What should we automate with AI first?
There is pressure to show AI progress, several candidate use cases, and no principled way to choose between them.
Choose a workflow that is high volume and low variance, where you already have a number you could measure. High volume makes the impact visible quickly; low variance means the edge cases that break AI systems are rare. The instinct is to pick the interesting problem, but interesting means variable, and variance is exactly where AI fails publicly. Back-office and operations workflows consistently outperform the sales and marketing pilots most budgets chase.
The baseline is the part teams skip, and skipping it is why so many pilots end in an argument. Before deploying anything, record what the workflow costs in hours today and what its current error rate is. Without those two numbers there is no way to show the pilot worked, and a pilot that cannot be shown to work does not get funded into production however well it performs. Measure for two weeks first — the cheapest step in the programme, and the one that decides whether the next gets approved.
Why first AI projects are so often the wrong ones
- Demo appeal and production suitability are inversely correlated. The workflow that demos best is usually the one with the most interesting variation, which is the one most likely to fail at scale.
- Volume is confused with importance. An important quarterly process is a poor first candidate: too few instances to learn from, too visible when it goes wrong.
- Nobody baselines. Without a before number the pilot cannot be defended in a budget review, however well it worked.
How to pick the right first automation
-
01
List candidate workflows with two columns: how many times a week it runs, and how much each instance differs from the last.
-
02
Discard anything running fewer than a few dozen times a week. You need volume to learn and to show impact.
-
03
Discard the high-variance ones for now, however appealing. Come back when the mechanism is proven.
-
04
From what remains, pick the one where you can already state a number — time taken, error rate, cost per unit — and write down today's value.
-
05
Check the data behind it: duplicates, dormant records, free text that should be structured. If that fails, fix it before building anything.
The right first workflow is usually boring, and boring is hard to fund. Nobody presents 'we automated invoice coding' to a board excited about AI. But the boring workflow is what proves the mechanism, produces a defensible number, and earns the mandate for the interesting one. Organisations that start with the interesting problem generally get a stalled pilot and a harder second conversation.
More on choosing an AI starting point
Because failures are public and variance is highest where humans are involved. Prove the mechanism internally, then move outward. Most organisations that succeed customer-facing did something unglamorous first.
Enough that a week of operation produces a meaningful sample — usually a few dozen instances at minimum. Below that you cannot tell improvement from noise, which means you cannot defend the result.
Then data is the project. That is worth knowing before spending on AI tooling, and the cleanup has a return independent of AI: better reporting, better forecasting, less reconciliation.
More on AI Transformation
Still not sure this is your problem?
A 20-minute fit check. We will tell you if it is something you can fix without us — that happens often enough that we lead with it.
Book a Strategy Session →Browse every answer.
The full Solution Desk — specific questions, straight answers, no gate.
Open the Solution Desk →