Enterprise AI without the vendor lock: a 90-day program for sovereign capability

Vendor-licensed AI is convenient and structurally fragile

Most enterprise AI today runs through a vendor relationship: a SaaS license, a hosted model API, a partner that owns the platform. It works, until the vendor changes pricing, the model gets deprecated, the data sovereignty requirement shifts, or the strategic dependency becomes a liability. Then the question becomes: how much of our actual capability lives outside our walls?

Sovereign capability does not mean self-hosting every model. It means owning the operating model, the evaluation, the governance, and the platform — so that the model layer is replaceable without rebuilding the institution around it. The frontier-hosted API is an input, not the entire capability.

The 90-day enterprise AI program runs four parallel tracks

Ninety days is enough to stand up a real capability if the four tracks run in parallel rather than sequentially. Each track has its own lead, its own deliverables, and its own gates. They converge at week 13 into an integrated operating capability your team runs going forward.

Program length: 90 days fixed-fee
Tracks: 4 parallel ops / platform / model / team
Roadmap horizon: 18 months use-case backlog scored
Outcome: In-house platform + ops + team

Track one: the operating model defines who decides what

Without a defined operating model, AI work happens in pockets, with no shared evaluation policy, no shared governance, and no shared release process. Each team builds its own gateway, its own eval set, its own approval flow. Six months later there are seventeen versions of 'AI governance' and none of them are accreditor-defensible.

The operating model defines: who owns AI strategy, who owns AI engineering, who owns AI safety, who owns AI accreditation. RACI for model release, eval policy, incident response. Forum cadence — the AI council that meets weekly, the safety review that meets per release, the strategy review that meets quarterly. The plumbing is unglamorous and entirely the difference between a program and a hobby.

Track two: platform stand-up gives you the runway

The platform is the technical spine — model gateway, retrieval layer, evaluation harness, observability, and the deployment pipeline. Deployed in your environment, under your accreditation, with your identity and data controls. Once the platform is up, every subsequent AI use case lights up faster, with consistent governance, against the same gateway and the same eval framework.

Platform components stood up in the 90 days

Model gateway — single entry point for all model calls, with auth, rate limiting, cost attribution, and routing across providers.
Retrieval layer — hybrid search over your tenanted document corpora with access control enforced before retrieval.
Evaluation harness — graded eval sets, regression detection in CI, and a UI for the eval team that does not write code.
Observability — full request trace, reasoning replay, and integration with your existing OpenTelemetry stack.
Deployment pipeline — model and prompt promotion through dev, staging, and prod gates, with rollback in under a minute.

Track three: foundation-model strategy avoids single-vendor dependency

A defensible foundation-model strategy uses a mix: hosted frontier models for capability, open-weights models for cost-sensitive or sovereignty-sensitive workloads, and bespoke fine-tunes for the use cases where domain depth matters more than raw capability. The mix is determined by capability fit and risk posture, not vendor relationships.

The model gateway from track two is what makes this strategy operable. Use cases declare requirements — capability tier, cost ceiling, data residency, latency target — and the gateway routes to the appropriate model. Swapping a frontier provider becomes a configuration change, not a re-architecture.

Track four: team enablement is what makes the capability outlive us

We do not staff the program with our engineers and walk away. The engagement model is paired delivery from week one: your engineers, operators, and leadership working alongside ours, on the actual platform and use cases, against the actual operating model. By week 13, your team has shipped, evaluated, and operated AI in production, with us reviewing rather than driving.

The enablement deliverables include a curriculum tuned to your stack, hands-on labs against the platform we just stood up, paired on-call shadowing, and a set of demonstrated capabilities — your team has run a model release, debugged an eval regression, responded to a safety incident, and presented a use-case proposal to the AI council. None of that is theoretical at the closeout.

Eval and governance are how the program scales without breaking

Eval and governance are the same conversation. The eval harness is what governance enforces. The governance forum is what reviews what the eval surfaces. Skip either and the program produces deployments your safety team cannot defend or evaluations nobody acts on.

We deploy a release-gate eval suite — capability evals, safety evals, refusal evals, regression sets — that every model and prompt change passes before promotion to production. The governance forum reviews releases that fail any gate, signs off on exceptions, and updates the eval suite when a new failure mode is discovered. The accreditor-facing documentation lives in the same system.

