Before signing off on an artificial intelligence project, it pays to ask a plain question: what problem does it solve, and what does it cost to leave that problem alone? Many initiatives start from excitement about the technology and only later look for a use. They tend to end as impressive prototypes that never reach production. Assessing the investment properly from the outset avoids that trap.

Define the problem and the cost of inaction

Describe the problem in business terms, without mentioning models or tools. Who feels the pain? Where in the process does it appear? How often? Then estimate what doing nothing costs: hours spent on repetitive work, slower responses to customers, errors that force rework, opportunities that slip away. You do not need precision, but you do need an honest sense of scale.

If that cost is low, the project may not be a priority, however interesting it looks. If it is high, compare AI with simpler options: a fixed rule, an integration between systems or a change to how the work is organised.

Check the data you have

An AI solution is only as good as the data it can reach. Ask yourself:

does the data exist and can it be accessed, or is it scattered across spreadsheets and inboxes?
is it current and reliable?
are there legal or contractual limits on using it, especially where personal data is involved?
who is responsible for keeping it accurate over time?

Often the first useful step is to organise and connect the data. That work has value in its own right and prepares the ground for whatever comes next.

Agree success criteria up front

A project without success criteria set in advance ends up judged on impressions. Before any development starts, agree with your team how results will be measured: time to handle a request, the share of correct answers in a sample reviewed by people, satisfaction among those who use the tool. Set a minimum acceptable level too. If the pilot falls short, the decision to stop becomes easier and less personal.

Account for running and maintenance costs

Building is only part of the bill. An AI solution in production carries ongoing costs:

model usage and cloud infrastructure, which grow with volume;
quality monitoring, since behaviour can drift with new data or new model versions;
updates to instructions, integrations and business rules;
support for users and handling of exceptions.

Include these in the analysis from the start. A FinOps practice helps you track spending and avoid surprises.

Name the risks

List what could go wrong and how you will contain it: wrong answers delivered with confidence, leaks of sensitive information, heavy dependence on a single vendor, resistance to change within teams. For each risk, define a control, such as human review of high-impact decisions, limited permissions, audit logs or training for the people involved.

Decide whether to stop or scale

A pilot exists to produce information, not to prove the idea was right. At the end, compare the results with the agreed criteria and with the expected running cost. If the criteria are met and the cost is sustainable, widen the scope gradually. If you came close, adjust and try again on a narrower scope. If you fell well short, stop and record what you learned: stopping early is a sound decision too.

To see how we approach these projects, visit our AI agents page. And if you have a specific idea you would like to assess, get in touch for a diagnostic conversation.