The One Rupee Test: Define Evidence of Completed AI Work

Editorial update: 14 November 2026. Unverified historical outcomes have been removed; this version presents practical planning guidance.

An AI project can produce messages, screens and passing checks without completing the business task it was intended to support. The useful question is what evidence would demonstrate one accepted piece of work.

The One Rupee Test is a proposed planning metaphor for that smallest meaningful outcome. This version does not present an unverified personal audit story or claim that a particular number of tests proved a deployment had failed.

Define the outcome before the activity

Write down the business task and its accepted finish. A drafted reminder, a sent reminder and a recorded customer response are different outcomes. Choose the one the project is actually responsible for.

Identify the destination where completion can be checked. That may be an approved entry, a saved document, a task record or a reconciled report. A summary saying the work is done is not enough by itself.

Separate the doer from the check

Ask the reviewer to inspect the relevant source and destination records. Verify that the result belongs to the intended customer, reporting period or transaction.

The reviewer should understand what the test does and does not cover. Checking one accepted result does not establish that all inputs, permissions or failure cases work.

Test the boundaries as well

Include a missing input, a duplicate, a conflicting instruction and an unavailable dependency where relevant. Decide what the workflow should do in each case. Stopping for clarification can be the correct result.

Keep external actions within the permissions agreed for the test. Use a controlled environment or authorised sample where an accidental action could affect a customer or business record.

Record evidence in a small completion card

Include: task definition, source references, intended result, actual result, check performed, reviewer, date and unresolved limitations. Distinguish a simulated action from one completed in the actual destination.

Retain useful test results, but interpret them alongside the business evidence. A software test can establish a property of the code it exercised without proving that the entire operating process is ready.

Decide the next step

If the smallest task succeeds, test representative variations and measure the human effort required to operate it. If it fails, identify whether the problem is data, instructions, permissions, integration or scope.

Do not expand access or automate more tasks solely because a demonstration looked convincing. Equally, do not discard useful progress because a limited test exposed a problem that can be corrected.

The goal is a clear account of what works, under which conditions and with what remaining uncertainty.

For help defining and testing a supervised workflow, explore Bizgenix custom AI solutions.