AI consultancy for manufacturing SMEs: start with a review
In one to two weeks we go through your office tasks and hand you a short list: the use cases worth a project, the ones that are not, and a price range for each. After that, things get built one process at a time.
Updated
You need this if
- The board keeps asking ‘what are we doing about AI’ and nobody has an answer with numbers in it.
- People use online assistants on their own, with no rule about what may be pasted into them.
- You are being pitched chatbots and AI agents and cannot tell which offers are serious.
- Support and sales hunt for answers in manuals and datasheets scattered across shared drives.
- You want to know where your drawings, prices and customer data would go before you try anything.
What you have at the end of the review
-
The task list
The repetitive tasks in the sales office, sales, purchasing and support, with their volume and the time they take, recorded with the people who do them.
-
Use cases, ranked
Three to five cases worth doing, ranked by expected gain and by difficulty. And the list of those we rule out, with the reason.
-
A trial on your documents
One case is tested on a sample of your real documents: you see what the model reads correctly and where it goes wrong.
-
The data note
For each case: which data leaves, to which model, where it is hosted, and what must never leave.
-
A priced roadmap
A price range and a duration for each case, in the order we would do them. The firm price of each project is fixed at the end of its short scoping.
-
A usage rule for the team
One page: which tools are allowed, what never gets pasted into an online assistant, who checks what.
Example: the review at an 80-person manufacturer
A typical run, for illustration. The company and the figures are made up.
- 1
Interviews with the sales office, two salespeople, accounts and support: fourteen repetitive tasks recorded.
- 2
Three stand out by volume: keying in orders received as PDFs, quotes for standard parts, technical questions from distributors.
- 3
Trial on 40 past orders: purchase orders produced by software are read well, scanned handwritten orders still have to be keyed in.
- 4
Two ideas are ruled out: sales forecasting (too little history) and a public chatbot (too few questions to justify it).
- 5
Debrief: three automation projects proposed in order, each with its price range and duration, plus the usage rule for the team.
How the review runs
Interviews
With the people who do the work, not only their manager. By video call; a visit is possible and priced separately.
2 to 3 days
Trial on your documents
Once the GDPR data-processing agreement is signed, we test one case on a real sample and measure what is read correctly.
2 to 4 days
Roadmap
Ranking of the cases, the data note, and a price range and duration for each project.
1 to 2 days
Debrief
We present the document to management and the teams. What you do with it is up to you, with us or without us.
one 2-hour meeting
What the price covers
The review is a fixed price, deducted from the first project we build for you. It is the same audit as our website audit, pointed at automation and language models: you buy one, not both. The other rows are starting prices; each project’s firm price is fixed at the end of its short scoping.
- AI review and roadmap
- £1,650 ex VAT
- Automation, per process
- from £3,400 ex VAT
- Assistant on your documents, with sources
- from £4,300 ex VAT
- n8n set-up and first workflow
- from £2,500 ex VAT
- Maintenance
- from £390 ex VAT a month
AI for manufacturing: what is worth a project, and what is not
Only 2% of UK manufacturers say AI is widely embedded across their operations, and more than half name skills as the main barrier (Make UK, June 2026). The gap is easy to explain: drafting a text needs nothing, automating a task means connecting it to your data. Here is how we rank six frequent requests.
| Use case | Our view | Why |
|---|---|---|
| Reading orders, price requests, invoices | Worth a project | Steady volume, a result you can check line by line, a person signs off. |
| Answering from your manuals and datasheets | Worth a project if the documents are up to date | The assistant quotes the page its answer comes from; it knows nothing that is not written down. |
| Translating and formatting product datasheets | Worth a project | High volume, quick review by someone who knows the product. |
| Public chatbot on the website | Rarely | Few questions, and the risk of a wrong answer given to a customer. A complete enquiry form and a clear catalogue do better. |
| Forecasting sales or breakdowns | Not with a language model | That is statistics on a long, clean history: a different trade with different tools. |
| An agent that orders, chases or negotiates on its own | Not yet | A mistake binds the company. By default, a person signs off. |
The cases that make the cut are then built one by one, as automation projects of 2 to 4 weeks.
