ESCALATING COST
The scope was never bounded, so the pilot kept growing to cover the next exception.
Every pilot is narrow, with its success criteria and its stop rule written before any code.
OUR METHOD
Every stage ends with something in your hands and a decision that is yours to make. Most clients start at one and climb as the evidence accumulates.
Each stage stands on its own. You invest in one, you get what it produces, and then you decide whether the next is worth it. We bill after we deliver, phase by phase.
| 01 /Fluency | 02 /Triage | 03 /Pilots | 04 /Production | |
|---|---|---|---|---|
| Engagement | 1 to 4 wks | 3 to 4 wks | 4 to 6 wks | 6 to 8 wks |
| You receive | People who can judge a use case, and a written list of candidate agents | A ranked, costed shortlist with a named owner per line | Measured benefit and real run cost, pilot by pilot | A runbook, monitoring, and two trained owners |
| You supply | An executive sponsor who can settle scope | One process owner per candidate workflow | Access to real data, under your governance boundary | Two people named to own the system |
After each of the first three stages you decide whether to proceed or stop. After Production, the system is yours.
Each duration is the working window for that stage on its own. They are not a schedule, and nothing commits you to running them back to back.
Gartner puts the cancellation rate above 40% by the end of 2027, and the three causes behind it are all the same kind of thing: a decision made too early, or a decision never made at all. The four stages exist to force each one at the point where it is still cheap.
ESCALATING COST
The scope was never bounded, so the pilot kept growing to cover the next exception.
Every pilot is narrow, with its success criteria and its stop rule written before any code.
UNCLEAR VALUE
The use case was never ranked against the alternatives. Nobody could say what it was worth.
Triage ranks the whole opportunity set on value, feasibility and time to impact, before anything is built.
INADEQUATE CONTROLS
Governance arrived after the build. Usually after an incident.
The rules are written at Triage, before there is anything to govern, and enforced in production.
Gartner, 25 June 2025.
01 / FLUENCY
Typically, tools get bought, training gets skipped, and usage stays shallow. The workshop fixes the order. We teach a method, everyone configures their own tools in the room, and nobody leaves without completing real work from their own role.
Then your people tell us which agents they want for the processes they own, and that is the list Triage starts from.
Fluency comes first for a reason. An agent handed to a team that cannot read its output is switched off inside a quarter.
When the workshop ends, the line stays open. An email address, for anyone at your company, for the questions that only surface once someone tries this on real work.
02 / TRIAGE
Triage takes the opportunity set function by function and ranks it on value, feasibility and time to first impact. You receive a costed shortlist with a named owner on every line: the ones we recommend you build first, and the ones we recommend you do not.
Most AI work that fails was built correctly and chosen badly. Triage catches that, at a fraction of the build it prevents.
The governance rules are written here too, before there is anything to govern: data boundaries, tool tiering, disclosure, and where a human has to review.
Nothing reaches a pilot without a ranked line, a named owner and a set of outcomes to build against.
1 MAP
Every repeated decision, document flow and handoff, function by function.
2 SCORE
Value at stake, feasibility with today’s models, and time to first impact.
3 CUT
Anything that fails the data test goes.
4 COST
Build effort, run cost, and the human time actually freed.
5 RANK
A ranked shortlist, an owner per line, and the pilots to build first.
03 / PILOTS
Triage does not pick one pilot. It picks the set worth building, and they run together. Running several at once is what makes the arithmetic honest. You find out which workflows earn their cost and which do not.
Each is a narrow workflow with its own success criteria and its own stop rule, both set before any code. We build by rapid prototyping: working versions in front of your users within days, then cut, keep or rebuild on the evidence.
They run against your real historical data, including the awkward records. A pilot built on a clean sample passes every test and then fails at the first real one.
We have built agents of this kind against real client data: research, finance, sales and customer intelligence.
See what we have builtWHAT THE PILOTS SETTLE
04 / PRODUCTION
The pilots worth keeping get hardened for real traffic: monitoring, logging, rollback and escalation paths. That engineering is done by our partner Delivery Devs, to a specification we write and sign off. Access, retention and audit are written down before the agent touches a production system.
Handover is a runbook and two trained owners inside your team. If we are still required in year two, the transfer failed.
This is its own stage and not the last week of a pilot, because exceptions are where most agents fail.
Ownership transfers with it: the code, the prompts, the configurations and the documentation.
What ownership coversWHERE AGENTS ACTUALLY BREAK
One test is applied before we agree to build anything at all.
SMARTER
Everything relevant, in view
An agent reads every record, filing and prior case before it answers. A person reads what the day allowed.
BETTER
The same standard, every time
No drift between the first case and the thousandth. Exceptions escalated instead of guessed.
FASTER
Minutes, and never a queue
Work measured in minutes instead of days, running whether or not anyone is at a desk.
If a workflow does not come out smarter, better or faster, we don’t build an agent for it.
Triage is designed to reach that answer cheaply. A considered no is a good outcome, and it ends the engagement at the stage you paid for.
None of them is a large commitment, and the work stalls without every one.
One executive sponsor
Someone who can settle scope inside a week, named in the engagement letter.
Where a question cannot be resolved, we stop the clock rather than build around it.
One process owner per workflow
A few hours each, from the person who actually does the work.
Where that time is not available, the workflow leaves the shortlist.
Real data, early
Under whatever boundary your governance sets, from the start rather than at review.
A pilot built on a curated sample passes every test, then fails at the first awkward record. It is the most expensive way to find a data problem.
Two named owners
Named at the start, to own the system after transfer.
If they change, the transfer is re-scoped rather than assumed.
Start with Fluency. It builds the judgement your people need, and it is the cheapest way to find out whether the rest of this is worth doing at all.
Or reach us however is easiest:
It is with a founder now. Not a coordinator.
You will hear back within two working days.