Work
Agents on the work, measured against a comparison.
Three ways we work, and the stages every engagement passes through. Each stage can end in a no, and the no is written down before we start.
01 · Deploy
Agents on one process, in three stages.
-
Measure
Find where the time goes
Six weeks. We follow one process end to end: which steps, which documents, which handoffs. You get a baseline you keep whatever happens next.
- Length
- Six weeks
- Ends in
- Go · Re-scope · Stop
-
Prove
Show the gain against a comparison
Twelve weeks. Agents go live on the process in waves drawn by lot. The sites still waiting are the comparison, and the decision rule is written down before the first wave.
- Length
- Twelve weeks
- Ends in
- Scale · Stop
-
Run
Operate it in production
We operate the agents in production, per site, for as long as they earn their keep, with the measurement still running. What we learn there, such as where people rely on the agents, check them and overrule them, improves our agents under your contract. It is never sold.
- Length
- For as long as it runs
- Ends in
- Your decision, on the numbers
Some clients start with one process, others with several at once. Each stage is scoped and priced to the deployment.
How Prove works
Sites go live in waves, with the order drawn by lot and the decision rule written down before the first wave. The sites still waiting are the comparison, so the gain is measured against them rather than against a forecast.
If the measured gain does not clear the rule, the work stops there. You keep the baseline and the comparison.
How deeply we check follows what it would cost you if we were wrong.
02 · Secure
Every agent is a new way in.
Every agent we deploy reads documents other people wrote. An invoice with one hidden line can read as an instruction. We attack agents the way an adversary would, before they go live.
- AI security reviewThe agent, its tools, its permissions and the documents it reads, tested for injection, data leakage and misuse.
- Agent red-teamingAdversarial testing of a deployed agent, with the findings attached to the deployment record.
Delivered with a specialist offensive-security partner.
03 · Transform
The whole business, not one process.
For an owner who has decided to change the business rather than one process: Measure, Prove and Run applied process by process, in an order set by where the time goes, with the margin as the measure.
The same gates apply, and each stage can still end in a no.
You see the diagnosis, and what would prove it wrong, before you sign.
Before any substantial engagement begins, we write this down and give it to you.
- DiagnosisWhat we currently think is wrong.
- InterventionWhat we propose to change.
- Expected outcomeWhat should change if we are right, in measurable terms.
- BaselineWhere you are today.
- CounterfactualWhat happens if you do nothing.
- Key assumptionsWhat has to be true.
- Disconfirming evidenceWhat would make us conclude the diagnosis was wrong.
- Measurement dateWhen we will know.
That document is the proposal.
When we say no
We will not sell work unless, on the evidence available at the time, we have a reasonable basis to believe the value to you materially exceeds the full cost of doing it: your people's time, your management's attention, implementation, infrastructure, migration and disruption. Not our fee.
Five outcomes, not two: accept, re-scope, defer with named prerequisites, refer elsewhere, decline.
- No accountable sponsor willing to be named and to hold decision rights
- No baseline, and no willingness to establish one
- The binding constraint is managerial, and our work would not address it
- A pilot with no pre-agreed success and failure criteria, no owner, and no production decision behind it
- Production deployment is impossible on regulatory, data-residency or contractual grounds
- We are not the right firm
Free · No email needed to run it
The Agentic Readiness Diagnostic
One workflow, seven or eight questions, about ninety seconds. Three readings rather than one score: opportunity, deployment readiness and safe autonomy. A named binding constraint. And an honest answer, including the answer that you should not do this.
Before it opens: independent certification
The diagnostic is built and has been audited against sixteen adversarial cases. It goes public after independent certification, not before.
The diagnostic encodes a practitioner heuristic, not a validated model. Testing whether its assumptions hold is one purpose of our research.
Start with one process.
Write to us with the process that eats the most time. The first call contains our qualification questions rather than a pitch.