Talk to the robots
How

One process.
Six steps.

This is one animation, not six. Scroll, and watch the same process get found, mapped, designed, built, tested and put to work.

The six steps

One process moving through six stations. It starts as one of dozens of workflows, most of them full of waiting and repeated manual steps. One is circled and the rest fade. It expands to show every human touch, system, decision, delay and exception. Those touchpoints then become an agent, a rule, an API — and two stay human — inside a guardrail perimeter. Live connections and a telemetry rail appear. A hundred test cases run: 96 pass and 4 become exceptions that get looked at. Then live cases flow, with a person taking the unusual ones.

automate. measure. improve.

  1. STEP 01

    Find it.

    We identify the opportunity together.

    And the last one matters most. If it cannot be measured, we cannot show you whether it worked.

    • repetitive
    • frequent
    • slow
    • manual
    • error-prone
    • expensive
    • hard to scale
    • measurable

    Opportunity + baseline

  2. STEP 02

    Watch it.

    We map what really happens today.

    Nobody designed the process you have. It grew. Seeing it drawn is often the first time anyone has.

    • Who touches it?
    • What do they do?
    • Which systems?
    • Which decisions?
    • Where does it wait?
    • Where does it fail?
    • What are the exceptions?

    Current process map

  3. STEP 03

    Design it.

    We decide what AI should do — and what it should not.

    Some touchpoints become an agent. Some become a rule. Some become an API call. Some stay a person, because that is where the judgement is.

    • AI workflow
    • Models
    • Knowledge
    • APIs / MCP
    • ERP / CRM integrations
    • Computer use / RPA
    • Human approvals
    • Permissions
    • Guardrails
    • Fallbacks

    Automation blueprint

  4. STEP 04

    Build it.

    We implement the workflow and measure it from day one.

    Measurement is part of the build, not something added later. The telemetry rail appears at the same time as the integrations.

    • Cycle time
    • Human minutes
    • Automation rate
    • Exceptions
    • Errors
    • Cost per case
    • Model cost
    • SLA
    • Business result

    Working workflow

  5. STEP 05

    Test it.

    Normal cases. Edge cases. Bad cases.

    Test, measure, improve, repeat. A few cases will always become exceptions — the goal is to know which, not to pretend there are none.

    • instructions
    • prompts
    • model choice
    • retrieval
    • rules
    • tools
    • thresholds
    • fine-tuning, when it helps

    Tuned workflow

  6. STEP 06

    Run it.

    Production is the beginning of measurement, not the end of the project.

    Most cases flow through. Occasionally one moves sideways to a person. That is the design working, not the design failing.

    • Quality
    • Cost
    • Latency
    • Errors
    • Exceptions
    • Business KPIs
    • Model changes
    • System changes
    • Human overrides

    Production + monitoring

A complete example

One lead. Two ways of working.

Today

An inbound B2B lead arrives. A person opens six windows and does this:

  1. Someone opens the form notification.
  2. Searches for the company.
  3. Checks the country.
  4. Finds the company size.
  5. Opens the CRM.
  6. Checks for duplicates.
  7. Creates the account.
  8. Creates the contact.
  9. Decides who owns it.
  10. Writes an email.
  11. Creates a task.
  12. Updates a spreadsheet.
  13. Waits.

Automated

The same lead, as one workflow. One branch remains: a high-value or unusual case goes to sales.

Automated: Lead received, then Validate, then Enrich, then Check duplicates, then Apply rules, then Create / update CRM, then Assign owner, then Prepare response, then Create follow-up, then Track outcome. One case travels the whole chain. The AI does every step and stops at Apply rules, where a person approves — high-value or unusual case → sales. The case ends as: Ready for sales.

The AI tracks what happened

ready for sales

The robot handles the steps. Sales handles the sale.

Automate. Measure. Improve.

Then do it again, on the next process.

Show the work — every figure, with its source
  1. Process-mining tools are explicitly designed to uncover how processes actually run and to identify automation and improvement opportunities, which is why this method starts with discovery rather than with a tool.
  2. Fine-tuning is listed as one optimisation among several, not a requirement. Most of the gain in practice comes from instructions, retrieval and thresholds.
  3. The pass rate shown in step 05 stops at 96 of 100 by design. A hundred out of a hundred is not something we have seen, and showing it would undo the argument the step is making.
  4. The thirteen steps in the example are the brief's own description of an inbound B2B lead handled by hand. They illustrate the shape of the work rather than quoting one client.
Next

Start with one process.

Show us how it works today. We will find what should stay human and what can be automated.

Talk to the robotsEmail us