European Insurance Group. Thirty percent saved through engagement re scoping.

A European insurance group operating in more than twenty countries with roughly fifty thousand employees runs OpenAI across ChatGPT Enterprise, the API and fine tuning.

The engagement had been scoped ambitiously and priced accordingly. Re scoping it against what the workloads actually needed cut thirty percent.

Almost all of that came from routing rather than from negotiating.

The customer profile

Fifty thousand employees across more than twenty countries, running claims handling, underwriting support and customer correspondence at volume.

Insurance is unusually well suited to AI assistance because so much of the work is document shaped. It is also unusually varied in difficulty: summarising a routine claim note and assessing a complex liability question are not the same task.

Pricing them as though they were is where the money went.

The opening position

The scope routed effectively all workloads to the most capable model available, with fine tuning layered on top and a commitment sized to match.

That is a defensible engineering starting point and an expensive steady state. The most capable model costs several times the mid tier, and a large share of the volume did not need it.

The approach

We tiered the workloads. High complexity assessment work stayed on the most capable model. High volume routine summarisation moved down a tier, and simple classification moved down again.

We then tested the fine tuning case honestly. Fine tuning is valuable where behaviour must be consistent and specific, and it is an expensive way to solve problems that better prompting and retrieval already handle.

With the routing settled, the commitment could be sized to what the estate would actually consume rather than to its most expensive possible shape.

The eleven moves

1. Segment the workforce by role. Document, inbox and meeting intensity decide who is genuinely addressable.
2. Separate seats from API. Two purchases with different economics and different failure modes.
3. Normalize every quote to cost per task. Token rates across vendors are not comparable.
4. Tier workloads by model. Reserve the most capable model for the work that needs it.
5. Model caching and batching. Restructuring a workload often beats anything won at the table.
6. Commit on the proven baseline. Never on the roadmap, which moves faster than the contract.
7. Keep a second provider viable. A working integration, not a stated intention.
8. Get price protection in writing. Including what happens when list prices fall.
9. Cover model deprecation. Establish your entitlement when a committed model retires.
10. Negotiate rate limits alongside price. A cheap rate you cannot consume at peak is not cheap.
11. Put data terms in the contract. Retention, training use and residency, not a console setting.

The commercial outcome

• Saving. Thirty percent against the original scope.
• Routing. Workloads tiered by genuine complexity rather than defaulted to the top model.
• Fine tuning. Retained where it earns its cost and removed where prompting and retrieval suffice.
• Commitment. Sized to realistic consumption after routing.
• Data terms. Negotiated explicitly, as regulated workloads require.

Model routing is the single largest lever in enterprise AI cost, and it is an engineering decision rather than a procurement one. The commercial conversation cannot fix a routing problem.

How we engage

• Enterprise AI scoping. A six week engagement that segments the addressable population, separates seats from API, and normalizes vendor quotes to cost per task. GenAI vendor services practice.
• Negotiation. We run the seat and API conversations separately, with price protection, model deprecation cover and data terms on the table. OpenAI enterprise procurement playbook.
• Generative AI procurement. The same discipline across OpenAI, Anthropic, Google and AWS Bedrock. AI platform contract negotiation.
• Vendor Shield. Always on cover across the AI platforms and the wider software estate. Vendor Shield.
• Run the numbers. The software spend assessment sizes AI spend against the wider portfolio.