(01)
The acceptance test is written first.
Before we build, we agree what counts as a good output. Document matches template. Required fields populated. Reviewer checklist passes. Specific. Measurable. Signed off.
What We Do
Plenty of agencies can build agents. Few will let you pay only when one delivers.
We design, build, and run AI agents for governance-led UK firms — one process, fully handled, no output, no charge. The capabilities aren't unique. The model is.
What You're Buying
Not software. A machine for one process.
A managed agent isn't a chatbot. It isn't an app your team logs into. It's trigger-based — a document lands, a case opens, a renewal passes. It runs a bounded sequence of steps and produces a structured output: a drafted engagement letter, a completed KYC check, a triaged claim. Ready for review.
It makes only the decisions you've defined, against rules you've agreed. Anything outside that goes to a person. That's what makes it safe inside a governed firm — and why the service is managed. We run the agent. You read the outputs.
Every output is a defined, countable thing. So it's billed as one. No output, no charge.
Trigger Steps Boundaries Output
The Service One service.
One service.
Everything in it.
Five work-streams go into every agent we run. None of them is optional, and none of them is sold separately.
- ( 01 )The thinking
Scoping and feasibility
- ( 02 )The core build
Agent engineering
- ( 03 )The connections
Systems integration
- ( 04 )The proof
Evaluation and governance
- ( 05 )The running
Managed operations
Not a menu. Every agent gets all five.
Processes That Qualify
Not every process qualifies.
The strongest agents run on processes with a particular shape. Five things we look for before anything gets built.
Some processes fail this list. Too rare, too variable, too much genuine judgement in the middle. Those aren't worth automating, and we'll say so. Some calls end with a clear no. We're fine with that.
-
A defined trigger
The work starts when something specific happens. A form is submitted. An email lands. A date passes. If nobody can say what starts the process, an agent can't either.
-
Rules you can write down
The decisions inside the process follow rules a competent colleague could put on paper. Where genuine judgement is needed, the agent hands over to a person.
-
An output someone reviews
The result is a document, a record, a decision — something a named person checks and acts on. Agents produce work for people, not activity for dashboards.
-
Volume that justifies it
It happens daily or weekly, not twice a year. Often enough that doing it by hand is a measurable cost, and often enough to prove the agent holds its standard.
-
A standard to test against
There's an agreed definition of what a good output looks like, so every output the agent produces can be graded against it.
(02)
We find the failures, not you.
Our engineers run the agent against representative samples until it consistently clears the bar. The failures get caught and fixed in our environment, before you ever see them.
(03)
Prompt engineering is real engineering.
A small change in how an agent is asked to do something can shift output accuracy by 1 to 2%. Across thousands of outputs, that compounds into something material. We test, refine, and lock. Nothing improvised.
(04)
Drift gets caught in our system, not yours.
Once live, every output is graded against the same criteria. If quality slips, we catch it before you do.
(05)
Every change is logged. Every version is reversible.
Defensible audit trail of every prompt, every model, every change. If a new version performs worse, we roll back.
In Practice
One agent, start to signature.
An engagement letter agent, the way it runs inside an accountancy firm.
(01)
The trigger
A new client is approved in the practice management system. Nobody presses anything. That's the point.
(02)
The steps
The agent pulls the right template, populates entities, services and fee schedules, and checks its own draft against the reviewer checklist agreed up front.
(03)
The boundary
A novel structure, a missing field, a judgement call — anything outside the rules stops and routes to a person. The agent screens. It doesn't decide.
(04)
The output
A drafted letter of engagement in the review queue, formatted to the template, fields complete. A partner reads it and signs it.
The team's morning becomes reading finished drafts, not assembling them. The invoice lists the letters that shipped. Nothing else.
Illustrative. Your process will differ. The shape won't.
The Model
Designed. Built. Run.
Design
Your process, mapped into a spec you'd recognise as your own.
Build
Senior engineers wire the agent into the systems you already run — CCH, Sage, Wildix, Dynamics, Salesforce.
Run
We operate it, monitor every output, and bill only when one ships.
Design, Build, Run →
The phases, the responsibilities, and where the risk sits — all on one page.
Where It Runs The sector changes.
The sector changes.
The shape doesn't.
Engagement letters in accountancy. Claims triage in insurance. Referral routing in healthcare. Different paperwork, same bounded machine underneath.
A Genuinely Free Conversation Not sure which way in?
Not sure which way in?
Start with the problem.
Bring the problem, not the category. Thirty minutes in, you'll know what an agent for it would cost, what it would deliver, and whether it's worth doing at all.