CompanyOS
The governed layer every AI workflow and agent reads from instead of guessing. Built inside the client's own stack, owned by the client, never sold as software.
Every AI project in a business starts from zero. That is the actual cost.
Rebuild the marketing workflow: someone re-explains the proposition, re-finds the pricing rules, re-assembles the customer data, re-invents the approval step. Then sales does it again. Then service does it again. Nothing compounds, so the fifth project costs what the first did and the business quietly concludes AI is expensive.
CompanyOS is the part that gets built once.
What the industry calls this
CompanyOS is our name for it. The category has several — context layer, ontology, semantic layer, enterprise knowledge graph, context OS — and they converge on the same thing: a shared, governed model of what a business's data actually means, built once and referenced consistently, rather than re-derived by every agent that touches it.
Palantir got there first and named it the Ontology: a business-aware representation of the whole data estate, working in real-world objects — Customer, Policy, Purchase Order, Field Service Ticket — with the relationships, the business logic and the permitted actions attached, and security wrapped round it. Commentators describe it, accurately, as an operating system for enterprise AI agents. Every major platform is now shipping its own version.
The problem is measured, and most firms have not fixed it
Constellation Research's Michael Ni put the stake plainly — "whoever controls runtime context controls the AI decision layer for enterprise data" — and was equally blunt about how far any one product gets a buyer: "vector memory isn't business meaning, business meaning isn't governance and governance isn't execution." That gap between the four is where the delivery work lives.
Who is actually delivering it
Three distinct tiers, and it matters which one a client is being sold.
| Tier | Who | What they actually give you |
|---|---|---|
| Platform-native context layers | Microsoft Fabric IQ (a business ontology any agent can query over MCP) · Snowflake Horizon Context + Cortex Sense · Oracle Unified Memory Core · Google Knowledge Catalog · AWS Context · Databricks Unity Catalog | The plumbing, inside a stack you probably already pay for. No two have converged on the same architecture, so expect to integrate rather than pick a winner. |
| Specialist context and graph vendors | Palantir Foundry/AIP · Glean (enterprise graph across documents, people, conversations) · Atlan · DataHub · Pinecone Nexus · Couchbase · Stardog, Timbr, dbt Semantic Layer, Galaxy at the lighter end | A purpose-built layer with governance and retrieval attached. Strong products; all of them still need someone to model your business into them. |
| The people who model it | Palantir's own forward-deployed engineers · the big SIs · boutique data consultancies · your own team, in every self-serve motion | The actual value. Analysts describe Palantir's forward-deployed engineering model as "a category of one" — which is a polite way of saying the software is the easy half. |
What it costs at the top of the market
Palantir publishes no commercial price list, but the UK government's G-Cloud 14 framework does. A Foundry term licence is listed from £66,000 per server core per year. A discovery package of up to three months runs £50,000–£250,000. Implementation and engineering services are billed separately at £150,000 per person per quarter. A single-organisation Foundry licence is listed at £3,000,000 a year. Palantir's own FY2025 10-K puts average revenue from its top twenty customers at $93.9m each.
At the other end, published mid-market pricing for a comparable connect-explain-act loop starts around €30,000 a year. The gap between those two numbers is not a quality gap. It is a gap in who does the modelling — and that is a services question, not a licensing one.
What is in it — five components
- The ontology, or canonical objects. One record per thing that matters — customer, product, person, project, contract, policy — mapped to the systems that hold it, with the relationships and the business rules attached. Today those live in four systems with three spellings and no join. This is the unglamorous majority of the work and the reason everything else is currently expensive.
- The context layer. The proposition, the segments, the tone, the approved claims, the pricing rules, the things nobody is allowed to say. Machine-readable, versioned, owned by the client's marketing director rather than trapped in a PDF and three people's heads. Analysts note most context platforms cover structured tables well and the messier unstructured material badly — this is the half that gets skipped.
