Almost every company has tried AI by now. Teams have chat assistants, a few pilots and a slide about the future. Far fewer can point to a line in the budget that changed because of it.
The gap is rarely the technology. It sits in how AI projects are chosen, measured and carried into everyday work. This article walks through the approach we use with clients to move from experiments to results the finance team recognises.
Key takeaways
- Start from a costly business problem, not from a model or a tool.
- Agree the metric and the baseline before you build anything.
- Fix the data the use case depends on before scaling it.
- Design the workflow around people, with clear points where they review or override.
- Treat evaluation, monitoring and cost tracking as part of the product.
Why most AI pilots stall
Pilots usually start with a capability: "we should use generative AI" or "let's try an agent". The demo works, people are impressed, and then the project stops at the edge of real operations.
It stops because nobody owns the outcome, the data in production is messier than the demo data, and the people expected to use it were never part of the design. None of these are model problems, and a better model will not fix them.
Start with the problem, not the model
The best AI use cases are found in the places where the business already loses time or money: work that is repetitive, high-volume and rule-heavy, or decisions that depend on reading large amounts of information.
Good candidates share a few traits:
- The task happens often enough that small savings add up.
- Mistakes are visible and can be caught by a review step.
- The information needed to do the task already exists in your systems.
- There is a clear owner who will measure the result and push for adoption.
Define value before you build
Before writing code, agree how success will be measured and record today's numbers. Typical measures are hours saved per case, time to resolve a request, conversion rate, error rate or cost per transaction.
A baseline turns a debate about whether "the AI is good" into a simple comparison. It also tells you early when a use case is not worth scaling, which saves more money than any successful pilot.
Get your data ready
AI systems are only as reliable as the data behind them. For most business use cases that means documents, tickets, records and knowledge bases that are scattered, duplicated and out of date.
Data work does not need to be a year-long programme. Prepare the data for one use case at a time: find the sources, clean and structure them, set access rules and keep them current. Each use case then leaves behind assets the next one can reuse.
Design for people, not just the model
Adoption decides whether AI creates value. If the new tool adds steps, hides its reasoning or sits outside the systems people already use, it will be ignored.
Build AI into existing workflows, show the sources behind each answer and make it easy for people to correct the system. Their corrections become the most valuable training and evaluation data you have.
Measure, govern and scale
Once a use case is live, treat it like any other product. Track quality against a fixed set of test cases, watch cost per request, and monitor for drift as your data and users change.
Clear governance helps you scale with confidence: who can approve new use cases, how sensitive data is handled and how decisions are logged. With those foundations in place, each new AI project becomes faster and cheaper than the one before.
Where to begin
Pick one problem that is costly, frequent and owned by someone who wants it solved. Agree the metric, prepare the data it needs and ship a version people can use within weeks. Measure it honestly, then decide whether to scale it, change it or stop.
If you would like a second opinion on where AI can pay for itself in your business, our AI solutions advisor is a quick place to start, or you can talk to our engineers directly.
Written byHassan YounusLead AI & Data Science Engineer