AI only matters when it creates leverage.
The colleague teaches itself: a procedure that keeps working is kept as a skill, scored every time it is used, and it gets better at your work whether or not anyone designs anything. Custom routines are the deliberate version. We decide with you where an agent genuinely creates leverage and where it doesn't, build the work that qualifies into named routines with an owner and a number to beat, and change the operating model around them so the thing actually gets used.
Capability is everywhere. Leverage is concentrated.
High volume, low variance — and the steps are written down somewhere, even if nobody follows them.
The work spans systems that don't talk to each other, and a person is the integration.
Throughput is capped by attention rather than judgement. The queue is the problem.
The same request arrives forty times a week with the same shape and a different account number.
Low volume, high judgement. The thinking is the job, and it is different every time.
The bottleneck is a decision nobody will delegate — not the keystrokes around it.
The system of record is the constraint. If the data isn't there or isn't trusted, an agent surfaces that faster than it fixes it.
The process is genuinely broken. Automating it makes a bad outcome arrive sooner, and arrive more often.
Deciding which column a piece of work sits in is the first thing we do, and we do it before anyone builds anything. Nothing on the right is off the table forever — most of it comes back once the process or the data underneath it is fixed. It is simply not where the first routine should go, and saying so in week one is cheaper for both of us than finding out in month four.
A skill is what it teaches itself. A routine is what we design on purpose.
Left alone, the colleague accumulates skills — a procedure that keeps working is kept, scored each time it is applied, and reused. That happens on its own. A routine is the deliberate version of the same thing: a named piece of work, scoped with the people who do it, carrying five things from the day it goes live.
A named human, not a function.
The person whose week changes if the routine stops working. Committees do not own routines; if the owner is a department, nobody is accountable and the routine quietly rots.
The tools it can reach, and nothing else.
A routine gets the task-level tools that job needs, each classified by blast radius. That boundary is a capability contract your security team reads, not a policy document nobody opens.
What stops and waits for a person.
Which mutations need a named human before they commit, and who that human is. The gate lives in the runtime, so it holds whether or not the model is having a good day.
A record that answers the question later.
Every action logged, replayable and reversible, attributed to the run that took it. An audit becomes a query instead of an exercise.
What this work costs you today.
The operational and commercial measures for this work, taken before the routine exists. A gain you cannot show against a number you agreed on beforehand is a gain you will spend the next year arguing about.
None of that is bolted on for the engagement. Approval gates, per-run token, cost and wall-clock ceilings, the kill switch and audit provenance ship with the platform — governance is already a property of the runtime. What a routine decides is how those controls are set for this particular piece of work, and who answers for it.
See the controlsThe routine is the easy half. What changes around it is the work.
The person does not disappear from the second row. They move to the one point in it where judgement actually changes the outcome.
Who does this now, and what do they do instead?
The honest version, agreed with the person doing it rather than announced to them. A routine that nobody's week accounts for does not survive its first busy month.
What reaches a human, in what form, and how fast?
An approval queue is a commitment to answer it. If the gate is set so tight that a person is rubber-stamping forty items an hour, the control is decorative.
Who owns the exceptions?
Most of the value and nearly all of the risk sits in the cases the routine hands back. That queue needs a name against it before go-live, not after the first bad week.
What does the team stop doing?
Usually a step that only ever existed to move data between two systems. If nothing gets retired, nothing got faster — you have added a participant, not removed a constraint.
We work these through with the people who do the job, not with an org chart. It is the part of a deployment that decides whether the technology gets used, and it is the part most AI programmes skip.
Prove it on one routine before anyone commits to ten.
Run it for real, with autonomy at zero.
The routine goes live against real work with every mutation held at the gate. Nothing commits without a named human, so you are watching the thing you will actually run rather than a demo of it — and a routine that is wrong is wrong in front of someone who can stop it.
The real work becomes the eval set.
The conversations from that period are captured as golden conversations for this routine and run as a merge gate from then on. A change that breaks how it handles a case you have already seen does not ship.
Measured against the baseline you agreed.
Against the number taken before the routine existed. We agree which operational and commercial measures count for this work up front, because choosing them afterwards is how AI programmes end up reporting activity instead of outcome.
Then it widens, or it doesn't.
Autonomy widens one mutation class at a time, as the evidence earns it. The second routine is faster to stand up than the first — the connectors, the capability boundary and the eval harness already exist — and if the evidence does not arrive, the honest answer is that this work belonged in the other column.
You do not have to buy this.
QuasiHuman is not inert without it. Tell the colleague what you need, correct it when it gets it wrong, and the procedures that keep working are kept as skills and scored every time they are used. Autonomy widens as the evidence earns it. That path costs you nothing but time, and for plenty of deployments it is the right one.
Custom routines are the same destination with us in the room from the first week: the triage done before anything is built, governance decided rather than discovered, the operating model changed deliberately, and a number on the table that you agreed before we started. It is a shortcut, not a prerequisite — and it is scoped per deployment, because the answer depends entirely on what your business actually does.
Bring us the workflow everybody complains about.
The one that eats a day a week, spans four systems and has a person in the middle keeping it upright. We will tell you whether it is worth automating before we quote you for automating it.