This is a method, not a pitch. Everything below can be run by an operations director with a sheet of paper and an afternoon. The figures quoted from other people’s pages were read on 21 September 2026 and are linked in Sources.

At a glance

  • Only 5 per cent of task-specific AI tools reached production in the largest published study of the gap, and the reasons are mostly process.
  • Frequency times duration decides whether a process is worth automating. Irritation does not.
  • Thirty minutes on paper is enough to map a process well enough to make the call.
  • Three conditions make a process un-automatable at any price, and one of them is free to fix.
  • Our own version is fixed at £2,000 over fourteen days, and it is not credited against a build.

Why does automating a broken process make it worse?

Automation makes a process faster and more consistent. It does not make it correct. If the sequence is wrong, you have just bought a machine that does the wrong thing at scale, reliably, and with fewer people watching it.

There is a second, quieter cost. A manual process has a person in it who notices. When the order form looks odd, somebody frowns and picks up the phone. Automate around that person and the frown goes away, but the odd orders do not. The exceptions that used to get caught informally now pass straight through, and you find out about them at the month end.

So the first question is never which tool. It is whether the process, as it runs today, is the process you would design if you were starting now.

How many automation projects actually reach production?

The most widely quoted figure is that 95 per cent of organisations get zero return on generative AI. It comes from MIT’s Project NANDA report, The GenAI Divide: State of AI in Business 2025, published July 2025, and it is worth reading what it actually measured before repeating it.

The report’s funnel for task-specific AI tools is the more useful number: about 60 per cent of organisations evaluated such tools, 20 per cent reached pilot stage, and 5 per cent reached production. The report names the reasons as brittle workflows, lack of contextual learning, and misalignment with day-to-day operations. All three are process failures, not model failures.

Now the caveat almost nobody publishes. The report states its own limits plainly: the figures are “directionally accurate based on individual interviews rather than official company reporting”, sample sizes vary by category, and success definitions differ between organisations. The evidence base is 52 structured interviews, 153 survey responses collected at four conferences, and a review of more than 300 publicly disclosed initiatives, over a six-month research period in the first half of 2025. That is a serious piece of work and it is not a census. Treat 95 per cent as a strong signal about a pattern, not as a measured rate for your business.

One further finding from the same report is directly relevant here: in its sample, external partnerships with configurable tools reached deployment around 67 per cent of the time against around 33 per cent for internally built tools. The report notes those are self-reported and may not account for confounding factors. We would add the obvious: it is also the finding most convenient to people who sell implementation, which is us, so weigh it accordingly.

The five questions that decide whether a process is worth automating

Answer these before you map anything. If a process fails two of them, stop there and pick a different one.

  1. How often does it run, and how long does it take each time? Both numbers, not one. A two-hour task that runs monthly is 24 hours a year. A ten-minute task that runs twenty times a day is more than 800 hours a year. The second one is worth real money and the first one is not.
  2. Is the input in a form a machine can read? A PDF from the same supplier every week is workable. A photograph of a handwritten docket is a different project with a different price.
  3. Are the rules writable down? If the person doing it can explain the decision in sentences that start with “if”, you have rules. If the answer is “you just know”, you have judgement, and judgement is expensive to encode and dangerous to guess at.
  4. What happens when the output is wrong? If the answer is “somebody notices next week”, you need checking built in, and that is a cost. If the answer is “it goes to a customer with a price on it”, you need a person in the loop, and that changes the design.
  5. Who owns this process today? Not who does it. Who decides how it runs. If nobody can be named, the honest answer is that there is no process yet, only a habit.

How do you map a process in thirty minutes?

Sit down with the person who actually does the work, not their manager, and draw it. One box per step, left to right, on a single sheet.

For each box, write four things: who does it, how long it takes, where the information comes from, and where it goes next. Then go back along the row and mark every box where the work waits for somebody. Queueing time is almost always larger than handling time and it is almost never in anyone’s estimate.

