From pilot to production: the checklist most teams don’t have

The reason most AI pilots stall is not a missing capability. It is a missing checklist. There is no agreed list of what has to be true before the feature ships, so the team discovers the requirements one at a time, usually after something breaks in front of a customer.

The stakes are visible in the industry numbers. Vendor partnerships reach production about 67% of the time, versus roughly one third for internal builds, a 2x gap. A large part of that gap is simply that experienced partners arrive with the checklist already written. Internal teams shipping AI for the first time have to derive it, and deriving it under production pressure is the hard way to learn.

Building rigor when nobody hands it to you

A founder and CEO of an agricultural-genomics company described what it is like to build a category with no reference to copy. His company reads soil health from the microbes in the ground using DNA sequencing, and early on there was simply no one to benchmark against. No prior art, no reference customer, no template. His team had to invent the rigor themselves, deciding what “good” even meant before they could claim to deliver it.

He made a distinction that maps cleanly onto AI production readiness. It was not enough to catalog what was in the soil, the taxonomy. What mattered was functional analysis, what the system actually does, so results could be compared across 56 countries and every crop. A checklist works the same way. Listing components is not the point. Defining what each one has to do, and how you will know it does it, is.

The lesson for AI teams is direct. When nobody hands you the production checklist, you do not get to skip it. You have to build it, deliberately, the way that founder built rigor into a field that had none. The teams that stall are the ones who assume the checklist does not exist because no one gave it to them.

He also had to work without reference customers, with no comparable company to point at and copy. That is the position most mid-market teams shipping their first AI feature are in, whether they admit it or not. There is no internal precedent for what production readiness looks like here, because the company has never done this before. That is precisely why the 2x partnership gap exists. Bringing in someone who has shipped this kind of feature before is a way to borrow a checklist that would otherwise take you a failed launch or two to write yourself.

The checklist, in plain terms

Before an AI feature ships, these should all have a clear answer. Most stalled pilots cannot answer half of them.

Data readiness. Does the feature run on production data, not a curated sample, and is that data clean enough?
Reliability target. What accuracy and uptime does this have to hit at real volume, stated as a number?
Ownership. Who is the named person accountable when it misbehaves?
Integration. Does it live inside the tools people already use, with the access and permissions that requires?
Monitoring. Will you know it is degrading before your users tell you?
Failure handling. When it gets something wrong, is the blast radius contained?

If any answer is “we will figure that out later,” that is the item most likely to strand the pilot.

How we approach it at Density Labs

When a team is staring at the leap to production without a map, we run the AI Readiness Assessment, our $2,500 engagement, and walk exactly this checklist against their pilot. We turn each vague area into a concrete yes, no, or here-is-the-work. The output is the list they did not have, written before the failures instead of after.

That founder built a whole field’s rigor from nothing. A single AI feature does not need a whole field, but it does need its list, and it needs it before launch.

You cannot check a list you never wrote.