Skip to content

Practice 02 · Intelligent Systems & AI Infrastructure

Give your workflow
a clearer next step.

Knowledge, tools and human review. Connected where work happens.

Follow one enquiry through the system.

A fictional enquiry, two sample documents and a prewritten reply. It all runs in your browser. Nothing is sent.

Interactive demo · Sample data
Example
  1. EnquiryReceive and read, not started
  2. RetrieveFind relevant notes, not started
  3. PrepareDraft from the notes, not started
  4. ReviewA person decides, not started

Ready. Run the demo to follow the enquiry.

Enquiry

Received 09:14
From
Alex R., Kestrel Field Supplies (fictional)
Subject
Customer portal and our CRM

Hi team,

We are planning a customer portal where our clients can see their orders and open requests. Every account already lives in our CRM.

Can the portal connect to it, or would we need to move our data first?

Thanks, Alex

Retrieved excerpts

Nothing retrieved yet. Excerpts appear at the Retrieve step.

Prepared response

Not started

No draft yet. The reply is prepared from the retrieved excerpts.

Human review

Review opens when a draft is ready.

Fictional people, companies and documents. The reply is prewritten; no model or external service is called.

See why the system answered.

The exact excerpts behind the draft, and the steps that produced it.

No sources used yet. Run the demo to see the excerpts the draft cites.

Capabilities

From the demo to your workflow.

The demo above runs in your browser on fictional fixtures: no model, no CRM, nothing sent. A production workflow adds what a demo cannot show: your documents and tools, evaluation on agreed examples, a trace of every run and review steps your team approves.

Use your knowledge

Retrieval over your own documents, so an answer arrives with the passage it came from.

  • Answers buried in documents

    Contracts, manuals, menus and past emails hold the answer. Retrieval finds the passage and cites it, so a person can check it.

  • Retrieval and indexing

    Search and indexing over the documents you choose, built into the workflow rather than a separate chat window.

Connect your tools

Adapters for the CRM, inbox, calendar and internal APIs your team already opens.

  • Information retyped between systems

    An enquiry lands in an inbox, then a spreadsheet, then the CRM. A reviewed workflow files it once and prepares the next step.

  • A CRM someone keeps current

    Enrichment, summaries and follow-up reminders are added to the CRM you already run, with a person approving what matters.

  • Inside the tools you have

    Adapters for CRM, ticketing, calendar and internal APIs, so nobody has to learn a new tool to use the workflow.

Keep people in control

Evaluation before release, review where it matters, and a trace for every run.

  • Evaluation first

    We write the evaluation set before the prompt. It runs on every change, and releases wait on the result.

  • Human review

    Review screens for the people who approve the work, placed at the steps that matter.

  • Traces and rollback

    Every run leaves a trace, and a change can be rolled back. Models can be swapped without rewriting the product.

Ownership

What your team receives.

What you own

The evaluation set, the prompts and the code are yours. Models can be swapped without rewriting the product, and the review steps are documented so a person always knows where to look.

  1. Workflow implementation

    The workflow codebase, typed, instrumented and versioned, with adapters for your tools.

  2. Evaluation examples

    A set of agreed examples and a scoring rubric that every change runs against.

  3. Review and rollback controls

    Review screens for the people who approve, and a documented way to roll a change back.

  4. Operator documentation

    Where to look, what each review step means, and what to do when a run needs attention.

Relevant work

Internal R&D · not client work

Atlas Intelligence

No client AI workflow is in production yet. The agent work we can show today is internal: Atlas, the CRM with enrichment and lead prioritisation that we run the studio on.

See studio R&D

Questions

We write the evaluation first: the failure modes and the scoring rubric come before the prompt. It runs on every change, releases wait on the result, every run leaves a trace, and a person reviews the steps that matter.

Yes, that is the usual case. We write adapters for your CRM, ticketing, calendar and internal APIs, so the workflow sits inside the tools your team already opens.

No. It is a deterministic demo with a fictional enquiry and fictional documents, built to show the sequence: retrieve, prepare, review. A production workflow runs on your documents and tools, with evaluation and review in place.

An AI Discovery Sprint takes two weeks at a fixed price of CHF 3,500. A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements. A production build is priced on the systems it touches, in one proposal with a fixed total, after the first call. Single-surface work typically costs CHF 8,000 to 12,000; larger systems CHF 15,000 to 25,000, with a ceiling of CHF 45,000.

Bring us the workflow that keeps getting retyped.

Describe where the information starts, where it ends up and who checks it. That is enough for a first call.

Discuss your workflowExplore the system