Every AI vendor now sells delivery. Decide what your company keeps after the handover
Engineers trained by model providers, partner software paid from commitments, AI capacity shipped in weeks: delivery is becoming a product. La Madre's view on what to rent and what to own.
Listen to this article · 5 min
AI-generated narration of the full article.
Read last week’s announcements side by side and a pattern stands out. The companies that sell models, software and hardware are no longer only selling products. They are selling the delivery of AI into your organization.
Anthropic is spending $100 million to train Frontier Deployed Engineers who lead real projects inside their own organizations and partner firms. OpenAI’s new marketplace lets eligible enterprises pay partners from their existing commitment, which removes a procurement step. Lenovo now sells private AI capacity in tiers that ship in 15 to 25 business days, with deployment and ROI services attached. And AWS offers agents that investigate incidents and review architecture inside the customer’s estate.
Our reading, and it is a view rather than a fact: the bottleneck in enterprise AI has moved from access to models to the capacity to deliver them into production, and vendors are competing to own that capacity. We said in our September synthesis that the model is no longer the hard part. This is the market’s response.
Why this is good news, mostly
Delivery capacity is scarce, and more of it is welcome. A team that could not get a use case through security review, procurement and infrastructure in under a year now has more ways to do it in a quarter. The vendors’ interest is aligned with getting systems live.
The risk is in what happens after go-live. A vendor-trained engineer finishes the residency, a partner’s contract renews or does not, a hardware tier needs its first upgrade, and a vendor’s agent keeps operating inside your estate. Each of those moments asks the same question: who runs this now, and do they know how?
Rent the build. Keep six things.
- Implementation capacity
- Platforms and models
- Infrastructure tiers
- Choice of what to automate
- Evaluation sets
- Identity and the agent file
- Business definitions
- Runbooks and on-call
- The exit plan
- Rent or buy: Implementation capacity · Platforms and models · Infrastructure tiers
- Keep in house: Choice of what to automate · Evaluation sets · Identity and the agent file · Business definitions · Runbooks and on-call · The exit plan
Can come from vendors and partnersMust stay with the enterprise
- The choice of what to automate. Vendors will propose use cases that fit their product. The decision of which workflows deserve an agent, using tests like the five we proposed for agent-ready work, stays yours.
- Evaluation sets. The cases and scoring that define “good” for each task are the only way to compare vendors, approve model changes and catch regressions. If a partner owns them, you cannot leave.
- Identity and the agent file. Every production agent needs its own identity, sponsor, route, budget and record of actions, as we argued in our piece on the control layer. The vendor’s agent runs under your identity system, not the other way round.
- Business definitions. As the Genie One architecture shows for fund finance, governed semantics are what make agent output checkable. They outlive every platform.
- Runbooks and on-call. Someone inside must be able to diagnose a failing agent at 3 a.m. without opening a support ticket first.
- The exit plan. Where configurations, conversation data and evaluation history go if you change partner, platform or model.
How to buy delivery without losing the run
Put knowledge transfer in the contract, with named internal owners who shadow the build and lead the second release.
Make the handover package the definition of done: owner, runbook, evaluation set, access review, rollback plan. Pay the final milestone against it.
Keep one internal engineer per production system who could rebuild the integration if needed. That person, not the vendor, is the continuity plan.
Review vendor agents like employees with access. An operational agent from a vendor is still a non-human identity in your estate, with permissions that need scoping and review.
What would prove us wrong
If vendors start publishing durable, portable handover standards (evaluation sets, agent definitions and runbooks in formats another vendor can run), the case for keeping all six in house weakens. We have not seen that yet. Until we do, assume that whatever you do not own, you will rent forever.
What to do now
- List every production AI system and name the internal owner of each of the six items above.
- Find the gaps: systems where a partner or vendor holds the evaluation set, the runbook or the only working knowledge.
- Fix the next contract, not the last one: add knowledge transfer, handover criteria and data portability.
- Grow the small internal cohort that can run what vendors build. It does not need to be large; it needs to exist.
The bottom line
The market is finally investing in the hard part of enterprise AI, which is delivery. That is a reason to move faster, not to hand over the keys. Rent the build when it helps. Keep the decisions, the evidence, the identities, the meanings, the runbooks and the exit. Those are what make every vendor replaceable, including the good ones.