Pre-Deployment Failure Mode Review (QA)

Pre-Deployment Failure Mode Review (QA)

1. 01

   Enumerate every system the bot will touch and classify each as UI-only, API-backed, or hybrid

   UI-only targets carry the highest breakage risk because selector drift and layout changes silently corrupt runs. Flag any workflow that depends on more than two UI-only systems for a second review before build.

2. 02

   Write down the expected output for three representative inputs before the first test run

   Pick one clean case, one edge case, and one malformed input. Recording the expected result in advance stops the team from rationalizing whatever the bot produces as correct.

3. 03

   Test the bot against a staging copy of client data, never production records

   A bot that writes to the wrong row in a live CRM can take days to unwind. Clone a representative dataset and confirm write operations are reversible before touching client systems.

4. 04

   Insert a human approval gate on any step that moves money, changes contract terms, or sends external communications

   Autonomous systems that act on predictions without review are the defining risk of 2026 deployments. Route high-stakes decisions to a named reviewer and log the approval timestamp.

5. 05

   Run the workflow five consecutive times and compare outputs for drift

   Single-pass tests hide timing and session-state bugs. Five runs surface intermittent failures that would otherwise appear in week two of a retainer, when the client is watching.

6. 06

   Document the failure signature for each known break mode and the first diagnostic step

   For a scraper, that might be a CAPTCHA wall or a changed selector. For a document processor, a new invoice template. Whoever is on call should not have to rediscover the diagnosis.

7. 07

   Confirm the client's IT or security contact has signed off on access scope before go-live

   Scope creep after launch is the most common cause of a stalled automation engagement. Get written confirmation of which accounts, folders, and endpoints the bot may reach.