Investigate the workflow, its data and its exceptions

Discovery creates a reliable account of how work is done today. A process chart from a manager is a starting hypothesis. Actual tasks reveal copied spreadsheets, informal approvals, missing fields, duplicate entry and the judgment that makes an exception safe to resolve.

Follow the complete task

Observe a representative operator from trigger to accepted business outcome. Capture active work separately from elapsed waiting. Record inputs, identity, applications, decision rules, interruptions, handoffs, rework and the final source of truth. Ask the operator to explain what would make them distrust an output.

For an initial discovery allocation, observe 12 tasks across three operators: six routine tasks, three frequent exceptions, two rare but consequential cases and one failed or abandoned task. This allocation reveals mechanisms; it does not estimate a population error rate. Expand observation until the important workflow variations and unresolved decisions are visible.

The total is eight minutes of active handling plus waiting. Removing four minutes of handling will not automatically remove an 18-hour approval delay. Conversely, an approval-routing change may improve elapsed time without changing active effort. Report both.

This emphasis on the progress people seek is informed by Jobs to Be Done. FIELD turns it into observation records and explicit acceptance conditions for a particular workflow.

Map exceptions before adding autonomy

Build an exception taxonomy: missing input, conflicting records, stale information, unsupported language, uncertain identity, insufficient authority, dependency unavailable, ambiguous external outcome and user cancellation. For each, define who resolves it, how much time it takes, and where the task waits.

An escalation is a designed path with a receipt, a queue and an owner. “Ask a human” is incomplete if the human cannot see the evidence or has no capacity to respond. Estimate the queue load from arrival rate and exception frequency, then inspect peak periods and staffing.

Inspect the data contract

For every source, record:

• Owner, permitted purposes, licence or contractual rights, and approved processing locations.
• Authoritative identifiers, event time, ingestion time, revision, amendment and deletion semantics.
• Completeness, missingness, duplicates, schema variability, representative languages and sensitive fields.
• Authorization rules, revocation latency, export limits, retention, deletion and derived-data obligations.
• Update frequency, API quota, pagination, bulk export, replay and reconciliation support.

Separate source truth from generated interpretation. An embedding, summary or model assertion must retain provenance and must not override a current authoritative record. A deletion plan must include caches, derived summaries, indexes, backups and documented retention exceptions.

Prove the difficult boundary early

Build a small vertical slice that crosses identity, network, data and final workflow output. A useful early probe could read one permitted record through the real proxy, deny another user, process a deletion and show the accepted result in the user’s existing application.

Record the actual API behavior on timeout and retries. Verify whether a successful HTTP response means the operation committed, was merely queued or can still fail. Test schema evolution and expired credentials. An integration that works on a developer laptop has not proven the customer’s runtime environment.

Use a dependency board with a real critical path. If security review takes ten business days and access cannot start before it, coding in parallel does not remove that wait. Find independent work, bring the dependency forward or adjust the commitment.

Distinguish fact, assumption and target

A discovery register needs an identifier, statement, classification, source, owner, date and next action. “The customer permits outbound model requests” is an assumption until the responsible owner approves the specific payload and route. “p95 under three seconds” is a target. “p95 was 2.7 seconds on this configuration and workload” is a measurement with a test record.

Use value-stream mapping to expose handoffs and delay, informed by DORA’s guidance. FIELD extends the same attention to the customer’s operational task. The output of investigation is an agreed first journey, the risks most likely to invalidate it and a test plan that addresses those risks before large commitments.