Then do the part people skip. Ask what happens when it goes wrong, and draw those branches too. In most operations the exception path is where the hours go, and it is invisible on an official process document because nobody writes down the workaround.

Finally, ask the person to tell you the last three times the process did not follow the drawing. If they can, you have found the real scope. If they cannot, you have drawn the official version and should try again with somebody else.

What does a good process audit document look like?

It has four things in it, and it fits on a few pages.

A map of the process as it actually runs, including the exception paths. A cost, with the assumptions written beside it so a finance director can argue with the assumption rather than the conclusion. A decision, in the form of a recommendation with a number against it. And a rejected list: the things you considered and decided not to do, with the reason. The rejected list is what separates an audit from a sales document, because a document that recommends everything has recommended nothing.

How do you price the manual work?

Take a loaded hourly cost. Salary plus employer costs, divided by working hours. A £28,000 salary is roughly £18 an hour loaded, and being ten per cent out on that figure will not change your decision.

Then multiply three numbers you measured rather than guessed: frequency, duration, and the number of people involved. Add the queueing time you marked on the map, because the work sitting in a queue is often the thing the customer notices. Write the total, then write the three assumptions that would change it most.

The honest version of this arithmetic has two more lines. First, the saving is only real if the freed hours go somewhere. Five minutes returned to forty people is not a saving, it is a slightly better day. Two days a week returned to one person is a saving, because you can point at what they do instead. Second, subtract the run rate: hosting, usage and maintenance do not stop.

Innovate247 publishes a version of this arithmetic with worked numbers, and it is the clearest example on the open web. It is also, like ours, a vendor’s arithmetic about a hypothetical business. That is the limit of every worked example, including this one, which is why the next section links something real.

What makes a process un-automatable?

Three conditions, and they are not about technology.

Too many exceptions. If more than a third of cases take the unusual path, there is no usual path. Automate the third that is genuinely routine and leave the rest alone.

Too many disconnected systems. Each integration is a dependency you do not control. Three is manageable. Seven, two of which have no usable interface, turns a process project into a systems project with a different budget.

No owner. This is the only one that costs nothing to fix and the only one that stops the project dead if you do not. Software makes a defined process faster. It does not create the definition.

When is the honest answer to fix the process and buy nothing?

More often than any vendor will tell you, because most of the pages on this subject are published by companies selling workflow software, and their revenue depends on the answer being yes.

Buy nothing when the process only breaks because two people were never told who does what. Buy nothing when the fix is deleting a step rather than speeding it up. Buy nothing when the process changes every few months, because you would be paying to set concrete around something still moving. And buy nothing when the total you calculated above is comfortably under the cost of measuring it properly. Keep your money and go and look at a different process.

What we charge to do this properly

The 30-minute fit call is free and nothing is purchased on it. The business-wide AI Constraint Audit is a fixed £2,000 over fourteen days, including three sixty-minute working sessions with the engineers who would scope any build.

You leave with four documents you own: a constraint map, an AI opportunity map, economic cases, and a ranked next step. You can hand them to another supplier and get a comparable quote. That is the test of whether the audit was a real piece of work or a sales asset in a different font. Full details, including the worst case, are on the AI Constraint Audit page.

The fee is not credited against a build. That is the mechanism that lets us say no.

A priced output you can inspect

Arithmetic about a hypothetical business is where every page on this subject stops. Here is one that is not hypothetical.

QSQuoter, a system we built and run, priced a first-floor build-over extension in Heaton Moor, Stockport, issued by Cheadle Construction on 14 July 2026, reference JAY-AA-20260713. Sixty-three measured lines, £98,214.72 including VAT, published in full so every line traces back to a dimension somebody can check. The process it replaced was quoting. The artefact is the proof, and what it proves is bounded: that we can build software which prices a construction job to an auditable standard. It does not prove hours saved for anyone else. What we cannot show you yet is stated on our evidence page.

