How a forward deployed engineer ships an AI agent to production

An AI agent can impress in a demo within days. Putting it into production, inside a company's tools, with its data, rules and users, is another matter. That is exactly the job of a forward deployed engineer: start from the field, build in the real environment and only stop when the agent is in use. This article lays out the method, step by step, as we apply it on our projects.

Why AI agents get stuck at the prototype

• The wrong use case: picked because it is impressive, not because it is frequent and costly.
• Inaccessible or messy data: duplicate documents, contradictory versions, knowledge that lives in people's heads.
• No integration: an agent that can neither read the CRM nor create a ticket is a toy.
• No evaluation: nobody can say whether the agent is right 60% or 95% of the time.
• Security discovered too late: access rights, personal data, human approval.
• Adoption forgotten: the tool ships and nobody changes how they work.

Step 1: start from a process, not a technology

The FDE starts by watching real work: who does what, how often, with which tools, where time is lost. The right use case is frequent, measurable and bounded: answering order-status questions, qualifying inbound requests, drafting a first quote, finding a clause in contracts. Our article on AI agents for small businesses gives concrete examples.

Step 2: map the data and tools

• Sources: documents, knowledge base, CRM, ERP, email, tickets, product database.
• Their quality: current, duplicated, contradictory? Cleaning is part of the project.
• The actions the agent must take: read, create, update, send, and which ones need human approval.
• Permissions: the agent should only see what the relevant user is allowed to see.

Step 3: build the foundation (RAG, tools, guardrails)

The agent rests on three building blocks. Retrieval (RAG) gives it the right information with sources, see RAG explained. Tools let it act in your systems through APIs or automations (see n8n workflows). Guardrails frame its behaviour: answer scope, polite refusals off-topic, handoff to a human, confirmation before any sensitive action.

Step 4: evaluate before deploying

An FDE does not ship on gut feeling. They build a test set from real questions and real cases, with expected answers, then measure the agent: accuracy, correct sources cited, rules followed, response time, cost per exchange. The test set is re-run after every change (new document, model switch, new rule) to catch regressions.

Step 5: secure it

• Accounts and API keys in the company's name, never the contractor's.
• Business plans from model providers that exclude training on your data. See protecting company data.
• Least-privilege access per tool, logged conversations, a defined retention period.
• Protection against malicious instructions hidden in documents or inbound messages.
• Compliance with the privacy laws that apply to you, such as GDPR for EU data.

Step 6: roll out gradually

1. Suggestion mode: the agent drafts, a human approves and sends.
2. Narrow scope: one channel, one team, one request type.
3. Expand when the metrics hold: more cases, more autonomy on low-risk actions.
4. Continuous monitoring: a dashboard of conversations, escalations and errors.

For a customer service agent, our guide to deploying an AI customer service agent covers this rollout in detail.

Step 7: drive adoption

This is where an FDE differs most from a plain developer. They train teams, explain what the agent does and does not do, gather feedback in the first weeks and adjust fast. People adopt a tool that saves them time in week one and improves when they report a problem.

Step 8: measure and improve

Metrics depend on the use case: handling time, share of requests resolved without a human, satisfaction, reported errors, cost per request. Review them monthly with an improvement plan. An AI agent is not a finished project, it is a living service.

We work the forward deployed way

Our AI agents service follows these eight steps: we embed with your teams, build in your environment, measure on your real cases and stay through adoption. Share your use case and we will tell you what can run in production, how soon and at what running cost.