A pilot on one process: how to test AI without a big budget

What the pilot is for

The pilot answers one question: does the approach work on your data and in your team? Testing that on one process is cheaper than on five at once.

• It limits the risk. If the hypothesis is not confirmed, you have lost weeks, not a quarter.
• It produces material. After the pilot you can see which data is missing and where the process is poorly described.
• It settles the argument. The “will it help or not” debate is replaced by a result you can look at.

How to choose the process

A good pilot candidate looks roughly the same in any company. For process ideas, see the list of 10 repetitive tasks you can automate with AI.

• It repeats often. Daily or several times a week — otherwise there is nothing to measure the effect on.
• It follows clear steps. Exceptions exist, but they can be listed.
• It has an owner. There is a person who can say whether an answer is right or not.
• It does not block the money. For a first pilot, better not to pick a process whose stoppage stops sales.

What stays inside the pilot — and what goes outside the brackets

Inside

One process, one channel, one interface language, a limited volume of data and one way of checking the result.

Outside

Integrations with every system, rare exceptions, interface design and company-wide training. That is the work of the implementation stage, not the test.

Why this matters

Every item added to the pilot “since we’re doing it anyway” stretches the timeline. A pilot that runs for three months stops being a pilot.

How to know the pilot succeeded

The criteria are agreed before the start, not after the results. Otherwise the discussion of the outcome turns into an argument about what we were expecting.

• The share of answers needing no edits. Counted by the process owner, manually, on a clear sample.
• Time per case. Measured before the pilot and after, without sophisticated analytics.
• Behaviour on the unknown. What the solution does when data is missing: admits it or makes something up.

Use your own numbers for comparison. Other people’s figures from presentations have nothing to do with your process.

What to do after the pilot

• Record the findings. Not just the result, but also the gaps in the data and in the process description.
• Decide the next step. Expand to the whole process, transfer the approach to a neighbouring one, or stop.
• Cost the operation. The cost of running permanently differs from the cost of the pilot — find it out before expanding.

Stopping after the pilot is a normal outcome. It cost weeks and saved you from an implementation that would have cost months.

Frequently asked questions