- Governed retrieval. Any workflow or agent can ask the business a question and get an answer grounded in its own documents, with the same access rules the human would have had. The current standard pairs a knowledge graph with vector search rather than relying on document retrieval alone.
- The evaluation harness. A standing test set that tells you whether a change made the output better or worse. Without this, every AI improvement is an opinion and every rollback is a guess.
- The measurement spine. Baselines and KPIs for each installed workflow, reported the same way, so the programme can be defended at budget time rather than believed in.
The interface standard is settled — which is why this is buildable now
MCP (Model Context Protocol) is how agents connect to business systems, and the argument about it is over. Anthropic donated it to the Linux Foundation's Agentic AI Foundation in December 2025, making it vendor-neutral; every major client supports it — ChatGPT, Claude, Gemini, Microsoft Copilot — and Microsoft's Fabric IQ ontology is explicitly queryable by any agent over MCP, not just Microsoft's own. Gartner expects 75% of API gateway vendors to ship MCP features by end of 2026.
Practically, that means a context layer built properly today is not locked to whichever assistant the client standardises on next year. Two years ago that was not true, and it is the single biggest reason this work is worth doing now rather than in 2024.
How it gets implemented
Five stages, each one paid, each useful on its own, each able to stop without stranding the client. Nobody signs a two-year transformation programme on stage one. The industry's own lesson is the same: start with one narrow lighthouse use case, not an enterprise-wide modelling initiative — and the market has moved from "we will build you a custom twin in eighteen months" to "we stand up a first version in weeks and iterate".
Fit test
Paid, small, and genuinely allowed to conclude "not yet". We inventory the systems, the state of the data, the real workflows and the appetite, then come back with a map, a phased cost, and an honest answer on whether the join is worth building.
Output: the map, the recommendation, and the number the programme would be judged on.
The objects and the joins
Canonical records and the relationships between them, modelled against one narrow lighthouse use case rather than the whole business. In the client's existing tools wherever those will carry it — buy the commodity, build only the join.
Output: one customer record, one product record, one person record, and reporting that reconciles to finance.
The context layer
Proposition, segments, tone, claims, pricing rules and prohibitions written once in a form both people and models can use — including the unstructured material the platform vendors handle worst. Owned and edited by the client's marketing lead, versioned like code.
Output: the client's own brand and policy layer, exposed over MCP, with the first workflow drawing on it.
Workflows on top
The Install, industrialised. Each new workflow reuses the objects and the context layer, so the second costs a fraction of the first and the fifth is routine. This is where the curve either bends or the whole thesis is wrong.
Output: installed workflows, each with a baseline and a reported KPI.
Agents on top
Internal first, customer-facing only when the evaluation harness says it is safe. Grounded, scoped, with a measured escalation path — the direct answer to the 57% confident-wrong number above.
Output: agents that hold up in front of a customer. The standard →
They run it
The client's team owns the context layer, the eval set and the roadmap. We stay embedded as long as they want the capacity, not because they cannot operate it without us.
Output: documented ownership, trained operators, and no dependency we manufactured.
Specific things you can do with it
The abstraction is the enemy of this sale. Concretely, once the layer exists:
Marketing and commercial
- "What did customers say about our pricing change last quarter?" — answered in seconds from call transcripts, support tickets and reviews, with citations and respecting who is allowed to see what. Today that question takes a person three days and comes back as anecdote.
- A campaign brief assembled, not written from a blank page — pulling the ICP, the positioning, the approved claims, what the last three comparable campaigns cost and returned, and the objections that actually came up in sales calls.
- A proposal drafted with the right case studies already in it, selected by sector and deal size from the credentials library rather than by whichever partner remembers one. This is The Delivery Multiplier with the library underneath it.
- One customer record, so two teams stop contacting the same person — and so churn risk, open tickets, contract value and campaign membership are visible in one place instead of four.
