AI and machine learning

Most of the risk is in choosing the wrong job for the model. A model can sort easy requests so well that everyone is delighted, and then quietly get the hard ones wrong, and nobody finds out until a customer does.

So we test the idea on your own work first. We keep some of your examples back where the model can’t see them, we score it only on those, and we tell you plainly when it isn’t worth building.

Most people start with an assessment: one week, one question, such as “can a model sort these 500 requests the way your best person does?”

Before the week begins, we write down what result would mean no. That way the answer can’t be bent afterward to look like a yes.

If the prototype earns more, we build it in weekly steps, and every step ends with a check that scores the model only on examples it hasn’t seen. You can stop after any week.

Not all of it has been measured yet. We have built one tool this way, and we have not yet measured it on a subtle brand, so we don’t claim more than that.

1. Assessment. One question, answered in a week, with what we expect and what would prove us wrong written down first.
2. Prototype. The smallest thing that could answer the question, working, for you to open and use.
3. Project. The rest, built in weekly steps, each ending on a date you know.
4. Ongoing support. After it launches, any week you want something changed, fixed or watched. No retainer.

• The brand character trainer: You load your writing and call passages on or off your voice. It shows how well a checker learned your ear, scored only on passages it hasn’t seen, with every miss listed. It refuses to give a number when fewer than 20 held-out labels exist. We haven’t measured it on a subtle brand yet.

A week costs $5,000: five working days, a price agreed before the week starts. An assessment is one week.

For one person, a lighter week of two to three days costs $1,000.

An assessment is one week, and we write down before it starts what result would mean no.

Model usage costs are separate, and we show them to you before anything runs. After the assessment you decide week by week whether to go on.

If the goal is everyday productivity across your staff, a team AI plan is the cheapest and fastest route: Claude Team is $20 per standard seat per month billed yearly. If your people live in Microsoft 365, Copilot sits inside the tools they use: the Business plan is $18 per user per month for existing Microsoft 365 customers on a first-year promotion that runs through December 2026. If you need to connect tools and automate well-defined steps, Zapier’s Team plan is $69 a month billed yearly for 2,000 tasks. We’re worth hiring when the job is specific to your work, and you want it tested on your own examples first, with a score that comes only from examples it hasn’t seen.

You keep the examples, the scores and the code. We show you what each model run costs before it runs, and where the data goes.