The seller and the operator are never the same person
The firms that sell enterprise AI and the teams that operate it are usually not the same people, and their incentives point in opposite directions. The seller is solving for the engagement: scope, hours, the next statement of work, the renewal. The operator is solving for Tuesday. Does it work. Can we explain what it did. Who gets paged when it breaks at 2am. Most delivery is built by the first group for an audience that turns into the second group the day after go-live. And that handoff, the one nobody puts on a slide, is where enterprise AI quietly dies. The pilot that demos clean and then never reaches a second use case is not rare. It is the default outcome when the people building it never had to carry it.
You learn different things on the hook
There is a set of lessons you only get by being on call for something. Drift you did not catch because there was never an eval baseline to drift from. An “agent” with no identity of its own, so when a record changed at 3am nobody could say which system or which run touched it. A model swap that quietly changed behavior because nobody treated a prompt edit as a release. A governance story that turned out to be a PDF instead of a control you could actually enforce. Our team came up inside regulated work, healthcare among it, where “it passed the demo” means nothing and “it survived the audit” means everything. So the things a seller treats as phase two are the things we put in scope first: evidence, approvals, the handoff, the runbook nobody opens until something is on fire. These are not abstractions to us. Most of the briefings we publish are about exactly these failures: deploy provenance you can replay, identities for the agents that act, eval baselines, the review roles that decide who approves what. That is what the team actually argues about over Slack.
Size stopped being the proof
Here is the part I will state plainly, because it is the real reason this company exists. Three engineers with the right tooling now beat the 200-person program. Not because of heroics. Because output per person went up by an order of magnitude and the cost of coordinating a large team did not fall by a single dollar. A big firm’s headcount used to be the proof it could deliver. You needed the bodies. Now the headcount is mostly the reason the bill is large and the date keeps moving right. I have watched a small senior team ship the thing while the big program was still aligning stakeholders on the charter. I will be honest that this is uncomfortable to say, because people matter and I believe that. But as a technologist I have to look at what the technology can actually do, and what it can do is make a small team faster than a slow one was ever going to be.
What to make a delivery partner prove
If you are buying, stop grading the deck. The deck is the easy part and everyone’s deck is good now. Make them build something on your data inside the first week or two, not promise it in a future phase. When they say the work is governed, ask to see the control running, not the policy that describes it; a screenshot of an audit log beats a paragraph about compliance. Put your own team in the room from the start, because the people who will run it should help design it, and a partner who keeps your team out is telling you something. Then ask the question that ends most sales calls: write down, in plain language, how we run this after you leave. The runbook, the training, the decisions explained well enough that the next change does not require a new contract. The whole point of the work is that you need us less over time, not more. That is a strange thing for a services firm to optimize for. We think it is the only honest one.
Key takeaways
- Operating experience is the scarce input in enterprise AI delivery, and most firms selling it have never had it.
- Size is no longer proof of capability; a small senior team with the right tooling now outpaces a large program.
- The real test of a delivery partner is what happens after launch, not how the pitch looked.