How we use AI

We’re not loyal to a vendor. We pick the model that fits the job and the budget, and we’ll tell you which one is running where in your project before it ships.

Most production work runs on the major commercial APIs. Some workloads run on smaller or local models specifically because they are cheaper and the task doesn’t need more. A meaningful part of our automation work uses no model at all — deterministic code is faster, cheaper, and more predictable, and we reach for it first.

If a model choice materially changes your running cost or your data posture, that’s a conversation we’ve with you, not a decision we quietly make.

We don’t send client data to consumer chat products. Production work runs on API and business tiers where training on submitted content is off, and we confirm the current terms for each vendor at the start of a project rather than assuming last year’s answer still holds.

Where it’s practical, automations run against your own accounts and your own API keys, so the data path belongs to you and you can revoke our access at any time without waiting for us.

We document which tools touch your data before anything is built, and that list goes in the project documentation you keep. If your procurement or security team needs to review it, that’s expected, not an inconvenience.

We don’t feed client brand assets, source code, or customer records into tools whose terms permit training on inputs.

Anything expensive to get wrong. Automations ship with review gates in front of irreversible or externally-visible actions — money moving, records being deleted, messages going to your customers.

Uncertain cases route to a person instead of being guessed at. A system that says "I am not sure, look at this one" is worth far more than one that’s confidently wrong 4% of the time.

Every automation we build logs what it did, so you can trace any individual decision after the fact. And every one has a kill switch that doesn’t require us to operate it.

On content work, a person reviews every asset before delivery. Nothing goes to a client straight out of a generator.

When we produce imagery or video with generative tools, we tell you so, and we flag what needs disclosing on each platform you plan to run it on. Meta, Google, and TikTok each maintain their own AI-content rules and they change; regulated industries add advertising substantiation requirements on top.

We won’t present generated imagery as documentary photography of your actual premises, staff, or results. For lifestyle and context, generated imagery is fine and we use it. For the hero product itself on a regulated or high-fidelity item, we shoot the real thing.

We don’t generate the likeness of a real, identifiable person without a signed release from that person. Everywhere else we use synthetic likenesses.

You do. Source code, prompts, configuration, reference sets, and documentation are yours on final payment, to keep, modify, or hand to another developer.

We don’t rent you access to something you paid us to build, and we don’t hold your deployment hostage. If you leave, you leave with the working system and the instructions.

Delivered content assets come with full commercial usage rights. We’ll tell you plainly where the law around generated content is still unsettled rather than pretending it’s not.

Yes, as a tool, in the same way we use a compiler and a code editor. It makes us faster. It doesn’t decide what your site says, who it’s for, or why anyone should buy from you.

What we don’t do is hand you generated output and call it a custom build. Every project is reviewed, restructured, and tested by a person who is accountable for it — and the time AI saves goes into the parts that actually move your numbers, not into the invoice.

If you have ever looked at a site and thought "that was made by AI", you have identified the thing we’re working to avoid.