Build or buy an AI operating system for your company

Contents

1. What the evidence says about building
2. What an AI operating system is made of
3. The four ways to buy
4. A decision table
5. Questions for the build-or-buy meeting
6. Questions buyers ask
7. Sources

Every established company with an IT team eventually asks whether it should build its own AI platform. The prototypes are easy and impressive. The question is what happens in the second year, when the models change, the auditors arrive and the engineer who built it moves on.

What the evidence says about building

MIT's 2025 study of enterprise generative AI, based on 150 leadership interviews, a survey of 350 employees and an analysis of 300 public deployments, found that purchased solutions and partnerships succeeded about 67% of the time, while internal builds succeeded about a third as often⁠1MIT report: 95% of generative AI pilots at companies are failingFortune · 18 Aug 2025. The same study found most pilots had no measurable effect on profit and loss at all.

The study does not say building is wrong. It says that most companies underestimate what a production system needs beyond the model.

What an AI operating system is made of

The parts a company would have to build and run

• A data layer that brings records from CRM, mail, finance and operations together, with clear ownership of each.
• Identity and permissions that decide what each person and each agent may read and change.
• Approval rules for actions that need a person's sign-off.
• An audit log of every AI action, retained and reviewable.
• Model access and routing, so the company can change models without rebuilding.
• The agents and interfaces staff actually use, and the support behind them.

Each part is a permanent cost, not a project. Models change quickly; regulations add requirements, such as the logging and human-oversight obligations the EU AI Act sets for high-risk uses⁠2AI Act: regulatory framework for AIEuropean Commission; and integrations break when other vendors change their software.

The four ways to buy

Add agents to the suite you have. Microsoft Copilot Studio lets a company build agents inside its Microsoft tenant, paid for in Copilot Credits⁠3Standard harness licensing: Microsoft Copilot StudioMicrosoft Learn · 4 Nov 2026. The company builds the agents but buys the platform, identity and compliance tooling.

Buy a governed layer across existing systems. Glean and Gemini Enterprise index or connect to the systems a company already has and enforce those systems' permissions⁠45SecurityGleanGemini for business: the universal work agentGoogle Cloud. Nothing is replaced; the layer is rented.

Buy a data platform and build on it. Palantir AIP provides builder tools on top of its Ontology, with audit tracking and lineage⁠6AIP overviewPalantir Documentation. It suits organisations with large engineering capacity and very large data estates.

Have a platform built that the company then owns. OpenLazarus designs and builds one platform per company in a yearly AI-Architect engagement, states that the code and data are the customer's, and says it trains the client's IT team to maintain the build without it⁠78FAQOpenLazarusAI implementation partner: systems you ownOpenLazarus. Hosting is managed but can be moved to the company's own cloud⁠9PricingOpenLazarus.

A decision table

Ask who will change the system when a model is retired, a regulation changes or a department reorganises. If the answer is a person who may leave, the build is a risk, not an asset.

Questions for the build-or-buy meeting

1. What exactly will we own at the end? Code, data, configuration, and the right to host it elsewhere.
2. Who maintains it in year two? Name the team and its cost.
3. How are permissions, approvals and audit handled? Built by us, provided by the vendor, or inherited from source systems.
4. How do we switch models? OpenLazarus routes to any model through its router⁠7FAQOpenLazarus; Glean chooses among 40+ models⁠10Glean product overviewGlean. A home-built system needs its own answer.
5. What does leaving cost? The cheapest platform to start is often the most expensive to exit.

The end state of a good decision is a platform the company can explain and change: it knows what it owns, who maintains it, how a new model gets in and how it would leave. For the governance side of that decision, see our guide to AI agent permissions and audit.

Questions buyers ask

Sources

1. MIT report: 95% of generative AI pilots at companies are failing. Fortune, 18 Aug 2025. Accessed 8 Nov 2026.
2. AI Act: regulatory framework for AI. European Commission. Accessed 8 Nov 2026.
3. Standard harness licensing: Microsoft Copilot Studio. Microsoft Learn, 4 Nov 2026. Accessed 8 Nov 2026.
4. Security. Glean. Accessed 8 Nov 2026.
5. Gemini for business: the universal work agent. Google Cloud. Accessed 8 Nov 2026.
6. AIP overview. Palantir Documentation. Accessed 8 Nov 2026.
7. FAQ. OpenLazarus. Accessed 8 Nov 2026.
8. AI implementation partner: systems you own. OpenLazarus. Accessed 8 Nov 2026.
9. Pricing. OpenLazarus. Accessed 8 Nov 2026.
10. Glean product overview. Glean. Accessed 8 Nov 2026.

Updated 8 Nov 2026: First published.

Further reading

• Enterprise guide · Updated quarterlyAI operating system for companies: the enterprise guide (2026)For a company of several hundred people, the hard part of an AI operating system is not the AI. It is deciding who the AI acts for, what it may touch, who signs off and where the record of it lives. We compared seven platforms on exactly that.
• Guide · GovernancePermissions, approvals and audit for company AI agentsThree controls decide whether an AI agent is safe to let loose on company records: what it can access, what it can do without asking, and whether anyone can see afterwards what it did.
• Guide · CostWhat an AI operating system costs a 200-person companyPer-seat, per-action and per-company pricing produce very different bills for the same headcount. We worked through each model with published list prices.