Workflow discovery for AI

Map the workflow.
Decide what
to automate.

We review case records with your team. You get a workflow map, approved rules and draft AI tests for your developers.

Travel & serviceCheck the booking rules before changing the customer’s flight.
Case recordsExpert reasoningAgreed rulesAI tests

Build an AI agent.
Prepare it for production.

Map one workflow, confirm its rules and give developers a brief they can implement.
Build an AI agent from a workflow

Before production, check allowed actions, test results, human handoffs and release responsibilities.
Review production readiness

Replacement example

Check compatibility
before ordering.

A customer needs a replacement part for tomorrow’s installation. The AI agent must check that the alternative meets the installation requirements before offering it.

Use the tabs to see the available records, the questions for a specialist, the rule and the draft test.

Replacement decision

What the records show

Check whether the replacement part fits.

The customer’s ticket says the replacement is needed for tomorrow’s installation. The stock record lists an available alternative, but neither record confirms whether it meets the installation requirements.

Use only the records available when the replacement was requested. What happened during the later installation cannot justify the agent’s earlier decision.
See the records and draft test

How we investigate
a workflow.

  1. 01

    Review real cases

    Review recent cases with the people who handled them. Include routine requests, rejections and cases that needed help.

  2. 02

    Map the steps and decisions

    Name the person or system at each step. Record the checks, the decision and the evidence behind it.

  3. 03

    Check the map with your team

    Have your team correct missing steps, uncertain links and conflicting accounts. Check the map against more cases.

  4. 04

    Choose what to automate

    Compare case volume, time spent, likely benefit and effort to build. Agree the rules and when a person takes over. Narrow the task or stop if automation is not worth the work.

Review exceptions.
Update the rules.

When someone steps in, ask why. Compare the case with routine work. Find the missing record, unclear rule or failed system action.

Decide what to fix and what should stay with a person. After approval, update the map, instructions and tests.

Arrange a follow-up interview

We aim to request an explanation in Slack or your agreed channel within one hour of an exception, then follow up for an answer within 24 hours. For complex cases, we aim to book an interview within 24 hours and hold it at an agreed time.

Draft AI tests
from case records.

Open case records in Claude or Codex. FieldSignal’s skills provide instructions for drafting AI tests, called evals. Review the expected actions and run the tests in your own tools.

Use case records in Claude or Codex
MCP

Connect your AI assistant to FieldSignal data.

CLI

Read FieldSignal data from your command line.

Skills

Give Claude or Codex instructions for drafting tests from a case.

Use your developers, or have FieldSignal build it under a separate agreement. Name who approves actions, measures results and maintains the automation.

Developer quickstart

What developers receive

Replacement decision
  1. Case recordsAn alternative part is available.

    The records do not confirm whether it fits the installation.

  2. Expert checkDoes it fit this installation?

    Ask a specialist to check the specification.

  3. Decision ruleGet confirmation before ordering.

    Keep the order on hold while a required check or approval is missing.

  4. Draft testRequest a specialist review before ordering.
  1. Workflow map

    Steps, people, systems and exceptions, linked to case records. Missing evidence is flagged.

  2. Expert reasoning

    What the expert checked, why they acted and what would change the decision.

  3. Decision rules

    When the agent may act, who must approve and when it must ask for help.

  4. Development brief

    Required records, allowed actions and handoff contacts. Each open question has an owner.

  5. AI tests (evals)

    Inputs, required actions and expected results. Your team reviews each test before use.

Workflow examples

Change the facts to check the next action.

Choose who runs discovery.

We run the discovery

Managed discovery

We investigate one workflow with your team. You receive a map, rules, draft tests, open questions and a recommendation on what to automate. Each open question has an owner. Your developers or FieldSignal can build it under a separate agreement.

See the discovery process
Your team runs discovery

Platform access

Your team investigates workflows for customers or internal use.

Connect through MCP or the command line. Draft tests with our Claude and Codex skills. Build and test in your own tools.

Use case records in Claude or Codex

Questions about discovery

What does FieldSignal do?

We review case records with the people who handled the work. Your process owner approves the rules. Developers receive the workflow map, rules and draft AI tests.

Managed discovery

Which records should I collect?

Collect the tickets, orders, messages and system records for one request. Ask the person who handled it what they checked and why. Compare routine cases with exceptions. Record missing evidence.

Use the workflow data checklist

Which workflows can I investigate?

Start with a repeated decision: replacing a part, approving an invoice, following up a ticket or changing an order. Check the records, approvals and exceptions. Decide whether the step needs a fixed rule, AI or a person.

Try the workflow examples

How do I find undocumented rules?

Ask an expert to explain a recent case: what they noticed, rejected and would have done differently if the facts changed. Check their explanation against the records. Have the process owner resolve conflicts and approve the rule.

Use the expert interview questions

Can FieldSignal build the automation?

Your developers can build the automation, or FieldSignal can build it under a separate agreement. Before building, agree the scope, system connections and who will maintain and run it.

Check the project responsibilities

How do I prepare an AI agent for production?

Approve the allowed actions, run tests on routine cases and exceptions, and check who handles failures. Record gaps and agree the pilot scope. FieldSignal prepares workflow evidence and draft tests; your engineering and operations owners must also complete their technical release checks.

Use the production-readiness checklist

What do developers receive?

Developers receive case records, approved rules, required checks, allowed actions, handoff contacts and draft tests. Access the data through MCP or the command line. Use FieldSignal skills in Claude or Codex to draft tests for review.

Use the workflow data

Access, schedule and data handling

Source access

List the systems and records needed for the workflow. Agree how we can access them. Use MCP or the command line to read FieldSignal data; agree any new system connections separately.

Development tools

Choose your development and testing tools. Adapt the reviewed cases to them.

Delivery schedule

We agree the schedule before starting. It depends on the workflow, available records and access to the people doing the work.

Data handling

Before sharing records, agree who may access them, which model providers may process them, where processing happens and when they will be deleted. Review current security details with us. Keep confidential records out of the booking form.

Book a workflow review.

Bring a case, its records and the person who handled it. We’ll agree what to investigate and deliver.

Agree the scope, responsibilities and price before starting.