How we work, and what it costs
Most AI proposals are deliberately vague about price because the answer is “it depends”. It does depend — but you can still be told what it depends on, which is what this page is for.
1. Discovery
One or two workshops with the people who actually do the work, not only the people who commission it. We map the workflow as it really runs, including the exceptions everybody has stopped mentioning. You get a written scope: what would be automated, what would not, what it needs from your systems, how success is measured, and an estimate.
Discovery is paid, and it is a fixed fee in the low four figures. The document is yours whether or not you build with us — including if the conclusion is that you should not build at all, which happens and is a perfectly good outcome.
2. First build
One workflow, end to end, two to six weeks, quoted as a fixed price from the discovery scope. Fixed price is possible precisely because the scope is narrow; anyone quoting a fixed price for “an AI transformation” is guessing, and you will pay for the guess later.
It ships in shadow mode: the agent does the work, a person approves everything that leaves the building, and we measure the gap between the two. Autonomy is widened afterwards, against numbers.
3. Then one of two things
We run it
A monthly retainer covering hosting, monitoring, evaluation, model updates and a defined amount of improvement work. Most clients choose this for the first year because the first year is when things change.
You take it
Full handover: source code, prompts, evaluation sets, infrastructure definitions, documentation, and a walkthrough with your team. No licence, no per-seat fee, no dependency on us.
What drives the price
- Number of systems to integrate. The single biggest factor. Two clean APIs is a different project from an on-premise system with no API.
- Where the data may live. Fully self-hosted costs more to set up and less to run. Frontier APIs are the reverse.
- Volume. Not for licensing reasons — high volume changes which model is economic, and therefore the architecture.
- Regulatory weight. Health and legal work needs review workflows and retention rules that ordinary marketing automation does not.
- Languages. Adding languages is cheap. Adding languages that need human review is not.
Running costs, honestly
Two components: infrastructure and model calls. Where a workload runs on models we host, the marginal cost per request is effectively zero and you pay for a machine. Where it runs on a frontier API, cost scales with use and we will show you the per-conversation figure before launch. Most systems are a blend, and we will tell you which parts are which and why.
No seat licences and no black boxes. You are buying a system, not access to one. If we part ways, the thing we built keeps working and someone else can maintain it. We think that constraint makes us build better software.
How we say no
Early and in plain words. If the volume is too low, if the rules are already deterministic, if nobody internally will own it, or if what you need is a form and a notification — we will say so on the first call. A project that gets switched off after three months costs us more in reputation than it earns.
Start with discovery.
A fixed fee, a written scope, and a document you keep — including if the recommendation is that you do not need us.