At 10:14 on a Tuesday, a founder watches a prospect abandon the product at the CSV import screen and reopen the spreadsheet they promised to replace. The team has spent 16 weeks on roles, notifications, Stripe billing, and a dashboard. None of it matters if the prospect cannot get a trustworthy result from the file they already use. That is the MVP test: does a real person cross the point where their existing workaround stops being easier?
A first release is not evidence because it reached production. It is evidence when it tests a named belief with behavior that can prove it wrong. “Operations teams will pay for automated invoice matching” becomes: “After seeing 50 invoices matched against purchase orders with fewer than three corrections, an accounts-payable manager will run the next weekly batch without concierge help.” Build only what is required to observe that behavior.
Start with one expensive problem
The first workshop should end with four fields, not a brainstorm board: user, trigger, current workaround, and measurable cost. “Small businesses need inventory management” says nothing useful. “A three-location independent pharmacy spends 90 to 120 minutes each weekday comparing two distributor exports before placing its 10:30 a.m. order” identifies the operator, deadline, source files, and economic damage. You can now test whether reducing that task changes purchasing behavior.
Next, name the assumption that would make the business nonviable. A marketplace may fail because suppliers answer fewer than 40% of buyer requests within 24 hours. A document-analysis product may fail because staff must verify more than one answer in ten against the original contract. A scheduling product may fail because coordinators will not grant calendar access. Teams avoid these questions because they are awkward; then they spend a quarter proving that users can click buttons.
Draw the smallest complete workflow
An MVP needs one path that finishes a job. For a freight-quoting product: upload a carrier rate sheet, select origin and destination, calculate a quote, flag exceptions, and send the result. A beautiful account page is irrelevant until a dispatcher can create a quote that survives an actual customer conversation. If the workflow stops before the outcome, the team is testing interface preference rather than product value.
Map the path as states, inputs, decisions, and exits before selecting tools. Specify what happens when the sheet has duplicate lane names, a required surcharge is blank, the address cannot be geocoded, or a customer returns three days later to revise a quote. Identify actions that require approval and actions that cannot be undone. These are not edge cases in B2B software; they are where trust breaks. A product that handles only the happy-path demo creates support tickets on day one.
Use the right amount of engineering
Speed does not require disposable code in every layer. For many early B2B products, a web client, API, relational database, object storage, job queue, and managed identity provider cover the first paid workflow. Splitting that into five services creates five deployment paths, five places for authorization drift, and a debugging trail that starts with “which log stream has it?” Keep the system boring until load, team boundaries, or regulatory constraints provide a measured reason to split it.
Some foundations cannot wait. Put organization and tenant identifiers on records before a second company enters the product; retrofitting isolation after data mixes is a migration and incident problem. Write audit events for exports, approvals, permission changes, and payment actions. Make queued work idempotent so a retry does not send two invoices or create two orders. For LLM features, store the source passage IDs, prompt revision, model ID, output, and reviewer correction. Without that trail, an incorrect answer becomes an anecdote instead of a debuggable defect.
Instrument the decision, not vanity
Define the events before launch and connect each one to the assumption under test. For the pharmacy workflow, measure: distributor file uploaded, reconciliation completed, exception count resolved, order exported, and a second weekly run completed without staff assistance. Add timestamps so you can see whether the task fell from 105 minutes to 18, not merely whether someone visited the page. Signups, sessions, and page views can explain acquisition; they cannot prove that a habit changed.
Pair event data with observed sessions. Watch users import their own files, keep a tally of where they hesitate, and record the work they perform outside the product. “This looks great” after a founder-guided call has almost no predictive value. A buyer who asks to add two coworkers, forwards an output to a supplier, or schedules next week’s run has incurred a cost to continue. Ask what happens if the product disappears tomorrow. “We would return to the spreadsheet” is useful only when you learn which tab, formula, and manual check they would restore.
Treat the first release as a fork
Review the release against the original threshold, not against the backlog. If users complete the workflow but fewer than 25% return within seven days, inspect whether the job occurs weekly, monthly, or only during emergencies; a retention problem may be a frequency mismatch. If users abandon before their first result, inspect identity setup, integrations, file quality, and permission errors before repainting the interface. If three customers tolerate obvious roughness and keep returning, trace their exact sequence of actions. Repeated behavior is more useful than feature requests detached from use.
The release should force a decision: deepen the workflow for the users who returned, shift to a better-defined customer, change the commercial model, or stop funding the premise. “Build version two” is not a decision. A startup with six months of runway cannot afford four months of motion that leaves the central belief untouched. Build the smallest product that can produce a credible yes or no, then spend the next quarter on the answer rather than the original idea.