Almost every company we talk to has already “done something with AI”: a test chatbot, a script that summarizes emails, a demo that impressed the board. The demo isn’t the hard part. The hard part is making that pilot survive contact with real operations: messy data, permissions, costs that grow with usage and users who expect it to work the same on Monday as on Friday.
In this article I share the approach we use to take generative AI from prototype to production inside business processes.
Start with the process, not the model
The wrong question is “what can we do with GPT or Claude?”. The right one is “which decision or repetitive task is costing my team time and money?”.
The best candidates usually share three traits:
- High volume and low risk per unit: classifying tickets, extracting invoice data, drafting first versions of replies.
- Scattered information: the knowledge exists, but it lives in PDFs, email threads and in the heads of two people.
- A human already validates the result: if someone reviews the work today, AI can prepare the draft and that person moves from doing to supervising.
If the process has no clear owner and no success metric (time per case, error rate, tickets resolved), it isn’t ready to be automated yet.
RAG: let the model talk to your data, not its memory
A language model knows a lot about the world and nothing about your company. The most practical technique to close that gap is RAG (Retrieval-Augmented Generation): before answering, the system searches your documents for the relevant fragments and hands them to the model as context.
A healthy RAG pipeline has four pieces:
- Ingestion: turn documents, tables and tickets into clean fragments with metadata (date, area, access level).
- Indexing: store those fragments in a vector database alongside keyword search. Hybrid search almost always beats either one on its own.
- Permission-aware retrieval: the system should only bring back what the person asking is allowed to see. This is the point pilots forget, and the first one that blocks a real deployment.
- Generation with citations: every answer should say which document it came from. If there is no source, the system should say “I don’t know”.
The quality of a RAG system depends more on how you prepare and retrieve the data than on the model you choose.
Integrate, don’t add another tab
AI creates value when it shows up where the team already works. Instead of an isolated chat, we connect the model to the CRM, the ERP or the help desk through APIs and automation tools such as n8n.
With protocols like MCP (Model Context Protocol) the model can use tools in a structured way: check an order’s status, create a task or read a CRM record, always through functions you define and control. The model proposes the action; your system decides whether it runs.
A pattern that works well:
// The model can only call explicit, typed tools
const tools = {
getOrder: { input: z.object({ id: z.string() }), run: crm.getOrder },
createTicket: { input: TicketSchema, run: helpdesk.create, requiresApproval: true }
}Read actions can be automatic; write actions that affect customers or money go through human approval, at least at first.
Measure before you scale
No evaluation, no production. Before opening the system to the whole team we build a set of real cases (50 to 200 are usually enough to start) with the expected answer, and we run it every time we change the prompt, the model or the way we retrieve data.
The metrics we find most useful:
- Accuracy against the expected answer.
- Faithfulness: does the answer rely on the cited sources or does it make things up?
- “I don’t know” rate: a system that never hesitates is a risk.
- Cost and latency per case: the biggest model is not always the right one.
Security and costs from day one
Three rules that save us surprises:
- Sensitive data: anonymize or mask personal information before sending it to an external provider, and review their retention policies.
- Prompt injection: treat all text coming from documents, emails or the web as data, never as orders for the model.
- Budget per workflow: log tokens and cost per use case. That way you catch a prompt that grew out of control in time.
A 90-day path
Our recommendation to go from idea to production:
- Weeks 1–3: choose a process, define the metric and gather the data.
- Weeks 4–8: build the workflow with RAG and tools, plus the evaluation set.
- Weeks 9–12: pilot with a small team, a human in the loop, and tuning based on the metrics.
AI doesn’t replace your team’s judgment: it multiplies it. When it is well integrated, your people stop copying and pasting information and get to focus on deciding.
If you want to identify which process in your company has the most potential, book a call with our team.
