Technical assistant over GMP equipment datasheets
94% queries resolved without human intervention
Pillar — Industrial Artificial Intelligence
AI Agents for Industry — A Complete Guide for Industrial SMBs
How to deploy an enterprise AI agent in an industrial SMB: RAG architecture on private data, cloud vs on-premise, GDPR compliance and real costs without per-conversation licensing. An honest technical guide based on real projects.
What is an enterprise industrial AI agent?
An enterprise AI agent is not ChatGPT with your logo: it is a system with controlled access to your data (catalog, manuals, ERP, machine logs), able to execute measurable actions within your business and auditable end-to-end. The difference with a generic chatbot is structural.
- Knows your actual catalog with prices, stock and availability — does not respond with generic internet data.
- Has customer memory (CRM) and reasons over order history, commercial terms and previous projects.
- Executes actions: creates quotes, opens tickets, schedules technical visits, notifies the salesperson.
- Quotes the exact source (manual, datasheet, email) — meets documentary traceability.
- Runs within your infrastructure: sensitive data does not leave the perimeter.
Example: Real ISIGECO case: an agent that answers technical queries over 600 industrial products by quoting the corresponding datasheet. Average response time: 4 seconds. Validated accuracy: 94% across 200 test queries.
Real use cases in industrial SMBs
These are the cases where an AI agent delivers measurable return in a mid-sized industrial company. Not experiments: each case has a business metric behind it.
- Web HMI with conversational assistant: the operator asks in natural language about a line's status and receives PLC data plus an action recommendation.
- Predictive maintenance: an agent correlating machine logs, failure history and OEM manuals to anticipate failures 7-15 days in advance.
- Datasheet technical assistant: engineers and salespeople ask the agent and get answers citing the exact PDF page.
- Lead qualification: the agent interviews the lead, identifies purchase intent, validates BANT and routes to the right rep.
- Quote automation: from a customer email, the agent extracts the BOM, calculates price under customer terms and generates the PDF.
Example: Typical metric after 3-6 months: 60-70% of technical queries resolved with no human intervention; average savings of 12-18 hours/week for the sales team; lead response time from 24h to 2 min.
RAG on private data (catalog, manuals, ERP)
RAG (Retrieval-Augmented Generation) is the standard technique to make the model answer only on your verifiable data and never «invent». The difference with fine-tuning is radical: with RAG you do not train the model with your data — you retrieve it in real time.
- Your PDFs (datasheets, manuals, certificates) are chunked and indexed in a local vector store or in a DPA-bound provider.
- Every agent response is built by first searching relevant fragments and only then generating text.
- Full traceability: each answer carries the exact citation (document, page, paragraph) and the operator can verify it.
- Change a PDF → after re-indexing, the agent is up to date with no retraining.
- Your data is not «fused» into an opaque model: it lives in your database, controllable and deletable.
Example: Typical ISIGECO stack: pgvector or Qdrant for the vector store, BAAI/bge-m3 embeddings (multilingual) or OpenAI text-embedding-3-large, orchestration via LangGraph or LangChain. Cost to index 1,000 technical PDFs: ~€30-50 in embeddings (one-time fee).
Cloud vs on-premise — when to choose each
No deployment is «better» in the abstract: it depends on the data the agent handles, the regulatory framework of the sector and the real query volume. Honest criteria below.
- Cloud (Anthropic Claude API, OpenAI API, Azure OpenAI): best latency, top-tier models, auto-scaling. Recommended if your data has no sensitive PII or confidential blueprints.
- On-premise with open-weight models (Llama 3.3, Qwen 2.5, Mistral): full control, data never leaves. Requires 1 NVIDIA GPU (from RTX 4090 or A6000). Quality: 80-90% of cloud.
- Hybrid: most operations in cloud and sensitive data (BOM with margins, customer drawings) stay local.
- Real on-premise cost: hardware one-off (~€6-15k), electricity ~€30/month. No per-conversation fees.
- Real cloud cost: €0 hardware, ~€0.003-0.015 per query. Economical up to ~50,000 queries/month.
Example: Practical rule: under 50,000 queries/month and no especially sensitive data → cloud. Regulated sectors (pharma, defense) or competitive catalogs → on-premise. Standard industrial SMB → hybrid (cloud for reasoning, local vector store).
GDPR compliance and sensitive data
Any AI deployment in a European company has to pass the GDPR filter and, from 2026, the EU AI Act. The good news: most industrial cases are low risk. The important news: it has to be documented.
