How to find the right AI use cases

Two failure modes

Ask a room full of people "where could AI help?" and you get one of two results.

The first is fifty sticky notes and nothing decided. Every workshop produces a wall of ideas — chatbots, dashboards, "something with AI in the customer portal" — and none of them get picked because there's no way to compare a customer service idea against a supply chain idea against a reporting idea. Energy goes in, nothing comes out, and the organisation concludes that AI strategy is just noise.

The second is quieter and more expensive: nothing happens at all. Leadership is waiting for the big idea, the transformative use case that justifies a programme. Meanwhile, the people doing the actual work have already started solving their own problems — someone in finance is running invoices through a free chatbot, someone in support has built a spreadsheet macro that half-works. It's not strategy. It's shadow IT with a language model attached, and it grows precisely because nobody above them is offering a better route.

Both traps come from the same mistake: picking one direction and trusting it completely. Pure top-down gives you ambition with no grounding — ideas that sound good in a board deck and fall apart on contact with a real process. Pure bottom-up gives you grounding with no ambition — useful fixes to small problems, none of which add up to anything the business will notice in a year's time.

The combined approach

The fix isn't choosing a side. It's running three modes together, each one correcting the others' blind spot.

Top-down sets the frame. Leadership names the outcomes AI is actually there to move — cost, revenue, risk, customer experience — and rules out anything that doesn't move one of them. This isn't about dictating solutions. It's about giving everyone downstream a test: does this idea touch an outcome we've agreed matters? If not, it's off the list, however clever it is.

Bottom-up fills the frame. The people closest to the work know which processes are painful, which decisions are made on guesswork, which reports get generated and never read. Give them a simple, always-open way to submit and score ideas — not a one-off workshop that vanishes into a slide deck, but a standing channel. This is where the real detail lives, and leadership rarely has it.

Middle-out filters and sequences. A small cross-functional group — not the board, not a single department — scores every idea against the same criteria: impact, feasibility, data readiness, strategic fit. This is the layer that stops good frontline ideas from dying in a queue and stops fashionable leadership ideas from skipping the queue.

A worked example

Say the outcome leadership names is "reduce cost to serve." On its own, top-down might land on something generic: "deploy an AI chatbot." Reasonable-sounding, hard to argue with, and likely to be built for a use case nobody has properly checked.

Bottom-up, on its own, surfaces something different: the support team flags that a huge share of calls are about order status, and agents spend most of that time manually cross-referencing three systems that don't talk to each other. Useful to know — but on its own, it's a ticket in someone's backlog, not a strategic priority.

Put the two together and a third idea appears that neither mode would have found alone: an AI-assisted lookup tool that pulls order status across all three systems automatically, surfaced directly to agents mid-call. It satisfies the top-down test (it reduces cost to serve, measurably, by cutting average handling time). It's built on bottom-up detail no leadership team would have guessed at unannounced. And it only exists because someone in the middle noticed the overlap between the outcome and the pain point. Neither the boardroom idea nor the frontline complaint would have produced it in isolation.

What good looks like

A written outcomes statement everyone in the organisation can quote back, not a phrase buried in a strategy document nobody opens. An idea intake that's open all the time, not a workshop that happens once and is never repeated. The same scoring criteria applied to every idea, no exceptions for whoever's sponsoring it. And a shortlist — five to eight funded initiatives a year, not fifty half-explored ones.

Pitfalls to watch for

Scoring criteria that bend under pressure. The moment a senior sponsor's pet idea gets waved through without scoring, the whole system loses credibility and the submissions dry up.

An intake channel with no visible output. If people submit ideas and hear nothing back, they stop submitting. Triage needs to be fast and visible, even when the answer is no.

Confusing volume with progress. Fifty ideas scored is not the same as five delivered. The point of scoring is to say no to most things, deliberately.

Next steps

1. Write down the two or three outcomes AI is meant to move this year — and get leadership to agree them in one sentence each.
2. Open a standing channel for ideas from anyone in the organisation, not a one-time call for submissions.
3. Agree scoring criteria before a single idea is scored against them, so nobody can retrofit the criteria to fit their favourite.
4. Put a small cross-functional group in charge of triage, and give them a cadence — weekly or fortnightly, not annual.

Where Florence helps

Florence is built around this exact shape. Any member of an organisation can submit an idea directly into their strategy. Editors triage submissions rather than letting them pile up unseen. And every project — whether it started with the board or with someone on the shop floor — is scored against the same criteria, so the shortlist that comes out the other end reflects genuine impact and feasibility, not who was loudest in the room.