How to implement AI in a company without burning the first attempt
A company's first AI project decides whether there will be a second one. When it fails, it rarely fails because of the model: it fails because of a badly chosen use case, inaccessible data, or because nobody agreed on what success meant.
- Published
- Updated
Why does the first AI pilot fail?
An AI pilot fails when the use case is picked for visibility rather than measurability: the case that impresses in a demo wins over the one with available data, describable rules and a known current cost.
The three reasons we see most often are always the same. First, the chosen process is undocumented, so there is nothing to compare the result against. Second, the data exists but is locked inside a system with no API, or in personal spreadsheets. Third, no one on the business side owns the outcome — only a sponsor who approves budget and disappears.
- Use case chosen for demo impact rather than the real cost of the process.
- Success criterion defined after seeing the first results.
- Scattered data with no owner and no access permissions resolved.
- No plan for the 20% of cases the system will not handle.
How do you choose the first use case?
The first use case has to meet four conditions at once: a repetitive process, data accessible today, a known current cost, and a tolerable error if the system gets it wrong.
A process that runs hundreds of times a month with stable rules is a better candidate than a strategic decision made twice a year. And a process where a mistake is caught at the next step is a better candidate than one where a mistake reaches the customer or the books.
| Condition | Good candidate | Poor candidate |
|---|---|---|
| Frequency | Daily or weekly | Annual |
| Data | In a system with an API or a queryable database | In email threads and personal spreadsheets |
| Current cost | Measurable in hours or errors | Vague, never calculated |
| Impact of an error | Caught at the next step | Reaches the customer or the books |
What does defining the success criterion upfront actually mean?
Defining the success criterion means writing down, before the first line of code, which number has to move, from what baseline, in what timeframe, and who measures it.
Without a baseline there is no result, only opinion. Before kickoff, measure how long the process takes today, how many cases arrive per week and how many are resolved incorrectly. That measurement takes a few days and is what later lets you say whether the project was worth it.
- A single metric: cycle time, resolution rate without human intervention, or errors per hundred cases.
- A baseline measured over the two preceding weeks.
- An explicit evaluation window, typically 6 to 12 weeks.
- A business owner who signs off on the criterion, not just the technology team.
How much human oversight is needed?
Any action with financial, legal or reputational impact needs human oversight, at least until there is accumulated evidence of how the system behaves in production.
The practical approach is to start in suggestion mode: the system proposes, a person approves. The log of approvals and corrections becomes the evaluation suite, and only with that evidence do you open the paths where the system acts on its own. Oversight is not distrust; it is how you collect the data that lets you automate more.
How long does it take to reach production?
A first, narrow use case usually reaches limited production between 4 and 8 weeks after kickoff, provided data access is resolved before the work starts.
In our experience the schedule breaks at the same point almost every time: permissions to internal systems. Request them during the diagnostic, in parallel with scoping, not when the team is already waiting to integrate.
- Weeks 1-2: diagnostic, baseline and access.
- Weeks 3-6: build, with deliveries every two weeks and evaluations from day one.
- Weeks 7-8: limited production with oversight and measurement against the baseline.
What is worth doing this week?
The cheapest next step is to list the three processes that consume the most hours and check, for each one, whether the data it needs is available today without prior development.
With that list and a two-week baseline, the conversation stops being about technology and starts being about the business. That is where an AI project starts to make sense, and stops depending on whichever model is fashionable this month.