The Work That Continues After Launch
Last updated:
It is not like maintaining a website
Conventional software keeps doing what it did. AI integrations degrade quietly as your inputs change, your content ages and providers update their models.
That is why the maintenance figure is higher than people expect and why skipping it produces a system nobody trusts within a year.
The monthly work
- Review the evaluation results and the correction rate trend
- Look at what reviewers corrected most, and why
- Check cost per item against last month
- Confirm the index is current and nothing stale has crept back
- Adjust thresholds if the evidence supports it
Two to four hours a month, done consistently, is the difference between a system that improves and one that decays. It is not difficult work; it just has to happen.
The occasional work
- Provider model changes — rerun evaluation, adjust if needed
- New document types or customers — extend the evaluation set
- Credential rotation, before expiry rather than after
- Annual review of provider terms and pricing
Who does it
Someone in your business owns quality — reading the numbers and deciding. Someone technical makes the changes, whether that is your team or ours.
Both roles need to be named. Unowned systems are the ones that quietly stop being useful.
What it costs
| System | Annual maintenance |
|---|---|
| Simple integration, stable inputs | 15–20% of build |
| Document processing, varied inputs | 25–30% |
| Customer-facing assistant | 25–35% |
Budget it explicitly at the start. Discovering it in year two, unbudgeted, is how good systems get switched off.
Frequently asked questions
Can our own team do this?
What if we do nothing?
Is retained support worth it?
Does maintenance decrease over time?
Have an integration nobody has looked at in six months?
It is probably worse than it was. A review is a couple of days and usually finds two easy improvements.