An enterprise AI infrastructure deal runs in seven stages. It starts with discovery with the head of ML or AI platform. It ends with one use case live in production. The CTO or CIO signs.

What happens at each stage?

This is our working map for inference, model serving, vector databases, evaluation tools and GPU cloud, sold to buyers with a platform team. Up to the letter of intent:

  1. Discovery with the head of ML platform and a staff engineer. Exit: the champion agrees to a benchmark and names the budget owner.
  2. Technical scoping with platform engineers and a security architect. Exit: the champion signs off a benchmark plan with pass marks.
  3. Proof of concept, then limited production. Exit: pass marks met on live traffic for the agreed period.
  4. Letter of intent with the champion, the CTO or CIO, and procurement. Exit: you know the economic buyer and the decision date.

After the letter of intent:

  1. Commercials with the CTO or CIO, finance and procurement. Exit: finance accepts the spend forecast at two traffic levels.
  2. Security, data protection and legal review. Exit: every reviewer signs off in writing.
  3. Go-live for one use case or one team. Exit: the review is held and the second workload scoped.

What makes an enterprise start a deal?

A model must leave the pilot and carry real traffic, or a bill has outgrown its budget. In Futurum's 2H 2026 survey of 1,636 enterprise technology decision makers, 46.9% reported AI spend above budget (Futurum, September 2026). Of those, 47.6% asked for more money. The release does not give fieldwork dates.

Watch for four triggers:

  1. A pilot model moves to production and needs latency targets, uptime and on-call support.
  2. A hyperscaler or model API bill overruns, and nobody can forecast cost per request.
  3. A customer contract or a regulator requires data to stay in one country.
  4. EU AI Act obligations apply to the buyer's system.

AI Act duties for general-purpose AI model providers have applied since 2 August 2025. The Digital Omnibus on AI took effect on 27 July 2026 (AI Act Explorer, July 2026).

It moved stand-alone high-risk uses in Annex III to 2 December 2027 (European Commission, October 2026). AI in Annex I products moved to 2 August 2028. A buyer that builds a high-risk system on your platform will ask what logs and documentation you keep.

What should discovery and scoping pin down?

Discovery finds the number the champion is paid to move, such as p95 latency or cost per GPU hour. We qualify with MEDDPICC in engineering terms:

  • Economic buyer: the CTO, the CIO, or the business unit head who owns the use case.
  • Competition: usually an in-house build on Kubernetes, a hyperscaler service, or the current model API.

For the test plan, record the models, peak requests per second, tokens per request, context length, regions and deployment option.

How should the proof of concept run?

Run it on the buyer's own workload, with pass marks written before the first test. Measure four things against the current system:

  • p95 latency at the target concurrency, with real prompt lengths.
  • Throughput in requests or tokens per second before latency breaks the target.
  • Cost per million tokens at the buyer's token mix, or cost per query for retrieval.
  • Accuracy on the buyer's evaluation set.

After the sandbox passes, send shadow traffic, then a set share of live requests.

What goes in the letter of intent and the commercial terms?

The letter of intent records the use case, pass marks, timeline, signers, economic buyer and decision process. Its clauses mirror an MOU or LOI for an enterprise payments deal, with benchmark results in place of payment volumes. Offer two commercial shapes:

  • Pay as you go while traffic is unknown, with spend caps and alerts.
  • Committed spend or reserved capacity after a few months of real traffic, at a lower unit price with agreed overage rates.

The FinOps Foundation's 2026 survey of 1,192 practitioners found that 98% manage AI spend (FinOps Foundation, 2026). That was up from 63% in 2025. Bring a spend model at two traffic levels to the first finance call.

What do security, data protection and legal check?

Expect these requests on where data goes and what trains on it:

  • A SOC 2 Type 2 report and ISO/IEC 27001 certification. Regulated buyers ask more often for ISO/IEC 42001 (IEC Webstore).
  • A data processing agreement, the subprocessor list, and the regions where GPUs and storage run.
  • The models you serve, their licences, and a written term that customer data does not train shared models.
  • A private deployment in the buyer's VPC for workloads security will not send out.

Send the security pack in stage 2. A data protection review that starts after the benchmark adds its full length to the deal.

How do you go live and expand?

Go live with one use case, with a review date set before launch. At the review, compare production latency, cost and accuracy with the benchmark. The second workload sells on the first one's numbers.

In payments, an operations or treasury lead champions the deal (the stages of an enterprise stablecoin payments deal). For the full buying group, see who buys AI infrastructure inside an enterprise.

To see the ML platform leads and CTOs at companies with a live trigger, start with five named buyers, free.