What this does not apply to

This method is written for UK businesses of roughly £1m to £50m turnover with processes that repeat: a contracts manager at a construction firm, a finance lead at a clinic group, an operations director at a distributor. It assumes somebody can be taken off the floor for half an hour.

It does not apply to one-off projects, where there is no process to measure. It does not apply to creative or advisory work whose value is in the judgement rather than the sequence. It does not cover regulated decisions about individuals, where the ICO’s guidance on AI and data protection adds requirements this page does not attempt to summarise. And it does not replace a formal controls audit, which is a different discipline with a different purpose.

When we are the wrong choice

If you already know which process, already have it documented, and already have a product that fits, buy the product. You do not need an audit and we will say so on the call.

If you want the audit to arrive at a conclusion you have already chosen, we are the wrong people, because the ranked list sometimes puts the favourite project fourth.

If the real problem is that two departments disagree about who owns the work, that is a management decision and no software resolves it.

If you need the answer this week, fourteen days is too slow, and you should say so rather than start.

Sources

Read on 21 September 2026.

The short answers

Each answer is the opening of the section it links to, so nothing here is written for a crawler that a reader cannot also see.

Why does automating a broken process make it worse?

Automation makes a process faster and more consistent. It does not make it correct. If the sequence is wrong, you have just bought a machine that does the wrong thing at scale, reliably, and with fewer people watching it.

How many automation and AI projects actually reach production?

The most widely quoted figure is that 95 per cent of organisations get zero return on generative AI. It comes from MIT's Project NANDA report, The GenAI Divide: State of AI in Business 2025, published July 2025, and it is worth reading what it actually measured before repeating it.

Which five questions decide whether a process is worth automating?

Answer these before you map anything. If a process fails two of them, stop there and pick a different one.

How do you map a business process in thirty minutes?

Sit down with the person who actually does the work, not their manager, and draw it. One box per step, left to right, on a single sheet.

How do you work out what manual work is costing you?

Take a loaded hourly cost. Salary plus employer costs, divided by working hours. A £28,000 salary is roughly £18 an hour loaded, and being ten per cent out on that figure will not change your decision.

When is the right answer to fix the process and buy nothing?

More often than any vendor will tell you, because most of the pages on this subject are published by companies selling workflow software, and their revenue depends on the answer being yes.

What is a business process audit?

A structured read of how a piece of work actually runs: every step, who does it, how long it takes, how often it happens, where the information comes from and what happens when it goes wrong. It ends in a decision about whether to change the process, automate it, or leave it alone.

Should we audit the process or just buy the software?

Audit first, unless the process is already well defined and a product obviously fits it. Buying first means you discover the mismatch after the licence is signed, and the usual result is a team doing the old process plus data entry into a new tool.

How long should a process audit take?

A single process can be mapped in thirty minutes by the people who do it. A business-wide audit that ranks several processes against each other and prices them takes longer. Ours is fourteen days and fixed at £2,000.

Can we run a process audit ourselves?

Yes, and for one process you probably should. The method on this page is the whole method. Outside help earns its fee when you need several processes compared against each other, when the answer has to survive a finance director, or when nobody internally can be neutral about their own department.

What does AGMM charge for a process audit?

The 30-minute fit call is free. The business-wide AI Constraint Audit is a fixed £2,000 over fourteen days and includes three sixty-minute working sessions with the engineers who would scope any build, plus four documents you own outright.

Is the audit fee credited against a build?

No. If it were, we would be paid more for recommending a build than for recommending you do nothing, and you would be right to discount the recommendation. Keeping it separate is what makes a no answer credible.

What if the audit says do not automate anything?

That is a valid result and it happens. You keep the four documents, you know what the process costs you, and you have spent £2,000 instead of committing to a build that would not have paid back.

Who should be in the room for a process audit?

The person who does the work, the person who owns the outcome, and somebody who can approve a change. Without the first you get the official version of the process. Without the third you get a map and no decision.