If the pilot fails
You keep everything: the process map, the blueprint, the evaluation set and the code. There is no proprietary lock-in to walk away from, and no exit fee.
How we work
Five stages from “this is broken” to operated.
No six-month discovery theatre. We map the process, write the blueprint, prove it on your data behind a human review queue, then harden it into production — with an exit plan you own.
Stage by stage
Deliverables, durations,<br />and what we need from you.
Each stage ends with something concrete you keep — a process map, a blueprint, a working pilot, a runbook — whether or not you continue to the next one.
- 01
Discovery
3–5 days
We sit with the people doing the work today, map the process as it actually runs — not as the SOP claims — and find where the hours and the errors are.
- Process map with volumes and cycle times
- Baseline metrics to beat
- Risk and data-access review
- 02
Blueprint
1 week
A written plan you can argue with: architecture, model choice, integration points, human-review design, evaluation criteria and a cost model per transaction.
- Solution blueprint
- Evaluation plan and target accuracy
- Fixed-scope pilot proposal
- 03
Pilot
2–4 weeks
A working system on your real data behind a human review queue. Small blast radius, real numbers, no slideware.
- Working pilot on production-shaped data
- Measured accuracy and cost per item
- Go / no-go recommendation with evidence
- 04
Production
4–10 weeks
Hardening: throughput, failure handling, permissions, monitoring, audit trail, runbooks and training for the team that will own it.
- Production deployment in your cloud
- Monitoring and alerting
- Runbook and operator training
- 05
Operate
Ongoing
We keep it working: evaluation runs on every model or prompt change, drift review, cost tuning, and a monthly session on what to automate next.
- Continuous evaluation and regression tests
- Monthly quality and cost report
- Quarterly roadmap of new automations
What we need from you
Four things, and the project moves fast.
- One process owner who can make decisions without a committee
- Read access to the systems in scope (sandbox is fine to start)
- A sample of real documents or tickets — a few hundred beats a few thousand
- A named reviewer from the team that does the work today
The arithmetic
Eleven steps in.
Three steps out.
This is where the value sits: not in a clever model, but in the number of times a person touches an item before it is finished.
Manual steps removed, exceptions kept human. Nothing with a financial consequence posts without either a match or a signature.
before · manual
11 steps- 01 Invoice arrives in a shared inbox
- 02 Someone opens the PDF and reads it
- 03 Keys 14 fields into the ERP
- 04 Checks the PO in a second system
- 05 Emails the warehouse about a mismatch
- 06 Waits for a reply
- 07 Notes the exception in a spreadsheet
- 08 Re-keys the correction
- 09 Files the PDF in a folder
- 10 Month-end: rebuilds the list by hand
- 11 Closes late again
after · automated
3 steps- 01 Invoice arrives — parsed, matched and posted
- 02 Only exceptions reach a human, with the mismatch shown
- 03 Every action logged, reconciled and reportable
Same controls, same audit trail, same month-end — with the keying removed and the exceptions collected in one queue instead of eleven inboxes.
How the routing actually works
Humans where it matters,
by arithmetic.
Every step has a threshold, tuned on your own evaluation set. Above it the action executes; below it the item goes to a person with the evidence attached. That is the whole safety story — not a disclaimer, a number.
routing rule
route(x) = HUMAN if p̂(x) < θ
AGENT otherwise - p̂(x)
- calibrated model confidence for step x, not the raw softmax score
- θ
- threshold tuned on your evaluation set — typically 0.82–0.94
- HUMAN
- the item lands in the review console with its evidence attached
- AGENT
- the action executes and is written to the audit log
Thresholds are per step, not per system: a wrong supplier name is cheap to fix, a wrong payment is not.
If you want to stop
Monthly engagements run on 30 days’ notice. We hand over repositories, credentials you own, and a written runbook so your team can continue.
If it works
We extend stage by stage — more document types, more steps, more integrations — each one with its own measured target rather than one giant rollout.
Next step
Start with the map, not the model.
Discovery is a fixed fee and takes three to five days. You will know what to automate, what it is worth, and what we would build first.
One business day — and the reply comes from an engineer, not a sales sequence