Service
Automation and AI features
Language models and automation put to work on specific jobs inside your product or your operations, judged on whether they save time.
- Scoped per project
- UK based
- Founder-led delivery
We build AI into products where it does a specific job, and we automate the processes around a business where a person is currently acting as glue between two systems.
Both of our own products use this in ways that are easy to explain. REQO transcribes your recording so you can edit the video by deleting text. FAM turns a spoken sentence into a calendar entry or a list item. In neither case is the AI the product. It removes a specific piece of friction that used to require expertise.
That is the bar we hold client work to as well. If a feature cannot be described as the task it removes, it usually should not ship.
Included
What you get
Every engagement includes all of this. None of it turns up later as an extra.
Feature scoping with an honest verdict
We work out whether the thing you want is a good use of a language model or a problem that ordinary code solves more reliably and more cheaply. Frequently it is the second one.
Natural-language input
Letting people type or say what they want instead of filling a form. Done properly this is the highest-value AI feature in most products.
Document and content processing
Extraction, classification, summarising, and structuring from PDFs, emails, transcripts, and forms, with the output checked rather than trusted.
Transcription, captions, and search
Speech to text, subtitles, and making spoken or written archives actually searchable.
Workflow automation and integrations
Connecting the systems you already run so data stops being copied between them by a person at a keyboard.
Guardrails, cost control, and evaluation
Rate limits, spend caps, fallbacks for when a model is unavailable, and a way to tell whether the output is any good. This is the part that separates a demo from a feature.
In detail
How we approach it
The uncomfortable first question
A good proportion of AI briefs describe a problem that a database query, a scheduled job, or a better form would solve outright: faster, cheaper, and deterministically. We ask that question first and we do not mind giving away the answer, because a feature that produces unreliable output damages a product more than a missing feature does.
Language models are excellent at understanding messy human input, summarising, classifying, and generating a first draft. They are a poor choice for anything requiring exact arithmetic, guaranteed consistency, or a defensible audit trail.
Designing for being wrong
Any AI feature will sometimes produce something wrong. The design question is not how to prevent that, it is what happens when it occurs. A feature that presents its output as a confident fact fails badly. One that presents it as a draft the user confirms fails safely and is often more useful.
FAM's voice commands work this way: what you said becomes a proposed entry you glance at and accept. That single design decision is why the feature is trusted rather than avoided.
Cost, latency, and the model landscape
Model pricing and capability move fast enough that hard-wiring a product to one provider is a liability. We build behind an abstraction, set spend caps, cache aggressively where the same question is asked repeatedly, and use the smallest model that does the job rather than the most impressive one.
Latency deserves the same attention. A feature that takes eleven seconds is not used, whatever its quality. Often that means streaming the output, doing the work in the background, or accepting a slightly worse model that answers immediately.
Automation without the AI
A large share of the time we save clients has nothing to do with models. It is an integration between two systems that were never connected, a scheduled job that replaces a manual export, or a form that writes to the place the data was always meant to live.
This work is unglamorous and it has the most reliable return of anything on this page.
Process
How it runs
Four steps from first call to launch. No surprises, no disappearing acts.
Call
One focused call to understand what you need, who it is for, and what success looks like. You leave with a clear, written scope and a proposal.
Prototype
You are clicking through a working prototype with most of the core features live, early in the project. Real screens in the browser, not static mockups.
Refine
We shape the prototype into the finished product, folding in every amendment you ask for. Nothing ships until you have confirmed the final build.
Deploy
We take it live on the hosting provider and domain of your choosing, fully wired up, and stay close after launch to keep it running smoothly.
Proof
Projects built this way
Case studies with the reasoning, not just a screenshot.
FAQ
Automation & AI questions
Which AI models do you use?
Whichever fits the job, the budget, and the latency requirement. We build behind an abstraction so the model can be swapped without rewriting the feature, because model capability and cost change every few months and being locked to one provider is an avoidable risk.
Will our data be used to train a model?
Not on the configurations we use. Business API tiers from the major providers do not train on submitted data by default, and we confirm that per provider before anything is wired up. Where the data is genuinely sensitive we will discuss what stays on your own infrastructure.
How do you stop it producing nonsense?
By scoping the task narrowly, giving the model the relevant data rather than relying on what it memorised, validating the output before it is shown, and designing the interface so output is confirmed rather than assumed. We also build an evaluation set so quality can be measured instead of felt.
What does it cost to run?
Ongoing model costs depend on volume and the size of the model. For most features it lands somewhere between negligible and modest, and we set spend caps and cache repeated work so it cannot surprise you. We will give you an estimated running cost before we build, not after.
Can you automate something that is not in our product?
Yes. Plenty of this work is internal: connecting systems, replacing manual exports, routing enquiries, generating documents. It never appears on your website and it is often where the largest amount of time is recovered.
Should we be doing anything with AI at all?
Possibly not, and we are happy to say so. The honest test is whether there is a specific repeated task you can name that is currently done by a person and does not need judgement. If you can name it, there is probably something worth building. If you cannot, wait.
Related
Other services
Next step
Start here
Tell us what you need built. You will get a written scope and a proposal before any work starts.
Usually within one working day