testology.io logotestology.io
← All case studies

Case study · AI QA and API testing

Making a confidential AI platform safer to ship

A fast-moving AI startup needed practical QA support without slowing product delivery. testology combined API automation, manual testing, regression checks, and AI-focused review so the team could keep releasing with clearer evidence and fewer unknowns.

Client
Confidential AI platform
Engagement
Freelance QA support, 10 to 20 hours per month
Focus
API testing, manual testing, regression, AI workflow assurance
Delivery
BDD scenarios, automated checks, reproducible defect reports

The challenge

The client was moving quickly, adding new API-backed features and AI-assisted workflows in a product where trust matters. They needed flexible QA help that could fit around a startup cadence: enough structure to protect releases, but not a heavy process that got in the way.

APIs carried the real product logic

Many important risks lived below the interface: permissions, response shape, data updates, error handling, and the way services behaved under ordinary user journeys. UI checks alone would not have been enough.

AI behaviour needed human judgement

For the AI implementation, the work could not rely on exact string matching. testology reviewed whether outputs were useful, consistent with the workflow, and safe enough for the context in which they appeared.

What testology delivered

The work combined repeatable automation with hands-on exploratory testing. Automated checks covered the stable, high-value paths; manual testing covered the places where product judgement, AI behaviour, and user experience still needed human eyes.

98
BDD scenarios reviewed across the test assets
187
API interactions exercised by automated scenarios
86
explicit response-time checks in the suite
15
negative or permission-focused checks
11
schema and contract-style validations

Scenario coverage by area

BDD scenarios were spread across core workflows, platform behaviour, access control, AI-facing flows, and supporting services. Areas are anonymised.

Scenario coverage by anonymised areaCore workflows: 25; Platform settings: 21; User/admin controls: 18; Auth & permissions: 12; Notifications: 12; AI output checks: 6; Live updates: 3; Specialist service: 1Core workflows: 25Core workflows25Platform settings: 21Platform settings21User/admin controls: 18User/admin controls18Auth & permissions: 12Auth & permissions12Notifications: 12Notifications12AI output checks: 6AI output checks6Live updates: 3Live updates3Specialist service: 1Specialist service10510152025

API methods exercised

The automation covered read, create, update, patch, and delete behaviour rather than checking only happy-path responses.

API calls by HTTP methodGET: 65; POST: 58; DELETE: 35; PUT: 16; PATCH: 13010203040506070GET: 6565GETPOST: 5858POSTDELETE: 3535DELETEPUT: 1616PUTPATCH: 1313PATCH

Quality signals built into the suite

Checks were not limited to “does it return 200?”. The suite also captured speed, contract shape, and failure behaviour.

Quality signal mixExecutable scenarios: 98; Response-time checks: 86; Negative checks: 15; Schema validations: 11Executable scenarios: 98Executable scenarios98Response-time checks: 86Response-time checks86Negative checks: 15Negative checks15Schema validations: 11Schema validations11020406080100

Automation and human judgement

The engagement balanced fast, repeatable checks with manual testing where AI behaviour and user experience needed context.

Testing approach balanceIndicative share of effort across API automation, Regression testing, Manual exploratory testing, AI behaviour review. The split is illustrative, not timesheet data.API automationRegression testingManual exploratory testingAI behaviour reviewQAworkstreams
  • API automation
  • Regression testing
  • Manual exploratory testing
  • AI behaviour review

The approach

1

Map the risk

Identify the workflows most likely to damage trust if they broke: access, data changes, service responses, AI-assisted outputs, and high-use user journeys.

2

Automate the stable checks

Build BDD-style API scenarios around repeatable behaviours, including response status, response body, schema shape, and basic performance expectations.

3

Test the AI manually

Review AI implementation behaviour with human judgement: whether the feature produced useful outputs, handled edge cases, and stayed aligned with the intended workflow.

4

Feed releases with evidence

Document bugs clearly, retest fixes, and keep regression checks current so the team could move quickly without losing sight of quality risk.

The outcome

The client gained a QA layer that matched the pace of a high-growth startup: lean enough to run part-time, structured enough to protect important releases, and practical enough for engineers to act on. Part-time has limits, though. At 10 to 20 hours a month not every release got a full pass, so the hours went to the flows most likely to hurt if they broke.

The goal was not to create a bigger test suite. It was to create a test suite the team could trust.

More reliable API releases

Repeatable checks helped catch broken contracts, unexpected response changes, permission issues, and regression risks earlier in the delivery cycle.

Clearer AI quality feedback

Manual review added judgement where automation alone was too brittle, especially around AI outputs, edge cases, and feature behaviour.

Better developer handoff

Defects were documented with enough context to reproduce, prioritise, fix, and retest without long back-and-forth conversations.

Building AI features into your product?

We test the APIs, workflows, prompts, and edge cases that decide whether users trust what you ship. We find it before your users do.