Why AI Can't Fix a Process Nobody Owns

There is a question we ask on the first call, and it is not about AI.

Who owns this process?

Ask it, then listen for the pause. Then a name, then a correction, then two names, then something like "well, it depends where it is in the month." That pause is the finding. Everything else in this guide follows from it.

What "nobody owns it" actually costs

A process without an owner has no one who notices when it changes shape.

Volume drifts up ten percent and nobody adjusts. A field that used to be filled in stops being filled in, and two steps later somebody starts working around it by hand. An exception path that was built for four cases a month is carrying sixty. None of that produces an alarm, because an alarm requires somebody to have decided in advance what normal looks like, and deciding what normal looks like is what an owner does.

So the decay is silent, and the detection event is a customer complaint.

Now automate it.

The automation runs the process as it was on the day it was built. The drift continues underneath. The exception path fills faster because the clean cases now arrive faster. The one person who used to catch the weird ones by eye is no longer in the flow, because the flow does not pass through her desk any more. Six weeks later the queue is four hundred deep and the first anyone hears of it is a customer.

The automation did not cause that. It just removed the last human who was accidentally holding the process together, and it removed her quietly, and it made everything else faster while it did.

That is the whole thesis of this firm: most companies do not have an AI problem. They have an operating problem that AI would have made faster and worse.

One piece of research points the same way, and it is worth reporting carefully rather than overselling. In McKinsey's March 2025 survey of the state of AI, of twenty-five organisational attributes studied, redesigning workflows showed the strongest association with respondents' reported EBIT impact from generative AI. Read the caveats with it: that is a survey, the impact is self-reported, an association is not a cause, and the respondent base is not the $15-40M American mid-market you are running. It does not prove the argument. It is consistent with it, which is a different and smaller claim.

The arithmetic, before anyone builds anything

Here is what we do instead of a demo.

Addressable annual value =
    (transactions/yr x minutes each x loaded hourly rate / 60)
  + (leads lost to slow response x close rate x contribution margin)
  + (rework and error correction hours x loaded rate)
  + (DSO reduction days x daily revenue x cost of capital)

Four lines. We run them out loud on a 25-minute call with the owner's own numbers, doing the multiplication aloud so they can hear it. It does two jobs at once. It sizes the opportunity, and it finds out whether the numbers exist at all -- which, when they do not, is the more useful of the two answers.

Below is the whole thing worked through. Take the structure, throw away the inputs, and put yours in.

Every number below is illustrative. It is a shape you can re-run on your own business, not a reported outcome.

The company in the example

A regional service business. Around $27M in revenue. Roughly 40 people, of whom about 22 do repeatable process work -- intake, scheduling, dispatch admin, billing, collections. Work arrives as a job, gets done in the field, and comes back to the office to be turned into an invoice.

Line 1 -- manual processing

Count the transactions, time one honestly, multiply.

40,000 work orders a year
18 minutes of hand keying each
$35 an hour, fully loaded

40,000 x 18 = 720,000 minutes
720,000 / 60 = 12,000 hours
12,000 x $35 = $420,000

Line 1: $420,000.

Two things to be careful about here. "Fully loaded" means wages plus payroll taxes plus benefits plus the share of overhead that person consumes -- not their hourly wage. We default to $35 until you give us a real figure. The loaded figure is higher than the bare wage by arithmetic: it is the wage plus everything added on top of it.

And 18 minutes has to be the honest number, including the ones that go wrong. Ask how long a task takes and you get the clean case. Ask how long the last one took and you get a different answer. Use a stopwatch on ten real cases if you can. Check what your CRM, ticketing tool, dispatch system and ERP already write: where there are timestamps, they tell you without anyone being watched.

Line 2 -- leads lost to slow response

600 inquiries a year go cold before anyone calls back
20% of them would have closed
$900 contribution margin each

600 x 0.20 = 120 lost sales
120 x $900 = $108,000

Line 2: $108,000.

This is the line that is easiest to get wrong, in both directions. Do not use revenue per sale -- use contribution margin, the money left after the variable cost of delivering it. And do not use your overall close rate for leads you never called. Those two are not the same population -- a lead nobody called is a weaker lead than one you worked -- so use a lower rate than your overall close rate, and write down which rate you used.

The count itself is a query, not an estimate: inquiries with no outbound contact logged within your own stated response window, for a year. If your system cannot produce that number, that is a finding.

Line 3 -- rework and error correction

3,000 hours a year fixing bad records, re-issuing invoices,
chasing missing paperwork
$35 an hour, fully loaded

3,000 x $35 = $105,000

Line 3: $105,000.

Rework goes uncounted, because nobody books it anywhere. It is somebody's Thursday afternoon. The way to find it is to ask each person in the process what they spend time fixing that should not have needed fixing, then be conservative with what they tell you.

Be careful not to double count. If the 18 minutes in Line 1 already includes fixing the record, it does not belong in Line 3 as well. Being honest about this takes a chunk out of the number, which is the point.

Line 4 -- cash timing

Collections run 9 days slower than they should
$110,000 of revenue a day
8% cost of capital

9 x $110,000 = $990,000 of cash tied up
$990,000 x 0.08 = $79,200

Line 4: $79,200.

Note what this line is and is not. It is the annual carrying cost of cash arriving nine days late -- worth real money, and not profit. If you put this in a report as profit, your CFO will find it, and rightly.

The nine days also has to be the part you can actually fix. Some of your DSO is the customer's payment terms and their internal process, and no amount of automation changes that. The recoverable part is the delay between work being finished and an invoice going out correctly.