The 18-month roadmap is what comes out of week 13

The closeout deliverable is not 'platform live' — it is a written 18-month capability roadmap, scored use-case backlog, hiring plan, and budget model. Every quarter has a defined investment, a set of use cases to ship, and a set of capabilities to add to the platform. The customer's AI council owns it from day 91 forward.

Roadmaps without ownership rot. Roadmaps without budgets are aspirational. Roadmaps without evaluation criteria become marketing. The Enterprise AI In-House program produces all three, integrated, signed off by the accountable executive at closeout.

The week-12 readout was the first time the CIO, the CISO, the CFO, and the head of engineering were all looking at the same AI roadmap with the same numbers. Everyone had been running their own. The platform was the artifact; the alignment was the actual deliverable.
— Chief Digital Officer, Fortune 500 industrial

What the program does not do

The 90 days does not solve every AI use case in the enterprise. It stands up the capability — operating model, platform, models, team — and ships the highest-leverage initial use case as proof. Everything else is on the 18-month roadmap, executed by your team against the platform we built together. We come back for specific engagements when it makes sense, but the steady state is yours.

It also does not bypass your accreditation, security, or procurement processes. Those processes start week one. The program is paced so the platform stand-up arrives at production-ready alongside accreditation sign-off, not before.

When 90 days is the wrong unit

If your environment requires accreditation that takes six months on its own, or you are pre-cloud, or your data foundation is not yet in a state where AI can do meaningful work, 90 days is the wrong unit. We say so in discovery and propose a foundation engagement first. The program works because the prerequisites are honest, not because we ignore them.

Frequently asked

What is sovereign AI?

Sovereign AI is the practice of owning the operating model, platform, evaluation, and governance for AI capability inside your organization, so that the model layer is replaceable without rebuilding the institution around it. It does not mean self-hosting every model; it means the strategic capability lives inside your walls, with hosted frontier models as one input rather than the entire stack.

What does the 90-day enterprise AI program actually deliver?

Four parallel tracks converge at week 13 into a running platform in your environment, a defined operating model with named owners and forum cadence, a defensible foundation-model strategy across hosted, open-weights, and fine-tuned models, a trained team that has shipped and operated AI in production, and an 18-month capability roadmap with a scored use-case backlog and budget model. Everything is owned and operated by your team going forward.

Why parallel tracks instead of sequential phases?

Because 90 days is not enough time to run operating model, platform, model strategy, and team enablement sequentially. Sequential delivery would push the team enablement track to month four or five, which is exactly when most programs stall — vendors leave, the capability lives only in their heads, and the customer's team never operationalizes it. Parallel tracks force enablement to start week one.

How is this different from hiring a vendor for an AI project?

A vendor project produces a deliverable; this program produces a capability. The deliverables of a vendor project depend on the vendor to operate going forward. The deliverables of this program — platform, operating model, team, roadmap — are designed to operate without the vendor. We pair-build with your engineers, your operators, and your leadership, and we are explicitly not the long-term operators.

What does the foundation-model strategy include?

A defensible mix of hosted frontier models for high-capability use cases, open-weights models self-hosted for cost-sensitive or sovereignty-sensitive workloads, and fine-tuned domain models for the use cases where depth beats raw capability. The model gateway makes the mix operationally swappable — a use case declares its requirements, and the gateway routes to the appropriate model. Single-vendor dependency is structurally avoided.

What does governance look like in this program?

Governance is a release-gate eval suite (capability, safety, refusal, regression sets) that every model and prompt change passes before promotion, plus a governance forum that reviews releases failing any gate, signs off on exceptions, and updates the eval suite when new failure modes surface. The accreditor-facing documentation generates from the same system. Eval and governance are the same conversation.

When is 90 days the wrong unit for this work?

When accreditation alone takes six months, when you are pre-cloud, or when your data foundation is not yet in a state where AI can do meaningful work. We say so in discovery and propose a foundation engagement first to address the prerequisite. The program works because the prerequisites are honest, not because we ignore them. Forcing 90 days against unmet prerequisites produces a capability nobody can use.