I'm sorry if this is a little long winded, but I want to give an earnest answer to it to nuance it a bit but let's just acknowledge from the start that it is based on GitHub Copilot, and in the end it's just a GPT and certainly gets things wrong.

However, when it comes to the workflows/actions, we haven't encountered much problems with hallucinations but we also have a rigid set of instructions to steer it (a common problem otherwise is things like outdated versions of actions, like using actions/checkout@v3).

Our main challenge there has been to have it prioritize our internal in-house actions/understand when they're a substitute since we require explicit whitelisting for third-party actions.

But it's also a bit of a black box in that you have little control over what the user asks (which might have a company specific intent/context not known to the LLM/not covered in our instructions) and what the recommended solution was until it doesn't work or breaks and they reach out to us.

One of the things it does well is directing people to the right actions and config for on-prem vs cloud solutions, and it has read access to non-protected internal repos and has generally done a good job answering questions/using it for reference if you provide one.

I think it also happens to be a good match because it's GitHub's platform, so actions and workflows is very native to its training data. But then I think about when I ask Copilot Chat about REST endpoints for something like "list all pending installations in an org?" and it just flat out makes up something like "GET /orgs/{org}/pending-installations" with zero connection to reality, so jury's still out on that one..

It does help however that the organization has strong guidelines and guard rails for most common needs for internal products.

We have no illusions about its capabilities, in many ways it's as much an experiment for us in seeing what can be done with LLMs (and our internal data) as it is a solution to a problem.

The conception was when we analyzed our Service Now history to get an idea for what kind of queries we were receiving and seeing how much of it was answerable with boiler plate replies/internal links. It has made a marked difference in offloading those incoming questions.

We're carefully optimistic about this specific use case, while our attempts to build something like a coherent modules/template scaffolder has failed to produce good/consistent results once you start adding a couple of requirements.

We also tried using it for automating updates non-intrusively, kind of like dependabot but for our IaC, but it made a mess of things and having it change only a specific line and return a file intact was more than a challenge. (This was a while ago, I believe it might have improved in that specific area today.)

I think it's a bit sad that the AI hype has resulted in a kind of contrarian reaction from many developers who would under different circumstances embrace new tech.

As a closing note I'm convinced that building these things keeps everyone in the company grounded and the hype/expectations at bay. It's difficult to tout AI as a magic bullet when you have a few LLM based solutions up and running and everyone can experience them (and their shortcomings) for themselves.