How we work

We build systems that have to stand up to review. That governs how we work, from the first review to the upkeep.

Three principles

Financial workflows need traceability, reproducible calculations and clear approvals. We build that in from the start.

  1. 1

    Traceable from source to result

    Every value keeps its link to the source document and to the checks that led to the result.

    A figure the system has produced points to the document and the place it came from. For calculated values, the formula and its inputs are shown, so the reviewer can go straight to the source.

  2. 2

    Calculations in code

    Language models can read and structure documents. Rules and calculations run in tested, version-controlled code.

    A language model may read documents and suggest which figures belong where. Calculations run in code with tests. That is why we can show how a value was reached and reach the same value again.

  3. 3

    Human judgement where it is needed

    When something is missing, contradicts another source or calls for a decision, it goes to the right person instead of the system guessing.

    Which checks apply is decided for each process and written into the scope of work. A stopped case stays with the person responsible until it has been dealt with.

What is yours

The code sits in your own repository and in a cloud account in your name from the first day. You receive an operating manual.

Ongoing support and maintenance run under an agreement that states what is included. Our fees are separate from your costs for cloud, models and software. Maintenance can be taken over by a named third party or by your own operations partner.

Where your data sits and who can see it

The system runs in a cloud account in your name, with storage inside the EU. Data is encrypted in transit and at rest, and it is not used to train models.

Which services are involved, where logs and backups are kept and what access support has are set out in an operating description that you receive before real data is used. Only named users have access, with roles that you control.

Before real data is used, a data processing agreement and a description of the processing are drawn up for your counsel to review. What applies to regulatory compliance is set out in the agreement.

The four steps

You decide after each step. No step presupposes the next.

  1. 1

    Review

    A 45 minute conversation about how the work runs today, followed by a written assessment within five working days.

    The review is a conversation to better understand how you work today. Together we go through the different areas of your business and build a picture of what recurs. That becomes the basis for what we recommend and why.

    You decide whether to go further. The assessment is yours to keep either way.

  2. 2

    A bounded first project

    One process is built with a clear scope, a fixed price and agreed goals. Where there is uncertainty that needs resolving, we start with a prototype on example data or masked data.

    The first project covers one process. Before the work starts, we agree together on a clear plan: what is included, what we need from you, what success looks like and the timeline.

    You decide whether it should be taken into production.

  3. 3

    Implementation

    What the first project produced is built out, connected to your systems and put into production. Scope and price are set by what that project showed.

    The implementation takes its scope and its price from what the first project showed, not from an estimate made in advance.

    You decide how the system is to be looked after.

  4. 4

    Operation and further development

    An ongoing agreement for standby, maintenance and further development once the system is running.

    Once the system is running, someone has to look after it. The agreement states who does what. If you already have an operations partner, we hand over with complete documentation.

What we do not do

  • We do not connect to your systems before there is a contract for the implementation. A prototype runs on example data or masked data.
  • We do not describe a prototype on example data as a system in production.
  • We do not let a language model calculate. Calculations run in code that can be tested.

How we handle data, calculations and approvals

The same thing on one page, to pass on to the credit committee, the risk function, the board or the auditor.

How we handle data, calculations and approvals pdf