The pilot that everybody quietly depends on
It started as an experiment. Someone keen, often in operations or customer service, built a Custom GPT that drafts replies to supplier queries, or checks quotes against a price list, or turns site notes into a report. It worked. Colleagues started sending them things to run through it. Now a real part of the week depends on it.
And it all lives in one person's account. The instructions are in their Custom GPT settings, the reference files are uploaded under their login, and the process is "send it to Sam and Sam will run it". When Sam is on holiday the work waits. Nobody else is quite sure what the instructions say. Management has been told the pilot was a success and keeps asking when it will go live, and nobody can say what "live" would even involve.
Why it gets stuck at this stage
The pilot proved the model can do the task. It did not prove anything about running the task as a business process, and those are different questions. A pilot in a chat window has no connection to your systems, so someone copies inputs in and outputs out. It has no record of what it produced last Tuesday. There is no way to tell whether a change to the instructions made it better or quietly worse.
The bigger blocker is ownership. Nobody commissioned it formally, so nobody owns the move to production. IT sees it as a personal experiment. The person who built it has a day job. The pilot is stuck because it is sitting in the gap between a successful experiment and a funded project.
What staying in limbo costs
| While it stays a pilot | The effect |
|---|---|
| One person runs it | Work queues behind their availability and holidays |
| Inputs copied by hand | Time goes on shuffling text between windows, and mistakes creep in |
| No log of outputs | When a customer disputes something, there is no record of what the AI wrote |
| Instructions changed ad hoc | Quality drifts and nobody can say when or why |
| Reference files out of date | Answers are based on last quarter's price list or policy |
Then there is the risk that sits on top of all of it. If that person leaves, the pilot leaves with them, along with the knowledge of why each instruction was written the way it was.
How we move it into a system you own
- We sit with the person who built it and extract everything: the instructions, the reference files, the examples they use, and the unwritten rules they apply when checking output.
- We collect real past cases, including the awkward ones, and turn them into a test set with the answers a competent person would accept.
- We rebuild the logic as a small service in your cloud account (Azure or AWS), calling OpenAI or Anthropic Claude through the API, with prompts stored in version control rather than a settings box.
- We connect it to where the work actually arrives, for example a shared mailbox in Microsoft 365, a form, or a folder in SharePoint, so nobody has to copy things in.
- We add a review step where a person approves or corrects output before it goes anywhere important, and the corrections are recorded.
- We run the test set every time the prompt, model or reference data changes, and we keep a log of every input and output.
The person who built the pilot stays involved as the subject expert. Their work is what makes the system useful; the rebuild just stops it depending on their login.
What the team works with afterwards
The task runs whether or not the original builder is in. Inputs arrive on their own, output lands where people already work, and anyone with the right role can review and approve. When someone asks what the AI said to a supplier three weeks ago, there is an answer.
Changes become safe to make. If you want to update the price list or tighten the tone, the test set shows whether anything else broke before the change goes out. That is the difference between a clever trick and something you can build the week around.
Is this your situation?
- A useful AI process runs from one person's ChatGPT, Claude or Copilot account.
- Colleagues send that person work to "run through the AI".
- Management has called the pilot a success but there is no plan for going live.
- Nobody else could say exactly what instructions the AI is following.
- You would lose the process if that person left.