Most businesses pick their first AI project the same way. Somebody saw a demo that looked clever, or somebody complained loudly enough in a Monday meeting.
Both are terrible reasons. The loudest process is rarely the most expensive one, and the one that demos well is usually the one facing customers, which is the hardest place to start and the worst place to be wrong.
The order you do this in matters more than the tools you pick. Pick right and the first project pays for itself, and everyone stops arguing about whether AI is worth it. Pick wrong and it quietly dies after four months, and the next person who suggests automating anything gets a look.
This is the method we use to rank processes inside an audit. Four questions, a bit of arithmetic, and a shortlist you can defend. If you haven’t yet worked out what your manual work actually costs, start there, because ranking without numbers is just opinion with a table around it.
Why the order is the whole decision
The failure rate on this stuff is not a secret.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, blaming escalating costs, unclear business value and weak risk controls. MIT’s NANDA initiative looked at generative AI inside companies and found 95% of pilots delivering no measurable impact on the P&L, with the small minority that worked getting there by changing how work actually flowed rather than by buying more licences.
Read the reasons in both and almost none of them are technical. Unclear value, costs that outran the return, work that never got embedded. Those are all symptoms of the same thing: the wrong process went first.
Meanwhile the adoption numbers keep climbing. The British Chambers of Commerce found 54% of UK SMEs now using AI in some form, up from 35% a year earlier. Plenty of businesses have started. Far fewer have a first project that paid for itself, and the gap between those two groups is mostly a sequencing problem.
The four questions that rank a process
Forget scoring matrices with 11 columns. Four questions decide it, and you can answer all four for any process in about 90 seconds.
1. How often does it happen? Per week, counted properly. Ask the person who does it, not the person who manages it. Managers underestimate frequency by a wide margin because they only see the times it goes wrong.
2. How long does it take each time? Also per instance, also from the person doing it. Include the faffing: finding the file, waiting for the system, reading back what you typed.
3. How much of it is the same every time? This is the one that separates a good candidate from a bad one. Three honest answers: same steps every time, mostly standard with the occasional judgement call, or needs a person thinking it through each time.
4. Has the process settled? If how you do this changed twice in the last six months, it isn’t ready. Anything built around a moving process gets rebuilt at your expense.
Questions 1 and 2 give you the size of the prize. Question 3 tells you how much of it software can take. Question 4 is a gate: fail it and the process comes off the list regardless of how big it is.
The arithmetic
Multiply the first two, then discount by the third.
- Same steps every time: assume software can take effectively all of it.
- Mostly standard, occasional judgement call: assume roughly 60%. AI gathers the context and drafts the work, a person checks and decides.
- Needs a person thinking each time: assume 25% at most, and usually less. This is the job, not the admin around it.
That last discount is the one people skip, and skipping it is how you end up with a ranking that puts your most skilled work at the top. A process that takes 5 hours a week but needs judgement every time is worth less as a project than one that takes 3 hours and never varies.
Then price it. Take the recoverable hours, multiply by 46 working weeks, and multiply by your loaded hourly cost. That last figure is not the wage. Once you add employer National Insurance, pension and the hours people genuinely work, it lands about a third above the wage on the contract, somewhere near £26.55 an hour on a median UK salary.
What this looks like on a real list
Say a 15-person business writes down its six worst offenders. Here’s the list as the owner would naturally order it, biggest time sink first.
| What happens | Times a week | Minutes each | Hours a week | How much is the same every time | Hours software could take |
|---|---|---|---|---|---|
| Producing quotes and proposals | 6 | 40 | 4.0 | Mostly standard | 2.4 |
| Answering inbound enquiries | 25 | 8 | 3.3 | Mostly standard | 2.0 |
| Pulling the regular reports together | 3 | 60 | 3.0 | Same every time | 3.0 |
| Retyping orders into the accounts system | 15 | 8 | 2.0 | Same every time | 2.0 |
| Chasing invoices and payments | 10 | 12 | 2.0 | Same every time | 2.0 |
| Compliance checks and evidencing them | 5 | 20 | 1.7 | Needs a person | 0.4 |
| Total | 16.0 | 11.8 |
Now re-sort by the last column and the order changes.
Reporting goes to the top, even though it was third on raw hours, because every minute of it is mechanical. Quotes drop to second, even though they’re the biggest number on the page, because a person still has to price the thing. Compliance falls off the list entirely: 1.7 hours a week sounds worth having back until you accept that the part needing a person is the part that matters.
Those recoverable hours across 46 weeks at £26.55 come to just over £14,000 a year, and nearly 12 of the 16 hours. The figures are illustrative and yours will differ. The shape usually doesn’t.
