# How We Use AI
> How RJJ Software uses AI: what it touches, what it does not, why any client can rule it out entirely, and the one podcast tool whose limits we cannot verify.


This page was last updated on September 1st, 2026


This page sets out how RJJ Software Ltd uses artificial intelligence (AI): where it is used in our work, where it is not, and what we will not do with it. No law requires us to publish any of this. We have written it because anyone deciding whether to work with us, or reading something we have published, is entitled to know.

## What We Use It For


Three things, and the list is short because that is the extent of it. We use AI assistance when drafting and editing our own writing, including the pages on this site. We use code assistance when we write software. And we use it to correct the transcripts for our podcast, which is set out in full further down.

Nothing else. It plays no part in what we recommend to a client or what we charge, and nothing goes on this website that a person has not read, edited and signed off.

## On Client Work, You Decide


Code assistance is the one use that touches somebody else's work, so it is handled differently.

We raise it during contract negotiation, before anything starts. Our preference is to use these tools and we say so plainly. But nothing runs on an engagement until the client has explicitly agreed, and not before we have walked them through the workflow: which models, which harnesses, and how the output gets reviewed. Clients get the specifics, because a conversation keeps pace with a changing toolchain in a way a published page cannot.

If a client would rather we did not, then we do not, and that covers the whole engagement rather than just their codebase. No code assistance in their repository, and nothing AI-drafted handed over as our work.

One limit is worth admitting. Spellcheck, autocomplete and predictive text are built into ordinary office software, and there is no meaningful way to switch all of it off; our documents start life as Markdown and pass through an office application or a converter to become a PDF. The commitment is the absence of code assistance and generative drafting, not a claim that no model has been near a file.

Whichever way that conversation goes, anything produced with AI assistance goes through my own editorial pass before a client sees it. That one is mine, and I do not hand it off.

## What Does Not Get Delegated


Editorial responsibility sits with people. The ideas, the opinions and the technical judgements are ours, and when something here is wrong, it is because we got it wrong.

Fact-checking is a human job. So is sign-off: nothing reaches a client or this website without someone having read it end to end and being willing to put their name to it. A model can draft a paragraph. It cannot be accountable for one, and we are not going to write as though it can.

## What We Do Not Do


We do not invent testimonials, reviews or client results.

We do not publish unreviewed output and pass it off as considered work.

And we do not clone or impersonate voices. That was not a hypothetical: when illness stopped our podcast intros being recorded, cloning was technically available to us and we turned it down, because having a synthetic Jamie welcome listeners to a show the real one could not speak on would have been a small dishonesty. The podcast section below describes what we did instead.

## The Modern .NET Show


[Our podcast](https://dotnetcore.show/) is where AI is most involved in what we do, and three separate things are going on.

We make transcripts in three passes. The finished episode goes through MacWhisper running NVIDIA's Parakeet model, entirely on the machine, so the audio is never uploaded to a transcription service. What comes out needs a lot of correction, so the text goes to Claude with a prompt carrying guest name spellings, the accents that could have thrown the recognition, and the technologies discussed. The transcript text does leave the machine at that step, and we would rather say so than let "runs locally" cover the whole process. Then someone on the team listens to the entire episode while reading the corrected transcript, editing as they go. That last pass earns its place. The failure mode of a language model correcting a transcript is fluent invention: a garbled passage gets smoothed into something plausible that nobody said, and only reading against the audio catches it.

An external editor, [Matthew Bliss](https://www.mbpod.com/), handles our audio post-production. We hand over the raw conversation audio; he hands back the finished file. When we asked him what AI touches that audio, he named one tool: [Accentize dxRevive](https://www.accentize.com/product/dxrevive/), which processes locally on his machine with nothing sent to a cloud service, and which he uses selectively rather than on everything.

It is not a noise filter. It rebuilds speech from patterns it has learned, and Accentize describe their own product as marking "a shift from traditional enhancement toward generative audio processing". They state two consequences plainly: the output is "no longer a pure version of the original signal", and "subtle changes in voice character may occur". Matthew's experience matches that, and he is more specific about when: the voice tends to shift on recordings that were difficult to start with, which is part of why he uses it sparingly.

He is clear that he cannot contrive words or sentences that were not in the original recording, and the plugin's controls bear him out: per its manual they set how much processing to apply, not what is said. There is nowhere to type words in.

What we cannot tell you is whether a model rebuilding damaged speech might alter a word by accident. Accentize do not document it, we have not found anyone who has tested it independently, and the audio profession is still arguing about where restoration ends. So we are not going to claim the words are untouched, because we cannot show it. We can tell you which tool is there, what its makers say it does, what Matthew sees in practice, and where the answer runs out. How he uses it is his own craft, and his to describe.

Six episodes, published between December 2024 and March 2025, open with a synthetic voice reading the intro. A stock voice from a text-to-speech licence we already held stood in when illness made recording impossible. It impersonates nobody, it introduces itself in the audio as not being the host, and all six say so in their episode metadata as well. [The full account is in a post about AI in podcasting](/blog/podcaster-questions/can-ai-powered-podcasting-really-boost-productivity-and-innovation/).

Until 2025 we also ran our own fine-tuned models for audio mastering, as part of [a podcast production service we have since closed](/blog/the-podcasting-years/).