If you have ever sat in a room where someone said “we should use AI for that” and everyone nodded, you already know the problem. Nodding is free. Building is not.
The missing step is not another tool demo. It is a simple question: what is this friction worth in pounds per year? Once you have a number — even a band — you can rank ideas. Without it, you get vibes, politics, and shelfware.
This is how we score opportunities inside an AI Diagnostic. It is the same habit I used as a CFO when the board asked “is this worth it?” — just applied to workflows instead of capex.
What we are actually scoring
Not “AI” as a concept. A concrete piece of day-to-day friction, for example:
- Copy-paste between CRM and a quote spreadsheet
- Chasing the same missing info on every new lead
- Re-typing invoices or job sheets into three systems
- Writing the same style of email reply twenty times a week
- Manually building a weekly ops pack nobody fully trusts
If you cannot describe the work in plain English, you are not ready to score it. Start with the workflow, not the model.
The five inputs (keep them boring)
We use five lenses. None of them need a data science team.
1. Time
Who does this? How often? How long each time?
Example: two people, 4 times a week, 25 minutes each → about 173 hours/year.
Fully loaded cost matters more than salary. A conservative £35–£60/hour for a skilled ops or finance person is often fair for UK SMEs once you include employment costs and the fact that this work steals time from higher-value work.
173 hours × £45 ≈ £7,800/year in time alone.
2. Cost and error
Mistakes, rework, missed discounts, late fees, duplicate orders, wrong quotes. Ask: “What goes wrong when this is manual — and what does that cost when it happens?”
Even a small error rate adds up. One bad quote a month at £200 of margin leakage is another £2,400/year.
3. Revenue
Not every automation grows sales. Some just stop the leaks. Where it does help revenue: faster response, fewer dropped leads, capacity to take more work without another hire.
Be honest. If the link to revenue is weak, leave it near zero. Inflated upside is how bad projects get approved.
4. Effort to fix
Rough build-and-embed cost: days of work, integrations, change for the team, ongoing babysitting. We score effort as low / medium / high and translate that into a simple cost band for a first version — not a three-year platform roadmap.
5. Risk
Data sensitivity, regulation, “if this is wrong someone gets hurt or fined”, dependency on a flaky API, or a process that changes every week. High risk does not kill an idea. It changes the order and the design (human-in-the-loop first, full auto later).
A simple working formula
For ranking, we do not need discounted cash flows. We need a decision-grade annual figure:
Annual value ≈ (hours saved × loaded rate) + error/rework avoided + revenue unlocked
Then we set that against effort and risk:
- High value + low effort + low risk → do now
- High value + higher effort → do next (maybe after a thin pilot)
- Low value or high risk with weak upside → do later or deliberately ignore
That last category is as important as the first. Half the value of a diagnostic is permission to stop chasing shiny ideas that do not pay.
Worked example (anonymised shape)
A services firm regenerates similar proposals by hand.
- Time: 3 people × 6 proposals/week × 40 minutes → ~624 hours/year
- Loaded rate £50/hour → ~£31,200 time cost
- Errors/rework: wrong scope left in a template twice a quarter → ~£4,000 margin impact (conservative)
- Revenue: faster turnaround wins maybe 2 extra jobs a year → thin, say £6,000 contribution (optional, only if credible)
Illustrative annual value band: ~£35k–£40k before revenue upside; higher if the win-rate story holds.
If a first version of structured proposal generation costs a few thousand to stand up and a day of training, the ranking is obvious. If it needs a six-month content project and legal rewrite of every clause, effort jumps and it may drop to “do next”.
The point is not the exact mid-point. It is that the room can now argue about assumptions, not abstract enthusiasm.
What we refuse to do
- Fake precision — £11,472.18 looks clever and is usually nonsense
- Vendor ROI slides — written to sell a licence, not your workflow
- Scoring “AI maturity” without naming the jobs people hate
- Counting only salary saved and ignoring rework, delay, and management attention
Good enough beats theatrical. Ranges beat false certainty.
How this feeds do now / do next / do later
After scoring a short list of opportunities we sort them:
- Do now — clear £ value, feasible with current systems, low enough risk to ship a thin version quickly
- Do next — strong value but needs data clean-up, an integration, or a pilot
- Do later / not worth it — weak value, chaos process, or risk that swamps the upside
That ranked list is the main output of an AI Diagnostic. Delivery is optional: we can build the priority items with the same AI-assisted approach we use on our own products, or you can take the plan and run it yourself.
Try it on one process this week
Pick one annoying weekly task. Ask the person who does it:
- How often, and how long each time?
- What breaks when it goes wrong?
- What would you do with the time back?
Multiply. Put the number next to a rough build effort. If the ratio feels embarrassing in either direction, you already learned something useful.
Want this done across the business, not one process?
An AI Diagnostic maps the friction, scores opportunities in £, and leaves you with a do-now / do-next / do-later plan — plus what not to automate.