How it works
What sits behind a reliable agent
A demo takes an afternoon. An agent you can leave running takes four layers of engineering, a test set built from your own cases and someone watching the numbers every month. Here is what that looks like.
Four layers
Every agent we build has the same anatomy
Control: who stays in charge
The layer that makes the rest safe to switch on. It decides what needs a person, records every step and lets you stop the agent with one click.
- Approval thresholds for money, data and anything irreversible
- A full log of every message, decision and action
- Live monitoring with alerts when quality slips
- A kill switch your team controls
Action: what it's allowed to do
Connections to your real systems, each one scoped to the smallest set of permissions the job needs. This is where an agent stops talking and starts working.
- Read and update orders, bookings, tickets and records
- Send messages, emails and documents
- Separate credentials for each tool, never shared logins
- Write access only where the process demands it
Reasoning: how it decides
The language model plus the instructions, rules and examples that turn it into your agent. We pick the model by testing several on your cases, and often use more than one.
- Your policies written as rules it follows, step by step
- A cheaper, faster model for simple steps, a stronger one for hard ones
- Plans multi-step tasks and checks its own work
- Models can be swapped without a rebuild
Knowledge: what it knows
Your documents, product data and policies, retrieved fresh at the moment of each question. Nothing is baked into the model, so updating a policy takes minutes.
- Connected to the systems where your knowledge already lives
- Kept current automatically as documents change
- Respects who is allowed to see what
- Every answer traceable to its source
Testing
No customer meets it until it passes the exam
Before launch, we build a test set from hundreds of your real conversations or documents, with the right answer agreed by your team. The agent is scored against it every time we change anything.
You set the pass mark. It goes live when it clears that mark, and not a day before. After launch, the same tests run every week, so a model update can't quietly make things worse.
The engagement
Four phases, each with a clear exit
You can stop after any phase and keep everything we have produced. Most clients don't, but the option keeps both of us honest.
01 · MAP · 1 TO 2 WEEKS
Find the work worth automating
We interview your team, measure where the hours go and rank the candidates by value and difficulty.
A ranked shortlist, a business case for the top pick and a fixed quote for the pilot.
02 · PROVE · 4 TO 6 WEEKS
Run it on real data
We build the agent, test it against your cases and run it alongside your team, or on a small slice of live traffic.
A working agent, its test results against the target and an honest go or no-go recommendation.
03 · LAUNCH · IN STAGES
Roll out with a safety net
Traffic moves over gradually, with people approving more at first and less as trust builds. Your team is trained to supervise it.
A production agent, dashboards, runbooks and a team that knows how to manage it.
04 · RUN · MONTHLY
Keep it sharp
We monitor quality, review failures, update knowledge and retest when models change. Then we look for the next process.
A monthly report against the agreed number, and improvements shipped every month.
Pricing
No surprises on the invoice
Every quote is specific to the process, but the structure is always the same. You know the cost of each phase before it starts.
Fixed price
One agreed fee per phase, quoted after the first call. If the pilot misses its target, you don't pay to scale something that doesn't work.
Monthly retainer
Covers monitoring, support, improvements and the monthly report. Sized to the number of agents and the volume they handle.
At cost, with a cap
Model and hosting fees billed at cost or straight to your own accounts. Estimated up front, capped, and reported every month.
Model agnostic
We don't resell anyone's model
No partnership quotas and no favourite vendor. We test the leading models on your cases and use whatever performs best for the price. When something better ships, we retest and switch.
- OpenAI GPT
- Anthropic Claude
- Google Gemini
- Meta Llama
- Mistral
- Open models in your cloud
Your move
Bring us the job nobody wants.
Thirty minutes, one process, and a straight answer on whether an agent should take it on. No slides and no obligation.