It works, but nobody knows how
A freelancer built an integration that uses ChatGPT to draft replies in your helpdesk, or to summarise calls into the CRM, or to answer questions on your website. It went in, it mostly worked, and the business started relying on it. Then the freelancer took on other work, stopped replying to emails, or simply moved on.
Now something has changed. The replies have got worse, or it has stopped altogether, or OpenAI has retired the model it used. You look for the code and find it is on a server in the freelancer's account. The API key is on their card, or yours, but nobody knows where it is used. There is no documentation. You are dependent on a system nobody in the business can see inside.
Why this happens so often with AI integrations
AI integrations are often quick to get working, which makes them attractive as small freelance jobs. The first version is a few hundred lines of code and a prompt. What gets skipped is everything that makes software maintainable: keys held in the client's own accounts, code in the client's repository, logs, tests, and a written note of how it works.
AI integrations also change underneath you in ways normal software does not. Model versions are retired, pricing changes, and outputs shift when the provider updates a model. An unmaintained integration does not just stay the same; it slowly degrades.
What the dependency is risking
| Gap | The risk |
|---|---|
| Code and hosting in someone else's account | You cannot fix, move or even switch it off |
| API key in an unknown place | Costs you cannot control, and exposure if the key leaks |
| Retired or changed model | Output quality drops or the integration stops working |
| No logs | You cannot see what it has told customers or staff |
| No documentation | Any change starts with reverse-engineering |
If it touches customer data, there is also the question of where that data has been going and under whose terms, which you currently cannot answer.
How we take over an orphaned AI integration
- We work with you to regain access: domains, hosting, repositories and provider accounts. Where the freelancer can still be reached, we request a proper hand-over; where not, we work from what is in your control.
- We move code into a repository you own, hosting into your own AWS, Azure or other account, and API keys into your own OpenAI or Anthropic organisation, then retire the old keys.
- We read the code and write down what it actually does: what data it sends where, which model and prompt it uses, what it writes back into your systems.
- We add logging of inputs and outputs, error alerts and a spending limit, so problems show up before customers notice them.
- We build a test set from real past cases, so the current behaviour is recorded and any change can be checked against it.
- We then give you an honest recommendation: keep it with light fixes, repair specific parts, or rebuild it properly, with the reasons for each.
Sometimes the original work is sound and just needs a home and some care. Sometimes it is fragile enough that rebuilding costs less than patching. We tell you which, and why.
What you have afterwards
An integration that sits in accounts you control, with documentation a new developer could follow, logs that show what it is doing, and tests that show whether a change broke it. If the provider retires a model, switching to the new one is a checked change rather than a crisis.
You also know what to ask for next time. A reliable ChatGPT integration company will hand over code, keys and documentation into your accounts as a matter of course, and explain how the system is monitored after launch. If a supplier is vague about any of that, it is worth asking why.
Is this your situation?
- An AI integration was built by a freelancer or small agency you can no longer reach.
- You do not have access to the code, hosting or API keys.
- Output quality has dropped, or it has stopped working.
- Nobody in the business can explain what data it sends to the AI provider.
- You are looking for someone reliable to take it over or replace it.