Custom AI development for your own use case
AI use case not listed? We'll look at it with you.
Custom AI development makes sense when neither an off-the-shelf tool nor a direct call to a general-purpose model solves your case, because the answer depends on your data, your process or your systems. Here is how to tell whether that is your situation, which technical routes exist, what data you need and how you know it works, before and after go-live.
If any of these sound familiar, we can help.
- 01You have your own data that could be put to use.
- 02A very specific process in your sector.
- 03You need to know if it's viable before investing.
What's included
We agree the exact scope with you in the proposal.- Manufacturing
- Agriculture
- Logistics
- Any sector with its own data
- 01
Feasibility
With your real data.
- 02
Prototype
Measurable results before going any further.
- 03
Integration
Connected to your systems.
- 04
Tracking
We check it keeps getting it right.
A model that estimates manufacturing lead times from order history.
It would first be tested on a year of historical data and, if it is accurate enough, handed to the sales team for quoting.
When you need AI of your own, and when you don't
Most businesses do not need a model of their own. Work up this ladder and stop at the first rung that solves the problem:
- An AI feature in a tool you already pay for: Microsoft 365, Google Workspace, your CRM or your accounts package.
- A SaaS product built for the task, if your process looks like everyone else's.
- A general-purpose model through an API, with clear instructions and your documents or records supplied as context. This covers a large share of real cases.
- Bespoke AI solutions: when the result depends on your history, your rules or several of your systems, and has to fit into how work is actually done.
- Training a model: only when the steps above have been tried and measured and fall short.
How this page differs from our other AI services
If you do not yet know where AI would pay off in your business, start with AI consulting for small business: it ends with a scored list of use cases and a plan. If what you need is an assistant that works across your CRM or ERP on behalf of your team, that is AI agent development.
This page is for everything else with a clear, measurable output: estimating manufacturing lead times from order history, classifying incoming requests by type and urgency, scoring quotes by likelihood of winning, checking photos from site against a standard, or forecasting demand by product and branch.
| Route | What it is | Fits when | Watch out for |
|---|---|---|---|
| Model API with context | A general model given instructions plus your data at the moment of each request | Reading, writing, classifying and extracting from text or images | Cost per request and dependence on one provider's versions |
| Fine-tuning | Further training of an existing model on your examples | Consistent format or style that prompting cannot achieve | Provider support can change; OpenAI has been closing its fine-tuning platform during 2026 |
| Classic machine learning | A smaller model trained on your structured history | Predictions and scores from tables: lead times, demand, risk | Needs enough clean history with known outcomes |
| Self-hosted open-weights model | A model you run on servers you choose | Strict rules on where data is processed | You carry hosting, updates and capacity |
What data you need
The question is less 'how much data' than 'which data'. Before any build we check:
- Examples with a known right answer: past orders with their real lead time, requests already classified by your team, quotes you know were won or lost.
- Representative cases: the difficult and unusual ones, not only the easy ones.
- Access: which systems hold the data, whether they have an API or an export, and who can authorise it.
- Personal data: whether any is involved, whether it is needed, and where it would be processed.
How you evaluate AI before and after go-live
We build the evaluation set first: a few hundred real cases with the correct answer, agreed with the people who know the work. OpenAI's own guidance on model optimisation suggests writing evaluations even before the prompts, and testing against inputs like those you expect in production. Every version of the system is then scored on the same set, against a baseline: how well your team or your current method does today.
Results are expressed in business terms: not an abstract accuracy score, but how many estimates land within an agreed margin of the real date, and which cases are flagged for a person to review. After go-live we keep measuring, because data drifts and models change. OpenAI, for example, has announced that two GPT-4o versions will be removed from its API on 14 January 2027. With an evaluation set, testing a replacement model takes days, not a rebuild.
How we would approach it
A manufacturer could start with a feasibility study on a year of order history: can a model estimate lead times better than the current rule of thumb? Only if the numbers say yes would it move to a prototype in the sales team's quoting screen, and then to monitored live use. Each step has a go or no-go decision based on the evaluation set, you see a working demo every two weeks, and the project has one fixed price; more on how we work.
Frequently asked questions
See all FAQs →What kind of projects fit here?
Any use of AI that doesn't fit the other categories. We'll talk it through on a call.
How do we know whether AI is the answer?
We start with a small, contained trial to see whether it works before investing more.
What about the privacy of our data?
We agree with you where it's processed and who has access, and put it in writing.
Do we have to train a model on our data?
Usually not. Most cases are solved by a general model with good instructions and your data supplied at the moment of the request, or by a small classic model for numerical predictions. Training or fine-tuning comes in only when measurement shows the simpler routes fall short.
How much data do we need to start?
Enough real examples with a known outcome to build an evaluation set, often a few hundred, and for predictions, enough history to cover your seasons and product mix. The feasibility study tells you whether what you have is enough before you spend on a build.
What happens if the model provider changes or retires a version?
Providers retire models on a published schedule and notify developers in advance. We keep the model behind a single layer of the system and re-run the evaluation set on the replacement, so switching is a tested change rather than an emergency.
What are the monthly running costs of custom AI?
Model usage, which depends on volume and on the model chosen; hosting; monitoring; and maintenance, including re-evaluation when models change. Market references and how to estimate them are in our guide to the cost of implementing AI in a business.
You might also be interested in
- AI and automationAI agents that query your systemsAsk in plain language and get the answer from your own data.
- AI and automationInternal AI assistants for your documentsYour team finds the answer in your manuals in seconds.
- Web platformsLegacy software modernisationWe modernise the software you already use without bringing the business to a halt.
Tell us about your project.
We'll reply within one working day. We start by understanding the problem, not by selling you a solution.