Prioritising use cases

The typical meeting produces twelve AI ideas and zero owners. A month later none are running and all are still on the whiteboard. Prioritising is not voting for the trendiest idea. It is picking one you can measure and switch off.

The matrix we use fits on one sheet: volume (times per week), pain (what the error or the wait costs), data (does it already exist and can you use it?) and risk (legal, reputation, irreversibility). Score high / medium / low. The winner is high on volume and pain, medium or high on data, and low on risk.

If the most popular idea is high risk and low data, it is not a pilot. It is a later conversation, or a no.

The four columns, with examples

Volume. Extracting 80 invoices a week beats “a quarterly report”, even if the report impresses the committee. Pain. A VAT miss hurts more than a mis-tagged internal thread. Data. If PDFs already land in a mailbox, you have data. If “we would have to digitise the 2014 archive”, there is no pilot this month. Risk. An internal FAQ cites and can be corrected. A public bot that talks prices cannot be corrected in time.

Write a real example in each cell. Empty cells are ideas, not cases.

One pilot, not a portfolio

A portfolio is how consultancies justify months. An SME needs a flow someone runs the next Monday. The second case is earned by the first number. Open three and you will have three demos and zero habits.

Honest exception: a GDPR inventory can run in parallel because it blocks the others. That is not a second AI pilot. That is hygiene.

Who scores (and who does not)

Score the people who suffer the process and who pay for the miss, with a facilitator who stops everything being a five. Leadership alone tends to pick the visible case (the website) over the dirty one (invoices). The vendor tends to pick what their product does well.

One hour is enough. If the matrix takes three workshops, you are negotiating egos, not prioritising.

From matrix to pilot sentence

The winner is written in one sentence: “when X arrives, we do Y, a human confirms Z, we measure W”. If the sentence will not come out, the matrix lied: you still have a theme, not a case. Cut again.

That sentence is also the scope you stop from growing. Every “and it could also…” is parked for the next cycle, not smuggled into the first.

Budget and programme

Priority decides the budget, not the other way around. If the voucher pushes a category that did not win, either cut the category down to the winning case, or say AI is not this quarter. Forcing the losing case because “we must spend the voucher” is how tool graveyards start.

Three typical ideas, scored

Website chatbot: medium volume, low pain (the questions are already answered by phone), weak data (FAQ unwritten), high risk. It does not win. Invoice extraction: high volume, high pain, data in the mailbox, medium risk if there is a queue. It wins. Meeting summaries: medium volume, low pain, data in recordings with no clear basis, medium risk. Parked.

Do this out loud with your own ideas, with numbers. If nobody knows the volume, that is this week’s job: count, do not buy.

The matrix is not a deliverable for a drawer. It is the cover of the pilot. When someone proposes “and also…”, it goes underneath and waits for the first number.

What to do with ideas that do not win

They are written under the matrix, with a date. They are not deleted: that stops them coming back in disguise next week. At day 90, if the first flow measures, you score again. Some will have died on their own. Others will have gained data (there is a FAQ now, there is an export now).

Do not create a 40-row “AI backlog”. That is a graveyard with a kanban. Five lines under the matrix are enough.

If two ideas tie, the lower-risk one wins, not the showier one. A tie usually means you still have not put volume numbers in. Count a week and run the matrix again. The showy idea can wait; risk does not fall with enthusiasm.

Frequently asked questions

What if leadership wants the chatbot no matter what?

It can wait, or you ship a cited FAQ with no chat. If it wins on ego and loses on risk, we say so. The voucher does not require a bot.

How many ideas should we list?

Five or six real ones. Thirty is a brainstorm nobody will score honestly.

Does the matrix replace a maturity diagnosis?

No. If you are at level 1, the winner may be “write the process”, not AI. The matrix assumes there is something to prioritise.

Next: where to start and a realistic budget. Request an assessment with three ideas and volumes.