Importance of Customer Involvement in AI Projects

A common assumption in AI work is that a project becomes a real AI project once it is described in the right language. The team has engineers with AI-oriented job titles, the kickoff uses the vocabulary of modern machine learning, and the work is labeled as an AI engagement. Under that surface, however, the project may run like any ordinary consulting job with a few new words attached. This article follows one case in which exactly that happened, and it teaches a way to tell the two apart.

Both columns can share the same titles and vocabulary. The difference lies in the lower rows, the parts of the operating model the customer and the team either share or leave out. The rest of this article uses a single case to show how each missing part shows up in practice.

The engagement was a portfolio-narrative generation system: an agent that drafted commentary for a regional financial-services firm's clients about their portfolio performance. The deal was 620,000 dollars for a fourteen-week engagement. At week four the project was visibly stuck.

Rather than rely on a general feeling that something was wrong, an examiner put five concrete questions to the project. Each question checks one part of the operating model. The questions cover evaluation approval, production data access, internal alignment, stakeholder involvement, and customer-facing visibility.

The five questions form a diagnostic path. Ask them in order of a stalled project. A clear yes with evidence keeps the engagement on track. A no is a warning sign, and several no answers together signal a cargo cult engagement. The path below shows how each answer routes the examiner.

The data question most directly exposes an engagement that looks productive from outside. The production data is the path the finished system must travel, yet the request meant to open that path was stuck. The data-access request was filed in week one and did not move, and the vendor team kept building against synthetic data it had generated itself. Over four weeks the engagement accumulated work against inputs the deployed system would never receive.

Synthetic data can be useful for early development. The problem in this case was that it was the only data the team used, and no one had confirmed that it resembled what production would supply. Assuming that synthetic data is sufficient for a production-ready system is a misconception this case corrects.

The second part of the operating model concerns who takes part in the design. The customer's compliance team was the group that would have to sign off on any portfolio commentary sent to clients. It had been told about the engagement but had not been drawn into the design conversation. The vendor planned to bring compliance in at the end of week eight, once the system was already working.

The source notes that this plan could only lead to one of two outcomes: a six-week delay, or a forced redesign, depending on what compliance found. Compliance was not the only group left out of the design. The eval set, which determines how the system is judged, was built entirely by the vendor and never approved by the customer's sponsor.

Customer involvement is therefore not a courtesy added at the end of a project. The people who must accept the result have to shape what counts as success from the beginning. Otherwise the team is optimizing for a target the customer never chose.

The final diagnostic question asks whether the sponsor can see the project's progress. The engagement had a Slack channel, which is a form of communication, but no dashboard. The sponsor had not seen a metric since the kickoff meeting. Her anxiety was rising for a specific reason: when her CEO asked how the project was going, she had nothing concrete to point to.

Informal conversation does not substitute for a regular, visible account of progress. A channel can hold many messages while the person responsible for the decision still has no measurable picture of where the project stands. Without that picture, the sponsor cannot tell whether the engagement is on track, and she cannot explain its status to people above her. A customer-facing dashboard gives the customer a steady view of the work, which is why its absence counts as a warning sign.