- If the agent handles personal data (customers, leads, employees), it must appear in the Record of Processing Activities (Art. 30 GDPR).
- With a cloud provider (OpenAI, Anthropic, Azure) you must sign a DPA and, if the provider is in the US, verify SCC and the EU-US DPF framework.
- EU AI Act: most industrial agents are «minimal» or «limited risk» (transparent chatbot). Only those making critical automated decisions are «high risk».
- Right of access, rectification and erasure: the user must be able to view/delete their conversations.
- Audit trail: every response must be reconstructable (prompt, retrieved context, model, timestamp).
Example: ISIGECO ships every agent with its simplified DPIA, its Record of Processing Activities and EU AI Act documentation so the client can prove compliance to AEPD if needed.
Real cost (without per-conversation licensing)
This is where most commercial proposals mislead. The number that matters is not «X € per conversation» but the total 12 and 36-month cost compared to actual savings.
- Base ISIGECO project for an industrial SMB agent: €8-25k of one-off development, depending on integrations (Holded, SAP, Odoo, custom ERP).
- Monthly maintenance: €200-600/month (monitoring, model updates, prompt fine-tuning, reindexing of new documents).
- Inference cost (cloud): ~€0.003-0.015 per query, billed by the provider (Anthropic, OpenAI). Can be passed through to the customer.
- No «per-user licensing» nor «per-conversation» fees: the code and (in on-premise) the model weights are yours.
- Typical ROI: 6-12 months if the agent covers ≥1 real use case (quoting, technical assistant or lead qualification).
Example: Real comparison: on-premise AI agent — hardware €9,000 + integration €12,000 + maintenance €300/month = €24,600 over 24 months. A full-time admin doing the same tasks: €35,000-45,000/year in Spain. The agent pays for itself in under a year.
Use cases
Automatic lead and RFQ qualification
Response time from 24h to 2 min
Web HMI with conversational agent
40% reduction in unplanned downtime
Quote generation from customer email
14 h/week saved for sales team
Frequently asked questions
How much does an AI agent cost for an industrial SMB?
A base cloud project is typically €8-15k of one-off development, with maintenance from €200 to €600/month. Inference (model consumption) is prepaid to the provider (Anthropic/OpenAI) and costs €0.003-0.015 per query. An on-premise case is €18-25k due to hardware but eliminates variable inference.
Do I need my own GPUs?
Only if you want fully on-premise. For most industrial cases connecting to a cloud provider (Anthropic Claude, OpenAI or Azure OpenAI) is enough, no GPU required. For full autonomy, an NVIDIA RTX 4090 (~€2,500) already runs quantized Llama 3.3 70B with decent quality.
Is it trained on my data? Does my data leave my infrastructure?
It is not trained on your data: we use RAG, which queries your document base in real time and only sends the model the fragments needed to answer. With a signed DPA, cloud providers do not train on your inputs. For absolute guarantee, an on-premise deployment ensures nothing ever leaves.
What happens if the model becomes outdated?
Model and document base are independent. If Claude 4.8 or Llama 4 is released, we switch endpoints and the agent keeps working with your data. Your documents reindex automatically when updated. There is no risk of being locked into a single version.
Can it connect with my SAP/Odoo/Holded?
Yes. ISIGECO integrates the agent with your ERP via its standard API (Holded, Odoo, SAP S/4HANA, Microsoft Business Central) or, if absent, via a custom connector. The agent can read orders, stock, customer terms and create new orders, quotes or tickets.
How is it different from ChatGPT Enterprise?
ChatGPT Enterprise is a generic interface: the model does not know your catalog, does not connect to your ERP and does not execute actions. An ISIGECO agent is built over your data, integrated with your systems and trained in your company's processes. The difference between an intern and a specialist.
How long until it's operational?
A functional MVP (1 use case, ~50-200 indexed documents, no ERP integration) is ready in 3-5 weeks. A full production agent with ERP integration, monitoring and admin panel takes 8-14 weeks. We work in two-week sprints with client validations.
Who maintains the system?
ISIGECO offers monthly maintenance (€200-600/month) covering monitoring, prompt tuning, indexing of new documents, model updates and support. Alternatively, we train your IT team to run it in-house: most of the code is standard Python and JavaScript.
Want to know if an AI agent will pay off in your SMB?
We give you an honest evaluation in 30 minutes: if the ROI is unclear, we tell you. If it is clear, we deliver a detailed proposal with cost, timeline and success metrics.
Request free evaluation