that on a development set it reached 86.7% accuracy on allocation decisions, against 55.5% for a large general model (September 2026).
List the twenty metrics and ten objects your leadership team argues about most. Assign one owner and one definition to each. Then pick one cross-system question that takes weeks today and make it answerable by an agent in minutes.
Other layers: Models and routing · Agents, traceability and evals · The learning loop
10 minutes
to scope a supply shortage, down from 3 months and 30 people
An enterprise data layer on Palantir Foundry unifying 33 ERPs (~3.5M data connections), feeding a Sector Control Tower for executives and a Program Digital Cockpit for program teams (source).
27 days
faster time-to-sign for candidates
A recruiting and hiring agent in Microsoft Copilot Studio, grounded in the Celonis Context Model (process-mining data on how hiring actually flows), used by recruiters and hiring managers (source).
2 months
for a 1% forecast-error gain that last took a year
A Genie-based research agent on Databricks that answers finance and store-operations questions about the labour-demand forecast, with explanations, replacing data scientists answering stakeholder emails (source).
86.7%
accuracy from a model trained on planners' decision records
A supply-chain allocation system where Foundry records each planner's final allocation, why they accepted or overrode the optimiser, what they expected, and what happened; those decision records train a specialised model (source).
Days → minutes
for fund-formation analysis and drafting, the firm says
The Fund Formation Engine on Palantir AIP and Ontology, bringing institutional knowledge, data and workflows for private-equity fundraising (documents, side letters, obligations, closings, compliance) into one system for every lawyer (source).
10 apps
on a legacy ERP in two months
Integrated data and ten operational apps on top of a legacy ERP with Palantir: BOM-to-cost explosion, order profitability, machine sequencing for work orders, purchasing-to-delivery links (source).
Thousands
of internal tools exposed to agents with zero service changes
An MCP Gateway and Registry that auto-crawls Uber's internal service definitions and exposes existing HTTP/gRPC APIs as agent tools, with tool-level authorisation and redaction for humans, services and agents (source).
25%
higher win rates in one division, vendor-reported
Consolidated a division assembled from four acquisitions around one AI-driven revenue platform (Gong, per Gong's release), starting from three decisions to improve: coaching, continuity and renewal risk (source).
A knowledge graph stores linked facts. An enterprise ontology adds governed definitions, owners and the actions each object allows. Agents need both meaning and permissions: what “active customer” means, and whether the agent may change an order. Hexion’s ontology, for example, lets agents carry a demand change into delivery plans and procurement options.
No. Start with the objects and metrics behind one valuable decision. Foundation Risk Partners cleaned one policy-level dataset before building two agentic systems; its owner Partners Group reported that the first use case went from concept to production in 14 weeks (August 2026).
Agents can only use calls, approvals and overrides that the company has recorded. NVIDIA reported in September 2026 that its model trained on planner decision records reached 86.7% accuracy on allocation decisions on a development set, against 55.5% for a large general model.
