Custom AI Solutions

Somebody spends hours a week reading, sorting or retyping information that arrives in a predictable shape

A decision gets made the same way every time by a person who has to be available for it to happen

You have been quoted for an AI project and cannot tell whether the number is reasonable

The demo that never becomes a product

A model that works on ten hand picked examples usually fails on the eleventh. We test against a sample of your real inputs, including the messy ones, before agreeing that a system is finished.

Cost that scales faster than usage

AI spend fails quietly, because nothing breaks until the invoice arrives. Cost is attributed per request and per feature from the first day, and spending ceilings are triggered deliberately in production before anyone relies on them.

Accuracy nobody agreed on

Good enough has to be a number written down before the build starts. We agree the acceptance threshold and how it will be measured, so finished is a fact rather than an opinion.

A paid discovery step comes first, usually the audit above or a shorter scoping workshop when the problem is already well defined. It produces the written scope, the acceptance threshold as a number, a sample of your real inputs including the messy ones, and the fixed price. Nothing is built before that document exists, because the scope conversation is what decides whether the project succeeds.

The build runs in weekly demos against your real data. Week one usually ends with the core loop working end to end on a narrow slice, so the shape is visible early. Integration with the systems you already run comes next, because that is where most of the value and most of the risk sits. Cost instrumentation goes in from the first request, attributed to a customer, a feature and a model, and the spending ceilings are triggered deliberately in production before you rely on them. Go-live is followed by handover: the repository, the provider accounts and the credentials in your name, a runbook, documentation and a recorded walkthrough.

The build price is fixed. The running cost scales with usage, and it is the number most quotes leave out. From operating our own AI products, three things hold. The model is rarely the expensive part; a text message or a minute of telephony costs more than the model behind it. The small model does most of the work once the workflow is designed properly, with the frontier model reserved for the few steps that need it, which makes a well-built system several times cheaper than the per-token headline suggests. And a spending ceiling that has never fired is an assumption, not a control.

Every system we build reports cost per feature and per customer from day one, so the running cost is a number you watch rather than an invoice you discover. Our running-cost guides give worked examples at current prices.

Custom is the right answer when the work has to happen unattended, at volume, inside your own systems: reading every application that arrives, answering the phone at 9pm, chasing every invoice, reconciling three systems that were never meant to talk. It is the wrong answer when a paid ChatGPT or Claude workspace would do, or when an off-the-shelf product already matches the workflow closely enough. Between the two sit the workflow tools, n8n, Make and Zapier, which are right for a single process with modest volume and wrong once reliability, cost at scale or data handling matter.

We say which of the three applies in the discovery step and we quote only for the one that fits. Half the value of a paid discovery is being told not to build.

AI screening that reads every application against a role, scores it and explains the score, with a recruiter deciding. A voice and message agent that answers inbound calls, chases applicants or patients, and books appointments into a real calendar, with the fallback designed before the voice. Document intelligence that reads invoices, contracts or referrals into structured records and flags the exceptions for a person. A customer-facing assistant grounded in your own knowledge base with escalation to a human where money or compliance is at stake. Each of these is a system we have built and run, and each is priced as a fixed scope.

How much does a custom AI system cost in the UK?

Augustova prices custom AI builds between £8,000 and £25,000 as a fixed price, agreed in writing before any work starts, with an explicit list of what is not included. Larger platform work is quoted per phase so the budget cannot drift mid project.

How long does it take to build a custom AI system?

Two to four weeks from agreed scope to a working system in production. Most delays on projects like these are decisions waiting on somebody rather than engineering time, which is why the scope and the acceptance criteria are agreed in writing before anything is built.

Who owns the code and the accounts?

You do, from day one. The code, the repositories and the provider accounts are yours, and at go live you receive documentation, credentials, a runbook and a recorded walkthrough so another developer could pick the work up.

What if AI is not the right answer to our problem?

We say so. Plenty of problems are better solved with ordinary software, a process change, or nothing at all. The discovery step is paid and its output is yours to keep, so an honest no still leaves you with a written scope you can take anywhere.