How to Prioritise AI Opportunities When Every Department Wants a Copilot

The hard part of an AI strategy isn't finding ideas. It's saying no to nine of them. Every department has a wishlist, every wishlist is defensible, and none of them are the same size. This post is about how to score AI opportunities so the conversation moves from opinions to numbers.

The framework below is the same one we use inside an AI Audit. It scales down to a five-person team and up to a 500-person operation.

The problem with the usual approach

The typical way teams prioritise: the loudest department wins. Or the founder's favourite idea wins. Or the vendor who ran the best demo wins. None of those are wrong every time, but none of them are right often enough.

The second most common approach is a fake matrix. Impact on one axis, effort on the other. Six ideas plotted. Everyone agrees on the picture. Nothing ships, because "impact" and "effort" were never defined.

A useful matrix needs three properties:

• Every score is in a real unit (dollars, weeks, percentage points)
• Every score has a defensible source
• The score set fits on one page and one meeting

The three axes that matter

Impact — expressed in annualised dollars

Not "high, medium, low." A number. If a support workflow currently costs 400 hours of team time per month at $50 loaded, that's $240,000 a year. If AI can absorb 60% of it, that's $144,000 of impact.

Include downstream effects when they're real: fewer escalations, faster response times, reduced churn. Don't include vibes. If the number requires a footnote longer than the number, it's not a number yet.

Effort — expressed in build weeks and integration risk

Effort has two components. The first is how long the build will take — this maps to money. The second is how hard the integrations are, because a two-week build with a nasty legacy CRM ends up being a three-month project.

A useful shorthand:

• Small: 2–4 weeks, one integration, off-the-shelf model
• Medium: 4–8 weeks, 2–3 integrations, some custom logic
• Large: 8+ weeks, custom model or heavy integration, dedicated infra

Readiness — the axis nobody scores

This is where most portfolios collapse. Readiness is whether the process is stable enough to automate, whether the data is clean enough to feed a model, and whether the team owning the workflow will actually adopt the new tool.

If a process changes every month, automating it is a moving target. If the data lives in five spreadsheets nobody trusts, cleaning it becomes phase one. If the team using it is already at capacity, adoption stalls even when the tech works.

Readiness gets scored on three sub-axes:

• Process stability (has the workflow changed in the last 6 months?)
• Data quality (can we sample 100 records and grade them ourselves?)
• Adoption readiness (does the team owning this workflow have 2 hours a week for a rollout?)

Scoring in practice

Take a spreadsheet. One row per opportunity. Columns:

• Opportunity name
• Annual impact ($)
• Effort category (S/M/L)
• Readiness (0–10)
• Confidence in impact estimate (0–10)
• Ranked score

The ranked score is impact-weighted by confidence, divided by an effort multiplier (1x for small, 2x for medium, 4x for large), then scaled by readiness. The math is less important than the discipline of filling every cell with a defensible number.

A mistake teams often make is trying to model this in a single formula and treating the ranking as truth. The ranking is a starting point for a conversation, not a verdict. The value is in what happens when someone challenges a score — you find out where the confidence is real and where it's cosmetic.

What separates the top two ideas from the next ten

Once the sheet is filled in, the top of the list usually has a clear shape:

• High-confidence impact — you've measured the current process
• Small or medium effort — no exotic infrastructure required
• Readiness score above 7 — the workflow is stable, the data is decent, the team is willing

Anything that doesn't hit those three usually gets pushed to phase two. Not because it's a bad idea, but because it's not the right first idea.

Common misjudgments

Scoring by ambition

The most exciting idea usually has the lowest readiness score. Rebuilding a core product with AI is a phase-three bet, not a phase-one project. Start with a workflow that already works, then improve it.

Ignoring adoption

The most technically impressive AI project we've watched fail was a document summariser that no team used because it lived in a separate tab from their inbox. Adoption is a readiness input, not an afterthought.

Confusing "we could" with "we should"

Almost every operational workflow can be partially automated. The question is whether the marginal automation is worth the maintenance burden. If a process runs 10 times a year, custom AI is almost never the right answer.

Scoring without a control

Before you commit to an AI approach, ask what the non-AI version looks like. Sometimes a well-designed form or a clean SOP saves 70% of the time an AI feature would — for $0 and no vendor lock-in.

Two examples from real audits

Example one: a services firm with 90 people

Fourteen ideas on the whiteboard. Top of the list after scoring: an internal proposal drafting copilot. Impact was clear ($180,000 a year in senior time). Effort was small (three weeks). Readiness was high (the proposal template was stable, the data was in one place, the sales team was asking for it).

Second on the list, which the CEO had originally called "the priority": an AI SDR that would replace their outbound function. Impact looked huge. Effort was medium. Readiness was 3 out of 10 — outbound didn't have a documented playbook, and the team was skeptical. Scored honestly, it dropped to a phase-two bet.

Example two: a logistics operator with 200 people

Their loudest ask was a chatbot for their customer portal. Scored honestly, the impact was small (the portal had low ticket volume). Meanwhile, an invoice reconciliation workflow scored highest — dull, unsexy, and worth $400,000 a year in finance team hours. That went to phase one. The chatbot got pushed to phase three.

How the ranking becomes a roadmap

Once the top three or four ideas are set, they get sequenced into a plan:

• Days 0 to 30 — one workflow live in production, with a clear before/after metric
• Days 31 to 90 — the second and third workflows built and rolled out, with adoption reviewed at the 60-day mark
• Months 4 to 12 — the phase-two ideas that were parked pending phase-one lessons

The sequencing isn't just calendar work. It's a bet that the lessons from the first workflow will change the plan for the second one. That's the whole point of doing phase one before phase two.

Frequently asked questions

How many AI opportunities should we score at once?
Ten to fifteen is the useful range. Fewer means you haven't been thorough. More means you've stopped being critical.

Who should own the scoring exercise?
Whoever holds the budget. Otherwise the ranking is advisory instead of decisive. In practice, an ops lead or COO tends to run the sheet, with the CEO signing off on the top three.

How often should the priority list get revisited?
Every quarter. Readiness changes fast — a data cleanup or a personnel change can move an idea from phase three to phase one in weeks.

Should we let an external partner do the scoring?
Only if they're accountable to the numbers, not the roadmap. A partner who ranks ideas honestly — including ones they wouldn't want to build — is worth the fee. A partner who ranks their favourite services at the top isn't.

Where this fits in a bigger plan

Prioritisation is the middle of an AI audit, not the whole thing. It sits after you've mapped the processes and before you've written the action plan. If you want the full sequence, our post on what an AI audit actually looks like walks through the two-week arc.

Working with us

The AI Audit uses this exact scoring framework, applied to your specific operation. Two weeks, five deliverables, and a Google Sheet you own — not another slide deck.