One process, not a platform

Why one process instead of a platform?

Because I have seen what a platform meeting does to a calendar. Six teams. A vendor. A map of systems that is out of date. Someone says 'we need a center of excellence'. Nobody says which pile will be smaller on Friday.

One process fits on a whiteboard. Cases in. Steps. Stop list. Owner. You can finish it. Finishing teaches you the actual miss rate. That number is the only strategy input that matters.

Anthropic's 2026 snapshot said most teams that deploy agents do it on multi-stage workflows, and a smaller share go cross-functional. That matches what I would recommend. Stay inside one function until the loop is boring. Cross-functional is how you inherit six policies.

What counts as one process?

A unit of work that ends in a state. Approved, rejected, needs a person, scheduled, paid. If you cannot name the states, you do not have one process. You have a department.

Good first processes are ugly and frequent. Invoice fields into a ledger with a reject queue. Clinic document first pass. Appointment no-show callbacks with a script and a stop on clinical questions. Internal access requests that are not security exceptions.

Bad first processes are 'customer experience' or 'knowledge'. Those are oceans. You will drown in meetings about tone. Tone is not a close state.

What does a week of hiring look like?

Day one is the SOP and the systems the worker may touch. If those are fictional, we stop. I will not spend your money decorating a guess.

The middle of the week is the loop: intake, structured output, log, gate, handoff of working state. We use real cases, not a happy-path script. Happy paths lie. Your last fifty tickets do not.

The end of the week is a live slice with a person on the rejects and a way to turn the worker off. Off is a feature. If you cannot turn it off, you did not hire. You fused.

That is the week. It is on /ai. Book 30 minutes on Calendly if you already know the process. If you do not, write the process and then book. I am not precious about the call. I am precious about the scope.

What do you refuse to add in week one?

A second department. A second model vendor because someone read a benchmark. A redesign of the source system. A public chatbot on the same contract. A board demo with fake data.

Each of those is a reasonable later project. In week one they are how you avoid learning. Learning is a miss rate on real cases. Everything else is set dressing.

I also refuse to connect a wallet or a clinical message path in week one unless the gate is already staffed and the limit is small. You can learn on drafts. You do not need to learn on disbursements.

• One queue.
• One policy document.
• One log.
• One owner and one backup.
• One off switch.

How do you know you picked the wrong process?

The cases are all unique and the team is proud of that. Pride in uniqueness is a warning. You cannot hire a worker for snowflakes. Keep the artisans. Hire for the mush that they hate.

The owner is 'the committee'. Committees do not take 2am rejects. A process without a person is a platform in miniature.

The close action is unclear. People argue about whether a case is done. If humans already disagree, the model will pick a side and sound sure. That is an argument amplifier, not an employee.

What happens after the first process works?

You keep it boring for a month. You review misses. You adjust the SOP. You do not immediately tour it as a transformation. Boredom is the success state.

Then you pick a second process that shares a system or a team, so the log and the rota can stretch. Adjacent is cheaper than a new religion in finance.

If the first process does not work, you turn it off. You keep the SOP. You have still bought clarity. Clarity is more valuable than a platform license you will not open in July.

Tell the team that off is a success if the miss rate was the reason. You learned. You did not fuse a bad worker to the company. The next process will be cheaper because the SOP habit exists. That is the only compounding I trust.

What does Friday of the week need to show?

A live slice, even if it is twenty cases. Real input. A person on the rejects. A log you can open without me in the room. If Friday is a slide, the week failed. Slides are Monday's job. Friday is a queue.

You should be able to say the escalation count out loud. 'Twelve closed, two stopped, one miss we caught.' That sentence is the product. If you cannot say it, you built theatre.

You should also know how to turn it off. I make people click the off path before I leave. If they cannot, we are not done. I would rather ship a day late than leave a worker nobody can stop.

If Friday looks messy, that is allowed. Messy and true beats clean and fake. Write the three things you will change next week. Then stop adding systems. The week is for one process. The next week is for the miss review. That rhythm is the whole method.

Send the owner a short note they can forward. What went live. What is gated. Who is on Saturday if a reject sits. Where the off switch is. If they cannot forward that note without you, the week is not finished. I stay until they can.