Partners
A partner badge tells you nothing about whether the thing we build will work.
So here is what each one actually changes about the work, including where it is the wrong choice.
The models we build on
Once a system runs on your real data, the model is a procurement decision.
Anything we build sits on your data, gets run by people who are not engineers, and outlives our involvement.
That narrows what matters to three things: how the model behaves when it is unsure, what the vendor does with the data, and whether you can still run it in two years. Benchmark scores speak to none of them. A model two points better on a reasoning test and confidently wrong in production is worse than one that stops and asks.
- Your data
- Business and enterprise API traffic is not used to train models by default. This is usually the first question your legal team asks, and often the one that decides the project.
- Behaviour when unsure
- Trained to flag ambiguity rather than invent an answer. In a finance or compliance workflow that is the difference between a system people trust and one they quietly stop using.
- Where we use something else
- High-volume classification, deterministic extraction, and anything with a hard speed limit are often better served by a smaller model or none at all. We do not put an LLM where a rule would do.
- What it is not
- Not a reseller deal. We take no margin on your model spend, and you contract with the vendor directly so the cost stays visible and the dependency stays yours.
Platform partners
The systems the work usually has to run through.
Most workflows worth automating already live in one of these. Partnership here means we know the platform well enough to build inside it rather than around it.
Spend, cards and bill pay. Finance workflows are a common first deployment because the before-and-after is unusually easy to measure: approvals, coding and reconciliation all have a countable cost today.
CRM and go-to-market. Usually where the messiest data lives, and where automating a process nobody trusts does the most damage, so we tend to fix the record before adding anything to it.
Outbound sequencing and deliverability. The obvious place to point a model, and the easiest to get wrong: personalisation that reads as generated does more damage than none at all. We use it for research and reply handling more often than for writing the message.