Why mid-market AI projects fail (and the 3 things that fix it)

I keep running into the same story about AI implementations at mid-market firms: started, paid for, never shipped. It shows up in postmortems, in write-ups, and in conversations with the operators who got left holding the result. The pattern is consistent enough to use as a diagnostic — ask three questions and you can usually predict whether an in-flight AI project will ship before the contract renews.

This post is the diagnostic, and the three things that fix the failure pattern.

The failure pattern

Mid-market AI projects fail in a specific shape. Not the way enterprise projects fail (governance, procurement, security review). Not the way startup projects fail (no users, ran out of money). The mid-market shape is:

The vendor delivered something. The team did not adopt it. Six months later, nobody is sure if it is still running.

Concretely:

• An AI chatbot got built and integrated into the help center. Three months later, the team that owns the help center is still answering tickets manually because the bot's answers were unreliable and they stopped trusting it.
• A document-processing automation got installed. It worked on the demo dataset. It does not work on the actual incoming documents because edge cases were not in the demo dataset.
• A "data assistant" tool got built. The dashboard is in someone's bookmarks. Nobody opens it. The original sponsor moved teams.

The vendor was not lying when they said the project was delivered. The project was delivered. It is just not producing value, because the missing layer was not the AI; it was the requirements work and the adoption work around it.

Why this happens at the mid-market specifically

Three reasons.

First, mid-market firms do not have a mature requirements function. Big enterprises have BAs and PMs whose job is to write rigorous requirements. Startups skip the requirements layer because the founder is the user and there are five users. Mid-market firms in the 50 to 500 employee range have neither: too big for the founder-as-user model, too small to have a BA staff. So vendors arrive into a requirements vacuum and write their own requirements, which do not reflect operational reality.

Second, mid-market firms are buying AI from vendors who do not understand the operational layer. Most AI vendors are technologists. They are good at building. They are less good at the AS-IS process documentation, the user research, the change management. So the AI gets built without the operational scaffolding that would let it stick.

Third, the buyer cannot tell the difference until 90 days in. A senior partner sells. A junior delivery team builds. The senior partner is no longer involved by the time the operational reality starts to bite. The junior team is not equipped to redirect the project. The buyer notices in month three that the velocity has died, but by then everyone is committed and the question becomes "how do we salvage this" not "what should we have done differently."

The three fixes

These are the three habits that move a mid-market AI project from "something got delivered" to "the team is using it."

1. AS-IS process documentation before any AI is specified

Every AI project should start with one or two weeks of pure process work. Sit with the operators who do the work today. Watch them do it. Write down the actual workflow, including the workarounds the official process does not cover. This document is the input to the AI design, not the org chart, not the ticket queue, not the vendor's prior work in similar industries.

If the vendor wants to skip this step because "we already understand your industry," the project is going to fail. The cost of two weeks of AS-IS work is trivial compared to the cost of a four-month implementation that nobody adopts.

2. Requirements that are observable, testable, and signed

Every implementation has a set of acceptance criteria that the buyer reviews and signs. Each criterion is something a non-technical user can verify by clicking. "The system handles edge cases gracefully" is not an acceptance criterion. "When a user uploads a CSV with mixed date formats in column B, the system displays a single warning and converts to ISO 8601 before processing" is.

The signed acceptance criteria are the contract. They are also the change-management script. When the project ships, you can hand operators a list of "the new system does these things in these specific ways" and they have something concrete to learn.

3. A 60-day adoption review with a redirect option

The contract should include a checkpoint at 60 days where the vendor and the buyer sit down and answer two questions: "is the system being used as designed" and "is it producing the value we expected." If the answer to either is no, there is a redirect option built into the agreement. The vendor adjusts the implementation, the buyer adjusts the workflow, or the engagement ends gracefully with documented learnings.

Most engagements do not have this checkpoint. The vendor declares delivery and walks away. The buyer is left holding a system that is not adopting and has no obvious recourse.

What this means for buyers

If you are evaluating an AI engagement right now, here are three diagnostic questions:

1. "Will you produce an AS-IS process document for the area we are targeting, before you propose what to build?"
2. "What does your acceptance-criteria template look like, and how do we sign it before development starts?"
3. "What happens at the 60-day mark if the system is not being adopted?"

If the vendor cannot answer these crisply, you are looking at a project that will follow the failure pattern.

What this means for how I work

I hold myself to the same three habits, because I have never seen the shortcut version end well.

The process documentation comes first, and it is not optional. The acceptance criteria get written and agreed before anything is built, rather than reverse-engineered afterwards from whatever shipped. And there is a point after delivery where the honest question gets asked out loud — is this actually being used, and is it doing what it was supposed to do — while there is still time to change direction.

If your AI implementation is in flight and you recognize the failure pattern, run the same diagnostic on it. Write down the AS-IS process for the area it touches. Put the acceptance criteria you thought you were buying next to what was actually delivered. Then be honest about which parts are salvageable and which need to be redone.