A Product From Problem to Operation
Follow a hypothetical product from customer discovery through a limited release and the decision to continue.
On this page
A product team wants to reduce the time a small finance team spends preparing invoices for review. The following case is fictional. It follows one set of decisions across the handbook; the numbers are illustrative and do not represent a measured result.
Find the work behind the request
The finance team asks for automatic invoice processing. Observation reveals several steps: opening an invoice, entering fields, checking totals against supporting information and resolving exceptions. Some invoices are straightforward. Others contain missing details or amounts that need a conversation with a supplier.
The product team chooses a narrower outcome: reduce preparation time for routine invoices while preserving the reviewer's ability to identify errors. Payment approval stays with the finance team. The group records the current preparation and review time, errors found and how long exceptions wait. It also records that this baseline comes from a small group and may not represent other teams.
The next question is whether preparation is the bottleneck. If the time is mostly spent waiting for suppliers, a drafting tool may help less than a better exception process. The team looks at recent work and interviews reviewers before choosing what to build. Customer understanding explains the investigation.
Compare a small set of options
The team considers a structured form with clearer validation, a tool that extracts fields for review, and an agent that prepares and approves invoices. The form is simpler but still requires manual entry. Extraction may save that entry while keeping approval with a person. Automatic approval would require a much larger authority and risk decision.
The team chooses to test extraction on a narrow invoice type. It does not commit to automation for every invoice. The next decision is whether reviewers can check the proposed fields accurately and with less effort than the current process. The discovery chapter describes how to frame the bet.
Prototype and define quality
A prototype shows each extracted field alongside the source passage. It uses permitted sample documents and does not connect to live payment systems. Reviewers try it and point out that a correct-looking total is hard to trust when the corresponding line items are missing. The team adds a visible exception state and asks a finance specialist to define which cases require manual handling.
The eval includes ordinary invoices, missing fields and conflicting totals. The team grades extraction separately from exception handling: copying the printed total correctly is not a pass if a material discrepancy goes unflagged. The detailed invoice evaluation shows how reviewers can disagree and improve the rubric.
Release within a boundary
Suppose the ordinary cases meet the team's proposed quality bar, but a conflicting-total case still fails. The team limits the pilot to a narrower invoice type, with a person screening exceptions before processing and reviewing every result. It names an owner for source data and the eval, and another for system access and recovery. No payment is made by the tool.
The result supports a limited learning decision, not a claim that every invoice will be handled correctly. The team tells reviewers what the tool can do, how to report a failure and where manual work continues. Delivery covers the release evidence and operating responsibilities.
Follow the work after release
Imagine the first pilot shows that entering fields is quicker, but reviewers spend extra time finding the source of a few extracted amounts. Some people return to the old process for those invoices. The team checks whether the source links fail, whether the documents differ from the test set and whether the review instructions are clear.
It improves the source display, adds a verified failure case to the regression set and compares the full preparation and review time with the baseline. It also checks errors and exception waiting time. If the workflow does not improve the outcome at an acceptable quality and workload, the team can narrow or stop it. If the evidence supports expansion, the next group still needs its own access, support and evaluation of differences in its invoices.
This is a product decision as well as a technical one. A faster extraction step has value only if the people doing the work can use and check it. Measuring the whole workflow describes how to connect the observations to a decision.
v3.1 · Updated Oct 2026