THE ADD-BACK · episode 07
SightingA portfolio company describes an autonomous process. What runs is a small group handling exceptions, surrounded by machine-generated messages, none of them in that process's cost centre.
The consensus readTeething. Volume will fall as the model improves.
The mechanismIt didn't. Exception volume ran 40/week at go-live and 190 by month 11, and 78% came from 9% of transactions — one structural gap, not a scatter of hard cases. Model quality was never the variable. The headcount left the schedule and didn't leave the building; it stopped being attributable to the process generating it.
The exposure1.5–2.0 FTE against a 3.0 FTE reduction — a factory roughly 50–65% the size of the headcount removed.
The testFind where failures land, count the frequent responders, look up their cost centres. If they all sit inside the automated function, you don't have one.
Hidden human factory
Exception-handling labour that survives automation, distributed across cost centres that never asked for it and don't report it separately.
The test. A subtraction. Count the FTE-equivalent hours consumed by exception handling, subtract the hours booked to the automated process's cost centre, and the remainder is the factory. If the remainder is near zero, the savings are clean and this doesn't apply to you.
The operating case below is a composite built to the shape of a real quote-to-cash automation. Cited research figures are real and sourced; the two are kept separate throughout.
| Line | Moves | Running |
|---|---|---|
| Claimed — An autonomous process, three FTEs removed, no residual manual handling | 3.0 FTE removed | |
| 1. The org chart is accurate changes the story Nobody is lying. The three roles are gone, the schedule ties, and the process genuinely runs without them for the large majority of transactions. A hidden human factory is not concealment — it is an artifact of cost accounting. Work is attributed to the cost centre of the person doing it, not to the process that generates it. The moment exception handling migrates out of the automated function it becomes invisible to the only measurement anyone runs. | −0 | 3.0 FTE claimed removed |
| 2. Where the load actually went Three people spend their mornings in the channel where failures land — one account manager, two in customer service. None is in the process cost centre; none has 'exception handling' in a job description. Two properties make it durable rather than transitional. It has no owner: nobody's headcount, nobody's metric, nobody's budget to fix, and no dashboard exists for a chat channel — which is why the line went from 40 to 190 a week over eleven months with nobody able to see the trend. And it has no name, which is why it is discussed only as 'we've been busy'. | −1.5 to −2.0 FTE | 1.0 to 1.5 FTE net |
| 3. Why it concentrates, and what that tells you changes the story 78% of exceptions came from 9% of transactions. A model-capability problem produces a long tail — a scatter of unusually hard cases. This is not a scatter. The 9% shared one property unrelated to linguistic difficulty: they required a fact not in the transaction record at all. Asked how that fact used to get resolved, the answer was that someone walked down a corridor and asked a colleague. Ninety seconds, no cost, on no process map. The automation didn't remove the work, it removed the corridor. Concentration means a structural gap you can find and price; a scatter would mean a capability problem you can only wait out. The dependencies that never fail are the ones nobody documents, which is precisely why automation removes them by accident. | −0 | 1.0 to 1.5 FTE net |
| 4. Where I was wrong — I looked in the wrong org where I was wrong I went looking for the factory inside the automated function — the reduced team, its remaining members, their time allocation. It is the obvious place and it produced almost nothing, because the two people left were genuinely doing less. The load had migrated outward, to adjacent functions, which is where it always goes and what I should have predicted from the mechanism. Work flows to whoever the customer or internal requester can reach, and after an automation the reachable person is rarely inside the automated team: it is customer service, account management or finance ops — functions that absorb rather than escalate, because absorbing is faster than arguing about whose job it is. What it cost: Three weeks searching the wrong org chart. The correction is now a rule applied first: search adjacent functions before the automated one, and search by who responds in the channel rather than by who reports to the process owner. | −0 | 1.0 to 1.5 FTE net |
| 5. The arithmetic residual A term without arithmetic is a phrase. Factory size = (exception volume per week × average handling minutes ÷ 60) ÷ productive hours per FTE-week, restricted to responders outside the process cost centre. Composite: 190 exceptions/week at ~22 minutes fully handled including wait and follow-up is ~70 hours/week, against ~32 productive hours per FTE-week for interrupt-driven work. Call it 2.0 FTE, effectively all outside the cost centre. At loaded mid-market cost, $180,000–$260,000 of annual cost the savings case does not contain. | +0 | 1.0 to 1.5 FTE net |
| Realised | 3.0 removed, 1.5–2.0 reappeared elsewhere — net 1.0 to 1.5 FTE |
The factory is roughly 50–65% the size of the headcount reduction, and every document supporting the adjustment is accurate.
Value at risk · multiple 9x, illustrative
$180,000–$260,000 of unaccounted annual cost at an illustrative 9x. Arithmetic exposed so you can substitute your own comps.
$1.6M to $2.3M of enterprise value — sitting inside an adjustment that will be presented as clean, because every document supporting it is accurate
When it surfaces. In a buyer's confirmatory diligence — the buyer's team is the first party in the chain whose economics reward looking for an offsetting cost. Nobody internally is paid to find a number that reduces their own reported savings.
THE ADD-BACK · episode 07 · diligence pack
Artifact: Raw export.
The only way to find responders, and it takes an hour.
Artifact: Named list with cost centres.
The rows outside the process cost centre are the factory.
Artifact: The counts.
You are testing concentration against scatter, and the answer determines whether the fix is structural or nothing.
Artifact: Written response, from operations rather than IT.
The answer is frequently a person's name, and that person is the corridor.
Artifact: Transcripts. Thirty minutes each.
The highest-yield hour in the pack, and it expires with attrition.
Disqualifier
"There's no such channel" is not a clean answer. Ask where failures go instead. There is always somewhere — and if genuinely nobody can say, that itself is the finding.
| Claim | Source | Sample | Class |
|---|---|---|---|
| Exception volume 40/week at go-live rising to 190/week by month 11; 78% of exceptions from 9% of transactions; 1.5–2.0 FTE absorbed across adjacent functions | Composite, built to the shape of a real engagement | n/a — composite | composite |
| ~32 productive hours per FTE-week for interrupt-driven work | Planning assumption, stated so it can be substituted | n/a — assumption, exposed | illustrative |
| 9x entry multiple | Illustrative mid-market comp | n/a — illustrative, arithmetic exposed | illustrative |
One standing caveat. Every number in this show is somebody else's measurement, and I'll tell you whose, with the sample. None of it is diligence on your deal. Do that yourself.