Where the model runs, and what is sent to it
It is the first thing an IT manager asks, and there is a precise answer for each case. The model only receives the text needed for the task (one order, one page of a manual), not your whole customer database.
| Option | Where the data is processed | When to choose it |
|---|---|---|
| A European provider’s API | Mistral hosts data in the European Union by default (Mistral). | By default, when the model is good enough for the task. |
| A US model served from Europe | Claude models are offered on Amazon Bedrock with routing kept within Europe (AWS); OpenAI offers European data residency, subject to conditions (OpenAI). | When one particular model reads your documents better. |
| An open model installed on your own server | Nothing leaves your server. | For drawings or data you will not send to anyone. Heavier to install and to maintain. |
Anthropic and OpenAI state that, by default, they do not train their models on data sent through their business offerings (Anthropic, OpenAI). We check this for the model chosen and write it into the data note, with the place of processing and the retention period. We call models through an option that processes data in the EU when the provider offers one. Accounts are opened in your name, and a GDPR data-processing agreement is signed before we touch any real data.
An assistant that answers from your own documents
Your manuals, datasheets, procedures and past answers are indexed. When a service engineer or a salesperson asks a question, the assistant finds the relevant passages, writes the answer and shows its sources: the document and the page. The assistant is set up to say when it finds nothing instead of making something up; we test that setting with your real questions.
- It is only as good as the documents. An out-of-date manual gives an out-of-date answer, with the source there to spot it.
- It reads drawings, diagrams and scanned tables poorly.
- It can get a summary wrong. The person asking keeps the source in front of them and remains responsible for what they tell the customer.
- To open it to customers or distributors, we go through the customer portal, where you know who is asking and about which machine.
It works best on a well-kept product catalogue, with one datasheet per part number.
When you do not need an AI consultant
- For everyday drafting, summarising or translating. A subscription to an off-the-shelf assistant and a usage rule are enough. That is not a project.
- If your ERP or CRM already offers the feature. Try it first: it is already connected to your data.
- If the problem is the data. Duplicate part numbers or prices kept in three spreadsheets are not fixed by a language model, but by a clean-up or a piece of bespoke software.
- If nobody on your side can give it two hours a week. A project with no internal owner stops at the demo.
Frequently asked questions
How much does the AI review cost?
£1,650 ex VAT, fixed, for one to two weeks of work. It is deducted from the first project we build for you. It is the same audit as our website audit, pointed at automation: you buy one, not both.
What does an AI project cost after that?
From £3,400 ex VAT per automated process, in 2 to 4 weeks. The roadmap gives a price range for each case; the guide to AI automation pricing shows where those figures sit in the market.
Is our data used to train the models?
Not by default at Anthropic and OpenAI, according to their terms for business offerings. For whichever model is chosen, we check this point and write it into the data note, with the place of processing.
Can the AI get things wrong?
Yes. A language model can misread a figure or state something false. So every case keeps a human check, automatic controls (prices, quantities, part numbers) and an error count during the pilot.
Do we need a recent ERP?
No. We need to be able to read from it and write to it through an API, a file or a connector. The review checks this, whether you run Sage 200, Access Dimensions, Pegasus Opera, Odoo, SAP Business One or an in-house system.
Do you only advise, or do you build as well?
Both, in that order: the review is consultancy at a fixed price, and what follows is build work that we do ourselves. You can also keep the roadmap and hand it to someone else.
Read next
- Service
AI automation
One repetitive task handed to an automation, with a person who approves. from £3,400 - Service
n8n workflows
Your automated workflows built with n8n and hosted on your own EU server. from £2,500 - Service
Audit
A review of the website, the tools and the manual tasks, with a costed roadmap. fixed price £1,650 - Guide
AI automation pricing
You pay for AI automation three times: to build it, to run it and to keep it working.
Start with a fixed-price audit
In one to two weeks you know what to build, in what order and in what price range. The audit fee is deducted from the first project we build for you.
£1,650 ex VAT 1 to 2 weeks See a sample report