The verdict in three sentences
A RAG AI assistant (retrieval-augmented generation) answers engineers' questions while citing the exact manual page, which makes it usable in industrial field service where a generic chatbot would make things up. The budget splits between initial development (pilot as a simple application, full rollout as a medium-complexity application) and inference at GBP 85 to 430 or CAD 150 to 750 a month, 2026 ballpark. The number one success factor is the quality of indexing of PDFs, diagrams and tables, far more than the choice of model.
Why field service at an industrial firm loses time
Take a packaging machinery manufacturer near Manchester (or in the Greater Toronto Area), 180 staff, 22 travelling service engineers, 2,400 machines installed at food processing clients. Documentation adds up to 14,000 pages: user manuals, wiring diagrams, bills of materials, service bulletins, job sheets. Some of it is in German or French, inherited from component suppliers.
An engineer facing a sealing fault on a bagging machine spends 25 to 40 minutes on average finding the right procedure, often by phoning a senior colleague. RAG changes the method: a plain-language question triggers a semantic search across indexed documents, the model drafts an answer from the passages found and cites them.
| Step | What happens | Watch point |
|---|---|---|
| Collection | PDFs, Word files, notes, job sheets | Exclude obsolete versions |
| Extraction | Text, tables, diagram captions | Old scans need OCR |
| Chunking | 300 to 800 word passages with metadata | Keep machine model and revision |
| Vector indexing | Database hosted in UK or Canada | Role-based access |
| Generation | Answer with clickable citations | Refuse when no reliable source |
| Field feedback | Engineer rates each answer | Monthly improvement loop |
Real project cost
Kolonell development scales with scope. A pilot on one machine range, 2,000 to 3,000 pages and a mobile web interface is a simple application. The full version over 14,000 pages, with permissions, multilingual sources, CMMS connection and usage dashboard, is a medium-complexity application.
| Item | Amount | Note |
|---|---|---|
| Pilot (simple application) | 2,500,000 to 4,000,000 FCFA (3,810 to 6,100 EUR) | 1 range, 6 to 8 weeks |
| Full version (medium application) | 4,000,000 to 7,000,000 FCFA (6,100 to 10,670 EUR) | 14,000 pages, CMMS, permissions |
| Inference and vector database | GBP 85 to 430 / CAD 150 to 750 per month (2026 ballpark) | Depends on query volume |
| Dedicated regional hosting | quoted | Common requirement for manufacturers |
| Kolonell maintenance | 35,000 / 75,000 / 120,000 FCFA per month (53 / 114 / 183 EUR) | Reindexing included depending on plan |
| Off-the-shelf knowledge SaaS | GBP 12 to 50 per user per month | Weak on technical diagrams |
With 22 engineers asking 10 to 15 questions a day each, monthly volume is around 6,000 queries. At that level inference stays under GBP 250 or CAD 450 a month with a mid-sized model.
Security and confidentiality: the IT director's questions
Wiring diagrams and BOMs are industrial know-how. Typical IT requirements in Manchester or Toronto: data and index hosted in the UK or Canada, no training of vendor models on your data, single sign-on through the company directory, query logging, role-based access (a subcontractor only sees the machines they service). UK GDPR applies in Manchester, PIPEDA and Ontario guidance in Toronto.
RAG has a compliance advantage: documents stay in your database and the model does not memorise them. A manual update propagates as soon as it is reindexed, with no retraining. For clients exporting to the EU, the EU AI Act treats this kind of internal search aid as limited risk, with a transparency duty towards users.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Tom, head of after-sales service at the Manchester manufacturer, sizes the situation before the project: 22 engineers, 4 jobs a day each, 30 minutes of documentation search per job on average.
With a 30% cut in diagnosis time, the gain is 9 minutes per job. Calculation: 22 × 4 × 9 min = 792 min a day, i.e. 13.2 h, times 210 working days = 2,772 hours a year. At GBP 48 loaded hourly cost for a field engineer, that is GBP 133,056 of freed capacity. Against a full version budgeted at the top of the range, 7,000,000 FCFA (10,670 EUR), around GBP 3,000 of annual inference and Premium maintenance at 183 EUR a month (2,196 EUR a year), payback is counted in weeks, and first-time fix rate rises by 8 points.
FAQ
Can the assistant read wiring diagrams?
It indexes captions, reference tags and associated BOMs, which covers 70 to 80% of diagram questions. Fine-grained reading of vector diagrams is still limited in 2026, so the answer points to the relevant page.
What if the assistant finds nothing?
It says so and offers to escalate to a senior engineer. This refusal behaviour is configured from the start and tested on 200 real questions.
How long for a pilot?
Allow 6 to 8 weeks for one machine range and 2,000 to 3,000 pages, then 4 weeks of field use before deciding to extend.
Do German or French manuals need translating?
No, semantic search works across languages and the answer is written in English with a citation of the original source.
Does it work offline at a client site?
Not fully, it needs a 4G connection. A degraded mode caching the 50 most consulted procedures covers most jobs.
Let's scope your project. We audit a sample of your documentation, scope a pilot on one machine range and deliver it in 6 to 8 weeks before extending. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
