Most of the writing about FDEs is abstract. Here's what the role looks like in practice for an AI deployment, drawn from how we and others in the agent ecosystem run it today.

Week 1: discovery, not building. The FDE shows up to the customer's standups. They sit in on operations calls. They ask questions that sound basic and produce surprising answers. Where does this invoice come from? Why are these two vendors handled differently? Who decided that threshold? What happens when the system disagrees with the controller? The point of week one isn't to write code. It's to map the real workflow against the documented workflow and find the gap.

There's always a gap. The documented process is the version that exists in onboarding decks. The real process is the version that runs every day, full of exceptions, workarounds, and unwritten rules. The agent has to work against the real process, not the documented one.

Week 2: integrations. Now the FDE writes code. They pull data out of the customer's systems. Sometimes that's a clean API call. Sometimes it's a SOAP endpoint that requires a specific certificate. Sometimes it's a homegrown application the controller built in 2013 that nobody else fully understands. Sometimes it's a vendor portal with a single login that doesn't support service accounts.

This week is the part of the FDE role that looks most like traditional engineering. It's also the part that compresses the most as tooling matures. What took two weeks of integration work in 2023 takes three days in 2026 because the integration surface area is mostly product now.

Week 3: capturing tribal knowledge. The FDE sits with the controller, the operations manager, the finance analyst, the warehouse supervisor. They walk through real cases together. The dual-invoice quirk where one vendor sends a PDF and a duplicate email and the agent has to dedupe them. The verbally-renegotiated discount terms where the master contract says one number but the email thread says another. The OCR exception where one specific supplier sends invoice scans rotated ninety degrees because their scanner has a Japanese locale setting nobody has changed in eight years.

None of this is in the system. All of it is in the operator's head. Week three is the work of getting it out of their head and into agent policy.

Week 4: first agent in sandbox. The FDE builds against the platform. Configures the agent. Wires the retrieval. Writes the policy. Sets the escalation thresholds. Runs the agent against sanitized real data. Watches what it does. Tunes. Re-runs. Tunes again. This is the empirical loop the non-determinism of agentic systems requires.

Weeks 5-6: production cutover. Read-only first. Then read-write with human-in-the-loop approval on every action. Then read-write with thresholds (auto-execute under X dollars, approval over X dollars). Then full autonomy with audit log. Each step takes a few days of watching, tuning, and convincing the security team and the operations team that the next step is safe.

Weeks 7-8: knowledge transfer. The FDE trains the platform admin, the operations team that will triage exceptions, and the security team that will review the audit log. The goal is for the customer's team to operate the agent without the FDE present. If the customer can't run it themselves by the end of the engagement, the engagement isn't done.

This is the deliberate distinction from a consulting engagement. Consultants extract value over the length of the engagement and disengage. The FDE engagement is designed to make the FDE unnecessary at the end of it.

The FDE isn't a consultant who parachutes in. They're a builder who embeds, joins the customer's Slack, attends their standups, learns their stack, and stays until the agents are live and outcomes are proven.