Adding up

Line 1  manual processing        $420,000
Line 2  leads lost               $108,000
Line 3  rework                   $105,000
Line 4  cash timing               $79,200
                                 --------
Addressable annual value         $712,200

Now the part that decides it

$712,200 is a large number. It is not the answer. The answer comes from comparing it to what fixing it costs.

Our full program is a $75,000 implementation plus $20,000 a month. Over twelve months that is:

$75,000 + (12 x $20,000) = $315,000 in first-year fees

We use a 3-to-1 hurdle -- the work has to be worth about three times what it costs before we will take it on. That is our own rule, not an industry standard, and we picked it because it leaves enough room for the estimate to be wrong and the project still to be worth doing.

$315,000 x 3 = $945,000 of addressable annual value required

And $712,200 does not clear $945,000.

So the honest answer for this made-up company is no at full price. We would say that on the call, in minute twenty, having done the arithmetic out loud with the owner rather than going away for two weeks and coming back with a proposal.

There is a yes underneath it. At our entry tier -- one workflow, end to end, $40,000, plus $5,000 a month:

$40,000 + (12 x $5,000) = $100,000 in first-year fees
$100,000 x 3 = $300,000 required

$712,200 clears $300,000 more than twice over. That is the shape of the right engagement for this company: narrower, cheaper, and comfortably inside the hurdle.

This is why the paid diagnostic exists. It is cheaper to discover the shape of the answer in three weeks for $12,500 than in nine months for $315,000.

Why 15 or more people, and not a revenue band

Run the hurdle backwards and the real qualifier falls out.

$945,000 of value required
at $35 an hour, fully loaded
$945,000 / $35 = 27,000 hours
27,000 / 2,080 hours a year = about 13 full-time equivalents

Thirteen FTEs of pure process work. Nobody spends every hour of every day on automatable process, so divide thirteen by the share of a role that actually is. At 85 percent of a role that is about 15 people; at 65 percent it is 20. That is the qualifier: 15 to 20 or more people in process roles before the full program clears.

Which is why we do not qualify by revenue. An $8M company running 40,000 service tickets has more recoverable value than a $30M company closing 180 large deals. Volume and process labour pay for this work. Revenue is a proxy for them, and a poor one.

Against revenue the hurdle looks like this, which is useful mostly as a sanity check:

If somebody quotes you a percentage of revenue as an automation saving without showing you which transactions produce it, ask them for the four lines. If they cannot produce them, you have your answer.

The part that matters most, and gets skipped

Everything above sizes an opportunity, before the work. It is a forecast made from the client's own inputs, and forecasts are allowed to be added together, because nobody is claiming the total already exists.

Benefit after the work is a completely different thing, and it is never reported as one number.

That distinction is the difference between a firm that survives to month twelve and one that does not. Here is how we report it. Four categories, always separate:

1. Incremental contribution profit -- new margin that exists because of the change.
2. Actual cash costs removed -- a payment that stopped. A cancelled subscription, a retired contractor, a courier you no longer use.
3. Capacity released -- reported as hours. Not dollars. Hours become dollars only if a role was actually eliminated or a hire was actually avoided -- and that is the client's decision, recorded, not our assumption.
4. Cash timing -- reported as days and working-capital effect. Not profit.

Fees are subtracted so the client sees net.

Look at what that does to the example above. Lines 1 and 3 -- the $420,000 and the $105,000 -- are hours, and in a report after the fact they stay hours until somebody actually changes the payroll. Line 4's $79,200 is a working-capital effect, not profit. Line 2's $108,000 is the only one that lands in contribution profit, and only for the leads you can show were recovered.

Sum those four into a single headline and you have handed a CFO a number that is not in the P&L, does not reconcile to anything, and mixes three different kinds of money. She will go looking for it. She will not find it. That is how a firm like ours gets fired in month nine, and it is entirely self-inflicted.

There is one more honest limit to state. When a client's numbers improve, other things changed too -- seasonality, headcount, pricing, a marketing push, a good quarter. We normalise the obvious confounders, use a staged rollout where it is feasible so there is something to compare against, and write down what we could not remove. A before-and-after comparison on its own does not establish cause. Any report that pretends otherwise is a sales document wearing a spreadsheet's clothes.

What to do with this on Monday

You do not need us for any of the following.

1. Pick one workflow. The thing your team does most often by hand. Not the most annoying one -- the most frequent one.

2. Get the four inputs. Volume from your system, minutes from a stopwatch on ten real cases, loaded rate from your finance lead, and lost inquiries from the log of what nobody called back. If any of them cannot be produced, write down which one. That absence is worth more than the number would have been.

3. Do the multiplication. Out loud, with the person who runs the process in the room.

4. Ask who owns it. One human being, end to end. If there is a pause, do not automate anything until there is a name. Writing that name down is free, and it is the highest-return thing in this guide.

5. Ask how you would know if it broke. If the answer is "a customer would tell us," you have found the second thing to fix, and it is cheaper than the first.

6. Compare the number to what a fix would cost. Any fix -- ours, an agency's, or two weeks of your own developer's time. If it does not clear three to one, do something else with the money.

Step four is the one that will still be paying for itself in two years, and it costs nothing at all. If the guide persuades you to do only one thing, do that one.

---

All figures in the worked example are illustrative and are there to show the shape of the calculation. They are not a benchmark, not a forecast, and not a reported outcome. The 3-to-1 hurdle is our own commercial rule. Our prices are on the approach page. If you want the arithmetic run on one of your workflows with your own numbers, that is a 25-minute call: assess your operation.