CSS frameworks · Framework 08 of 08
Four Ways It Breaks
AI systems fail from four directions and only four: the model changes, the prompt drifts, the data changes, or a vendor changes a connection. Reliability watches all four and ignores everything else.
Why it matters
An AI workflow that worked in March can quietly stop working in June with nothing in your business having changed.
A model update shifts how drafts read; six months of small edits leave the prompt contradicting itself; a new field appears in the job record; a vendor moves an API.
You find out from an angry customer unless someone is watching the four places it can come from.
How we apply it
- At handoff, each of the four sources gets a check: a monthly re-run of your output tests, a prompt diff, a data schema check, and connection monitoring with alerts to us.
- Each month the Reliability report shows all four columns, including the months with nothing in them, so you can see the watching happen.
- When something drifts, we log the incident by source, fix it inside the plan's hours, and note the fix in the report.
- Quarterly we check whether the workflow still pays at your current volume, and say so either way.
The number
| Measure | Cadence | Baseline |
|---|---|---|
| Incidents caught before a customer saw them, by source | Monthly | Zero incidents at launch |
| Incidents a customer or tech reported first | Monthly | Target zero |
| Output-check pass rate | Monthly | Launch report |
| AI cost per completed job | Monthly | Launch report |
Who is accountable
If a drift from one of the four sources reaches a customer and our report that month shows that column clear, the fix and the next month of Watch are free.
We reply within one business day on Watch, and say in writing if the workflow stops paying for itself.
Where you meet it
Watch, $1,000 a month after launch.
A common question
Four Ways It Breaks, asked and answered.
Isn't this just maintenance?
Maintenance fixes what broke. This watches the four places breakage comes from, so the fix happens before a customer notices, and you can see the empty columns in the months nothing did.