The four answers you’ll get
Every process on your list lands in one of four places. Naming them stops the argument, because the fight is almost always between “automate it” and “leave it alone” when the honest answer is one of the middle two.
Automate first. High volume, same steps every time, settled process. Software does the whole thing and a person sees the exceptions. This is where the payback is fastest and it’s what we’d put at the top of an audit.
Assist, keep a person. Real volume, but judgement in the middle of it. AI pulls the context together and drafts the output, then routes it to someone to approve. You get most of the time back and none of the risk of a machine deciding something it shouldn’t.
Worth doing, not first. Genuinely automatable, but the volume doesn’t justify going first. Bundle it into a later phase once the big one is running and paying for itself.
Leave it alone. Low volume, or a person has to think every time. Automating it would cost more than it returns. Anyone who tells you otherwise is selling something.
If you run the same six processes through our what to systematise first scorecard, it does exactly this arithmetic and hands you the ranked list with the verdict against each line. It takes about 3 minutes and the link you get back keeps your numbers, so you can send it round before anybody argues.
Five processes that look like great first projects and aren’t
The one the loudest person complains about. Volume of complaint and volume of work are different measurements. Count before you commit.
The customer-facing one. It’s the most visible, so it’s the most tempting, and it’s the one where being wrong costs you a customer rather than 10 minutes. Do the back-office one first, get the wins banked and the trust built, then go near anything a customer touches.
The one only the owner does. Usually the most valuable-feeling and the least documented. It lives in one head, with a set of exceptions nobody has ever written down. That makes it slow and expensive to build and easy to get wrong.
The one sitting on messy data. If your customer records live across three spreadsheets, two inboxes and somebody’s memory, tidying that comes first. Building on bad data just makes the bad data faster.
The one that changes every month. Still moving means still cheap to leave alone. Come back when it settles.
What a good first project actually looks like
Five things, and a genuinely good candidate has all five.
- It happens at least a few times a week, every week
- The steps are the same each time, and someone could write them down in an afternoon
- The data it needs already lives somewhere a machine can read
- Nothing breaks for a customer if it’s wrong on Tuesday
- Somebody in the business will notice the time coming back
That last one gets forgotten and it’s the reason plenty of technically fine projects get quietly dropped. If the hours you save are spread across nine people at 20 minutes each, nobody feels it and nobody defends it when the budget gets tight. Concentrated savings survive. Scattered ones don’t.
This is a capacity plan, not a headcount plan
Worth saying plainly, because it’s what people in your team are actually wondering.
That same British Chambers of Commerce research found 95% of SMEs using AI reported no impact on workforce size over the past year, and 86% said job roles were unchanged. That matches what we see. Businesses that take 12 hours a week out of their admin don’t shrink by a third of a person. They put the time into work that was already being neglected: following up leads properly, onboarding customers faster, answering the phone.
Rank your processes on that basis and you’ll pick better ones, because you stop looking for the job you can delete and start looking for the work nobody should be doing by hand.
What the right first pick gets you
Founderise had a founder personally handling every step of delivery. One process, taken properly: 12 hours a week back, a 3.5x increase in margin, and 9 weeks from starting to a working product. The order mattered there. Delivery went first because it was the thing standing between the business and every new customer.
MidShift has served over 20,000 professionals through an AI guidance platform, with 92% faster progression through the process it replaced. That volume was never reachable one at a time by hand. Again, one process, chosen because it was the one repeating at scale.
Neither started with a strategy document. Both started with a ranked list and the nerve to only do the top item.
What to do this week
Write down every process your team does more than twice a week. Aim for 10 lines, not 3. Then get the frequency and the minutes from the people doing the work, not from the org chart.
Discount each line by how much genuinely needs a person, sort by what’s left, and draw a line under the top one. Not the top three. The top one.
Then check it against the five-point list above. If it fails on data or on stability, take the next line down. That single process, with a pound figure next to it and a reason it’s first, is worth more than any AI strategy anyone will try to sell you, and it takes an afternoon to produce.
Related reads
- The cost of manual work: what repetitive admin really costs
- AI for small business: the complete guide
- How AI agents handle client onboarding while you sleep
- Build vs buy: when custom software is worth it
- Case study: how Founderise automated delivery
Want the ranking done properly, on your real processes?
That’s what the AI audit is for. Two weeks, fixed price, every repeatable process in your business mapped and costed in hours and pounds, ranked by what to do first, plus one working proof of concept running on your own data. If you go ahead with a build within 90 days, the audit fee comes off it in full.
Book a free call and we’ll tell you honestly which process should go first, and whether it’s worth doing at all.