AI development
that does a job,
not a demo
Most AI development stalls because nobody agreed what the AI was for. We start from the job — the question your customers keep asking, the content nobody has time to write, the process a person re-keys every morning — and build the smallest thing that does it properly.
* Assistants, agents and automation, grounded in your own data and checked before anything reaches a customer.
Where AI actually pays
Two places it earns its keep
AI is worth adding in two places: in front of your customers, where it answers the questions your team answers over and over, and behind them, where it does the repetitive work that keeps a catalogue, a feed or an order queue moving.
Everything else is a demo. We would rather tell you a job is not worth automating than build something impressive that nobody uses after the first month.
The work is ordinary software development with a model inside it. It needs the same things any other system needs — data it can reach, a clear definition of correct, a way to tell when it is wrong, and someone who can fix it. That is why this sits next to custom software development and API development and integration rather than apart from them.
If the aim is to be found and quoted by AI search rather than to run AI yourself, that is a different piece of work — SEO and GEO services covers it.
Two halves of the same work
01
In front of customers
Assistants that help someone choose, understand or troubleshoot — answering from your catalogue, your policies and your documentation, in your tone.
02
Behind the scenes
Product content at catalogue scale, classification and matching across messy feeds, triage and routing, and the steps in a process that a person currently repeats by hand.
What we build
What
we build
Six kinds of work, all of them things we have shipped and still run. Each one starts as the smallest useful version, measured against what the job looked like before.
Related: Custom software development · API development and integration
Shopping and support assistants
An assistant in the page that helps someone choose — matching what they describe to what you actually stock, explaining the difference between two products, and handing over to a human when it should. Grounded in your catalogue and policies, so it cannot promise a product or a delivery date you do not have.
Product content at catalogue scale
Descriptions, benefits, meta titles, usage and warning copy across thousands of products, written to one house style and checked against the manufacturer's own information rather than invented. Structured so your team reviews and approves rather than retypes.
Classification, matching and clean data
Mapping products onto a taxonomy, extracting attributes from supplier feeds, spotting duplicates across catalogues and normalising names, brands and URL slugs. The unglamorous work that decides whether search, filters and feeds behave.
Agents wired into your systems
Where an answer is not enough and something has to happen: an order progressed, a refund raised, a ticket routed, a record updated. We give the model the narrowest set of tools that does the job, and every action it can take is one you could audit afterwards.
Search that understands the question
Retrieval over your own content and catalogue, so a search in a customer's words finds the right product or the right page. Often more valuable, and far cheaper to run, than a conversation.
Guardrails, evaluation and cost control
Validated output, refusals where the model does not know, rate limiting and abuse handling, logging you can inspect, and a test set that tells you whether a prompt change made things better or worse. This is the part that decides whether an AI feature survives contact with real traffic.
Running
in production
Four pieces of AI work we built and still support. We do not name clients on this page, but we will talk you through any of them.
Our own
CogniBot, on this site, is the same pattern: it answers from this site’s own content, and it tells you when the answer is not there instead of filling the gap.
An online retailer whose assistant replaces the person on the shop floor
Customers used to get their advice across a counter. We built the assistant that does it online — asking what they use now, explaining the trade-off between two options, and pointing at the product that fits, with the guidance rules the client already published behind it rather than the model's own opinion.
A health marketplace with thousands of products to describe
A supplement catalogue where every product needs a brand, a clean URL, a meta title and description, benefits, directions and warnings — all consistent, all verifiable against the manufacturer. We built the pipeline that drafts it against the brand's own sources and hands it to a human to approve, plus the shop-side assistant customers ask.
A wholesale platform where the process runs itself
Suppliers and resellers trading through live catalogues, with orders, deadlines, cancellations, refunds and payment release handled by the system rather than by an inbox. Product data arrives messy from many suppliers and has to land on one taxonomy before anyone can search it.
This site
A grounded assistant built on our own content, with validated output, rate limiting, prompt caching and a refusal path — small enough to read in an afternoon, which is the point. Ask it something we have not published and watch it say so.
From idea to
something live
AI work goes wrong in a predictable way: a promising demo, then months of nobody being able to say whether it is good enough. We put the measure in place before the build, so that question always has an answer.
01
Agree the job and the measure
One job, and how we will know it is being done well enough to ship.
02
Get your data in front of it
Catalogue, documents and policies, in a form the model can actually use.
03
Build the smallest useful version
Narrow scope, checked output, a refusal path, and a cost you have seen.
04
Measure it on real traffic
Watch what people really ask, fix what it gets wrong, widen the scope only when it earns it.
Clear
answers
What people ask before any of this starts — including what we cannot promise.
Almost never. Most of the value comes from using a good general model well: your own content in front of it, a narrow job to do, and a check on what comes back. Training or fine-tuning a model is a decision we would justify to you before spending anything on it.
The assistant only answers from information we put in front of it, and we validate the shape of every answer before it reaches a customer. When the answer is not in what it was given, it says so rather than guessing. You can see this working on our own site — CogniBot answers from this site's content and nothing else.
You do. Model accounts and API keys are yours, the prompts and evaluations live in your repository, and the data stays in systems you control. Nothing here is designed to be difficult to take somewhere else.
Usage is metered per question, so the running cost is a function of traffic and how much context each answer needs. We design for that from the start — caching what repeats, keeping context tight, and rate limiting abuse — and we tell you the expected monthly figure before you commit to launching it.
Next step
Tell us the job,
not the technology
Describe the question your team keeps answering, or the work somebody repeats every morning. We will tell you whether AI is the right tool for it, what the smallest useful version looks like, and what it would cost to build and to run.