How we work

Written, async, fixed scope. A person approves every action.

No calls, no open-ended projects. You describe the work in writing, we agree one fixed scope, and the agent we build never acts on your accounts without your approval.

Start in writing

01 / AGREE

Fix the scope in writing

One workflow, one inbox or source, the document layouts covered, the fee and the pass/fail tests. All agreed before you pay.

See services

02 / BUILD

Test before you rely on it

We measure field-level accuracy and cost per document on an agreed evaluation set before handover. You see the results, including the misses.

Evaluation method

03 / RUN

Your team stays in charge

Staff approve, edit or reject every item in a review queue. You get a written runbook, and Agent Care if you want ongoing fixes.

See Order Intake

Human approval is not optional

Our agents prepare work; people decide. Nothing is sent to customers and nothing is written to your accounting system until someone on your team approves it. Uncertain or unmatched items go to an exception queue instead of being guessed.

Your accounts, least-privilege access

You own every account the agent uses. We ask only for the access the workflow needs and document it in the runbook. You can remove that access at any time.

Data handling

Documents are processed through the Claude API, with Anthropic as our sub-processor. We do not train models on your documents. Retention and deletion are agreed in writing for each engagement; trial documents are deleted within 30 days. See the privacy notice.

What counts as done

The agreed workflow runs from the agreed inbox or source to the review queue and export.
The agreed evaluation set has been run and the results are shared in writing.
Every action that changes data outside the agent requires approval.
You have the runbook: how it works, how to pause it, who to contact.

Communication

Everything happens in writing: scope, questions, progress, results and support. That keeps a clear record for both sides and lets us work across time zones.

Demonstrations and claims

We are a new studio. We show no testimonials, no customer logos and no results we have not measured. Samples are fictional and labelled as such. They demonstrate a method; they are not customer outcomes or a guarantee of business results.