The shelf tool
Bought at end of quarter discount speed. Two workflows shipped, then the two people who understood it got busy. Still billing, no longer running.
The tool is almost never what went wrong. We find the leak first, put a number on it, and only then build something that beats the number. Everything below is the method, including where you can walk away from it.
Automation usually dies somewhere between the demo and an ordinary Tuesday. Three versions of it we keep walking into.
Bought at end of quarter discount speed. Two workflows shipped, then the two people who understood it got busy. Still billing, no longer running.
The demo was genuinely good: clean data, the happy path, applause in the room. Production is where the scanned PDFs live. It was scoped to succeed, not to survive.
It shipped, it runs, it is even elegant, and the operations budget never moved. Nobody had checked where the hours were actually going.
of enterprise GenAI pilots return nothing measurable to the P&L.
MIT NANDA, 2025of GenAI projects were predicted to be abandoned after proof of concept.
Gartner, 2024of early RPA projects failed to deliver, before AI was even the excuse.
EY, via Forbes, 2018Those are our category’s numbers, not somebody else’s. Worth knowing what sits underneath the first one: the failures traced to workflows and diagnosis rather than to the technology.
Nobody in it is a villain. Incentives just point where they point, so read the last line on each, ours included.
Licenses, renewed yearly
Adoption, whether or not it saves you an hour
Billable hours
A longer build. The meter is the business model
Recommendations
A deck you approve. Delivery is someone else’s problem
A measured outcome, month to month
Your hours back, fast. It is the only way we keep the engagement
We trace one workflow end to end and count every manual touch. The painful step usually sits upstream of the complaint.
Minutes per run, times runs per week, times what an hour really costs. Every input is shown, and a guess is labeled as a guess.
In your accounts, under your logins, starting with your worst inputs. Done means it runs without us, watched and documented.
One per step, yours to keep whatever happens next, and written for whoever has to use it. Including the person who replaces us.
One step is a third of it. That is the step we price first, and it is rarely the one anybody complains about.
Three of these were measured and one is your own estimate, so the only number you have to take on trust is one you supplied.
Written for whoever has to trust it at 2 a.m. and has never met us. A vendor you can leave is a vendor you can check.
A second engagement, unrelated to the one above. Same method, on a task nobody had ever put a number to.
a year, on one task nobody had timed
Measured, rounded and anonymized. Yours will differ.
That sentence costs us projects. It’s also the only reason a stranger should believe the rest of this page: a method that can tell you no is worth more than one that always says yes.
Five tools in, one path out, with one approval kept exactly where human judgement matters.
Built where you can see it, suspend it, or take it over.
Your worst inputs go in first. The easy ones are never what break.
Anything customer facing or money moving waits for a person.
AI is used narrowly, and gated wherever a wrong answer would cost you.
A vendor retires an API on a Friday night. A supplier redesigns their invoice template and tells nobody. Nothing failed on your side either time, and neither was preventable, so we design for the drift rather than for the demo.
Every run is watched. A failed step retries, then routes itself to a person.
Failures take one item down, never the pipeline. The other 41 invoices carry on.
A named person on every workflow. When it cannot decide, it asks.
This month’s exception is next month’s rule, and one short note a week tells you what ran and what was caught. That is the maintenance.
If a tool dies, the manual path is still documented in the Runbook.
The same arithmetic we’d do in step two, with no email gate and no “book a call to see results.” The sliders start at the shape we see most often; drag them to yours. If it comes out small, keep doing it by hand. Sincerely.
1,152 hours · 7.2 months of a person
“They turned a detailed business process into a working system.”
A small operations practice, and this page is not pretending otherwise. No four hundred person delivery organization, no analyst report with our name in it, one published client. If that disqualifies us we understand, and the method above is yours to take anyway.
Being early is not only a discount on credibility. It changes what you get, structurally, and here is that trade in both directions.
A bet on a firm without a famous logo wall.
The principals, not a pod. Whoever is on your first call is who ends up in your systems.
Patience with a client roster that is still short.
Outsized care. Our next client will read about you, so a mediocre outcome costs us more than it costs you.
A month at a time. Nothing longer is asked for.
A vendor who has to earn it again every four weeks, with deliverables that already work without us.
Two hours of your process owner’s time to start.
Speed. No account layer and no workshop billed as a deliverable. Diagnosis starts the week you say yes.
Six things people say on the first call. Open any of them.
Whatever failed was almost certainly built before anyone diagnosed it, and scoped to survive a demo rather than a Tuesday. That is what the first two steps are for, and why the worst inputs go in first. You are never asked to trust our optimism, only arithmetic you can check on inputs you approved.
Operations rarely needs intelligence. It needs reliability. Most of what we build is deterministic and frankly dull: data moving between systems on a schedule, with receipts. AI earns a narrow place, reading a messy document or drafting a reply for a human to approve, and it is gated wherever a wrong answer would cost you something. When a vendor leads with the model rather than your workflow, that is the tell.
The first two steps tell you what to build, and the Leak Map is yours to hand to whoever you like. What clients actually pay for is everything after that: the edge cases, the monitoring, the 3am failure, the upkeep as your tools drift underneath you. If somebody on your team has real ownership and real time for that, doing it yourselves is a good answer. The calculator above tells you what their hours are worth.
A hire is the right call when the work needs judgment forty hours a week. It is the wrong call when the work is rekeying and chasing, because you would be paying a salary for something a system does without sick days, and whoever you hired will spend their first year resenting it. Most strong teams end up doing both. The hire you eventually make gets a better job for what was automated before they arrived.
Systems are built in your accounts, under your logins, on permissions you grant and can revoke in an afternoon, inside the tools your data already lives in. Nothing of yours is copied onto our systems. Where part of a build has to run on our infrastructure we tell you which part and why before we start, information passes through it rather than being kept, and the Runbook documents every connection we make. The security page sets this out in full.
Everything keeps running, because nothing lives with us. The systems are in your accounts, the Runbook explains what runs and how to switch it off, and your team has already operated it without us in the room, which is part of the definition of done. Worth knowing that the usual legal default runs the other way round: a vendor owns what it builds unless the contract says otherwise. Ours says otherwise, in writing.