Each one is shown the way it arrives, on the services page. Most clients start with one agent on one painful workflow, then expand once the numbers hold.
Invoice · scanned
total 14,976.00 · .98
Ticket #48213 · closed
Charged twice for my August plan?
Refunded the second charge of €29.00.
confidence 0.93
Assistant · HR policy
Can I carry leave over?
Up to 5 days, if approved before 15 Dec.1
2 sources
Workflow run #8,127
DocuSign completed 0 ms Extract fields 1.8 s NetSuite customer 0.9 s Post to #sales-won 0.2 s
Business plan · 8 weeks
approved by analyst
Proof of delivery
confirmed
Prototype · 14 screens
2 changes after build
Accuracy goes in the contract
Every agent we ship is accepted against a written test set and a service level. If it doesn't hit the numbers, it isn't done.
01p95 latency
Time to a complete, grounded answer, measured in production, not in a notebook.
02Accuracy
Scored against a labelled set of real cases your own experts agreed on.
03Grounding
Share of claims traceable to a real source document, log line or record.
04Escalation rate
How often the agent correctly hands a case to a human instead of guessing.
Latency distribution
p95 target · under 15s
p95 means 95 of every 100 answers land left of that line. The shape here is illustrative. The threshold is the number we sign.
The artefact
This is what you get signed
Every build ends with an acceptance report against your own labelled test set. If a row says FAIL, the agent is not accepted and we keep working. Below is the format that document takes, filled with representative figures from an incident-triage engagement.
Acceptance report
Incident triage agent
Specimen
Client
Test set
labelled incidents
Scored by
Client domain experts
Agreed targets against measured results
Metric
Target
Measured
Result
p95 latency
< 15s
11.4s
Pass
Accuracy
≥ 0.75
0.78
Pass
Grounding
≥ 0.80
0.82
Pass
Escalation rate
≤ 0.15
0.11
Pass
Targets are agreed during the first week, before a line of code is written.
Our open sample agents each publish a real report in this format, produced by their test on made-up data: see the finance agent's report.
~99% precision on banking details>95% recall across entity types10,000 tickets a day, sorted60 / 30 / 10 closed, routed, held<15s p95 to a grounded answer0.78 accuracy on the labelled set0 engineering tickets for routine charts~99% precision on banking details>95% recall across entity types10,000 tickets a day, sorted60 / 30 / 10 closed, routed, held<15s p95 to a grounded answer0.78 accuracy on the labelled set0 engineering tickets for routine charts
01/ 04
Four agents, accepted on their numbers
Scroll to move through them, or skip ahead.Swipe sideways to move through them.
Each one has a working sample on GitHub that runs on made-up data, with its own acceptance test you can re-run.
01
Regulated finance
One wrong digit is unacceptable
Risk analysis over contracts, reports and transactions, pulling out counterparties, dates, amounts and banking details from a mix of clean text and scans, across inconsistent legal formats and synonyms.
Hybrid design: deterministic rules for structured fields
Fine-tuned transformer for organisations and document IDs
Banking details validated against official directories
precision on banking details, where errors are unacceptable
>95%
recall across the tracked entity types
-90%
target reduction in document processing time
API+UI
integrated into the existing back office
02
Customer support
10,000 tickets a day, sorted
The challenge was never volume alone, it was sorting it. Closing the routine cases automatically is only safe if the agent also knows, reliably, which cases it must not touch.
Answers grounded in your help centre and past tickets
Confidence thresholds that decide close vs. escalate
Clean handover, the specialist sees what the agent found
Every escalation logged and scored, so the split improves
A data team running 200+ ETL pipelines. Every failure cost an engineer 30 to 45 minutes of manual triage across Grafana, ClickHouse logs, Confluence and runbooks, before any fixing started.
Agent triggered by the monitoring alert itself
Retrieval across logs, metrics, runbooks and past incidents
A BI platform with 50+ dashboards. Every "can you add a chart for this segment?" became a two-day ticket for data engineering, and there were more of them every week.
Natural-language to SQL over the live warehouse schema
Handles filters, time windows and period-over-period growth
to answer a new question, down from a two-day ticket
50+
dashboards the agent can reason across
0
engineering tickets for routine chart requests
Self
analysts serve their own questions
00/ 04
First call to production in weeks
01
Audit
1 week
We pick the workflow, build the test set with your experts, and agree the SLA and the price of the build.
02
Build
4 to 12 weeks
Weekly demos on your real data. Accuracy tracked against the test set from day one.
03
Accept
1 week
Go-live only when the agreed numbers are met. Your team trained, everything documented.
04
Run
ongoing
Monitoring, drift checks, retraining and new use cases on a monthly retainer.
You can stop after any stage. Nothing is locked in.
Trust
Your data stays yours
GDPR, signed DPA, EU hosting or a fully self-hosted deployment.
You own the code
Full IP transfer on final payment. No lock-in, no black boxes.
Accuracy is measured
Labelled test sets and a written acceptance report, every time.
Humans stay in the loop
Every action that carries risk waits for a person to approve it.
How pricing works.
It starts with a free call. The first week after that is paid and ends with a test, a plan and a quote for the build. Every build is quoted against the job we agreed, once we have seen your real documents or requests, so the number you get is about your work, not a line from a price list.
Start here
A call
30 minutes
We find out what is wasting your team's time and whether an agent is the right fix. No preparation, no obligation.
Free
The first week
1 week
The test, a plan and a quote. It comes off your build if you go ahead.
Credited
An agent
6 to 10 weeks
One job, live and passing the test.
Quoted
Paperwork
6 to 10 weeks
Pulling details out of documents.
Quoted
An answer assistant
5 to 7 weeks
Answers from your own documents.
Quoted
The plumbing
4 to 6 weeks
Pipelines, integrations, automations.
Quoted
Keeping it running
monthly
We watch it and keep it accurate, or your team takes it over.
Optional
Hosting and model costs are billed at what they cost us. Nothing is locked in: you can stop after any stage.
Let's measure it
It starts with a free 30-minute call. If there is a job worth doing, we start with a paid first week that comes off the build.