Where to put AI first, if you have one quarter
Most AI projects fail because they start in the wrong place. A four-question filter that separates the processes worth automating from the ones that only look like it.
Most AI projects that fail don’t fail technically. They fail because they started in the wrong place — usually the most spectacular one, rather than the most boring one.
The pattern is consistent: someone in management sees a demo, the team picks the most visible process, and three months later the result is a tool two people use. Not because the AI didn’t work, but because the chosen process had too many exceptions to be worth automating.
The filter below has four questions. If a process doesn’t clear all four, it isn’t your first project — it can be your third.
- Pick the first AI project on frequency and repeatability, not on how impressive it looks in a meeting.
- You automate execution, not decisions. Consequential decisions stay with a human, with a validation point before the action.
- If you can’t write the rule in two sentences, the process has too many exceptions for a first project.
- A process with no baseline can’t demonstrate a result. Measure for two weeks before you build anything.
Four questions, in this order
- 01
Does it happen at least weekly?
Below that frequency the economics never close: build and maintenance effort exceeds the time saved. A quarterly process, however ugly, is cheaper done by hand. Frequency beats complexity.
- 02
Does the output look the same every time?
If the output is a quote, a summary, a record update, a classification — good candidate. If it differs every time and needs contextual judgement, AI can help with preparation but can't close the process.
- 03
Can you write the rule in two sentences?
The most brutal and most useful test. If it takes a page to explain when it applies and when it doesn't, the process has too many exceptions. That doesn't mean it can't be automated — it means it isn't first.
- 04
Do you have the before number?
How many requests a week, how many minutes each, who does them. Without the baseline, in three months you won't be able to say whether it worked, and the project becomes a matter of opinion — and opinions lose to budgets.
What clears the filter and what doesn't
| Process | Clears it? | Why |
|---|---|---|
| Answering repetitive customer questions | Yes | Frequent, similar output, the rule writes short, the volume is already measured. |
| Quoting on standard configurations | Yes | Repetitive with explicit rules. A human validates the final price before the quote goes out. |
| Filling in the CRM after a call | Yes | Daily, structured, boring. Exactly the profile that gets ignored because it impresses nobody. |
| Negotiating with a major account | No | Rare, high stakes, full of context. AI can prepare the material; the decision stays entirely human. |
| Annual strategy | No | Happens once a year and depends on judgement. It can never pay back the build. |
Three decisions for this week
Ask for three candidates, not one
Ask department heads which repetitive activity eats the most time each week. Ask for three answers, not one — the first answer is almost always the most visible, not the most suitable.
Measure for two weeks before you build
Volume, time per unit, who does it. Two weeks of measurement costs almost nothing and completely changes the conversation at the end of the project.
Decide where the validation point sits
Before the build, not after. What is the last action a human must approve? The answer to that question determines the architecture, not the other way around.
Questions we got on this edition
We have low volumes. Is it worth it?
It depends on the stakes per unit, not the volume. Ten high-impact decisions a week justify more work than a thousand inconsequential operations. What doesn't work is the combination of low volume and low stakes — automation never pays back there.
Why not start with something visible, like a chatbot?
You can, and sometimes it's the right answer — a chatbot on a site with real traffic clears the filter comfortably. The problem is choosing visibility over frequency: a spectacular tool for a process that happens twice a month looks good in the deck and dies in six months.
How long does a first project that clears the filter take?
In our experience, two to six weeks from decision to production for a well-chosen process. If the estimate runs past a quarter, that's usually not technical complexity — it's a sign the process wasn't defined clearly enough.
Let's run your processes through the filter.
A 30-minute conversation with your department heads in the room. We come out with three ranked candidates and the validation point settled for the first one.