- Answer-engine content generated from real customer language, with ratings markup attached — the output half of Ground Truth, running continuously rather than once a year.
- The monthly board report, assembled automatically against definitions the finance director already signed off. That is The Growth Model without the fortnight of analyst assembly.
Operations and risk
- Compliance evidence generated from queries rather than collected by hand. A documented case: SOC 2 and ISO 27001 audits that previously took 80+ hours of manual evidence collection each quarter, produced from graph queries instead.
- "Which customers are exposed if this supplier fails?" — a question that is trivial with a modelled object graph and effectively unanswerable without one.
- A new hire querying the business instead of interrupting the two people who know — the onboarding cost that nobody puts on a spreadsheet.
- Service failures ranked and routed from the same corpus marketing is mining for content, so ops and marketing stop arguing from different evidence.
Reported outcomes for enterprise knowledge-graph deployments cluster around 25–40% efficiency gains and roughly three times faster analytics development cycles, with much higher headline ROI figures circulating from vendors. Treat the headline numbers as directional and vendor-sourced; the 80-hours-a-quarter compliance example is the shape of claim worth repeating, because it is specific and checkable.
How we would actually do it
- In the client's tenancy, on the client's licences. Every asset lives in systems they already pay for and already trust. No E5 environment, no E5 login, nothing they cannot take to another supplier the day they want to.
- Buy 80%, model the join. The catalogue, the warehouse, the vector store, the model access and now the context platform are all commodities with real vendors behind them. What no vendor sells is your business modelled into them. That is the only thing worth our engineering time.
- MCP-first, so nothing is locked. The layer is exposed over an open standard governed by the Linux Foundation, not wired to one assistant. If the client changes model vendor, the layer survives.
- One lighthouse use case, not an enterprise modelling programme. The documented failure mode is breadth. We model the objects that the first workflow needs and no more.
- Embedded specialists, not a project team. The people who deliver it sit inside the client's business — which is the model, and also the only way to learn how the work is really done.
- Governance from stage zero. Data handling, permissions, retention and the approval path designed in at the start, not retrofitted when procurement asks. In regulated sectors this is the gate, not a nicety.
The objections, named first
| Objection | Answer |
|---|---|
| "This is a platform. You said you don't sell software." | Correct, and it is the sharpest tension in the offer. The difference is ownership and shape: built in their tenancy, on their licences, sold as an engagement with fixed scope per stage, no recurring licence to us, no product roadmap we control. If it ever becomes a thing we host and rent, we have broken our own model. |
| "You are not an enterprise-grade engineering firm." | Also fair, and the reason the doctrine is buy-the-commodity. We are not writing a database, a catalogue or a model. We do the modelling, the context design and the measurement — which is precisely the part Palantir bills at £150,000 per person per quarter and every self-serve vendor hands back to the customer. |
| "Why not just buy Fabric IQ / Snowflake / Glean?" | They should. Those are the components. None of them models your business for you, and no two have converged on one architecture — analysts expect buyers to integrate rather than pick a winner for several quarters yet. Our work sits on top of whichever they already own. |
| "Two-year programmes die." | Which is why it is not sold as one. Five stages, each independently useful, each with its own number, and an explicit right to stop. The market itself has moved from eighteen-month twins to a first version in weeks. |
| "Prove the demand first." | Agreed. No client should be sold CompanyOS cold. It is what a client asks for after The Install has worked twice and they want the third one faster. |
Who it is for
- Who buys it: an MD or CEO who has already seen one workflow work and is asking why the next one costs as much as the first. Or, increasingly, one whose agent gave a confident wrong answer in front of a customer. Rarely a first purchase.
- Size: big enough to have several systems that do not talk, small enough that no internal team is already doing this, and too small for the seven-figure floor. Roughly £10m–£200m turnover.
- How it is judged: time and cost to stand up workflow n+1 versus workflow n. If that curve is not falling, the layer is not real and we should say so.