Put AI to work on
the customer’s mess.

A decade ago an FDE integrated software with software. Today, an enormous share of deployments have the same core: take a capable LLM and put it to work on the customer’s specific, messy data. The customer has documents no search can navigate, a workflow that eats an analyst’s afternoon, a decision made by reading unstructured text. The FDE’s job is to aim modern AI at exactly that — and the toolkit for doing so has settled into three reliable patterns.

This handbook is a map of that toolkit, written for the FDE who needs to choose and apply the patterns, not invent them. Each pattern has its own deep handbook on this site; here we cover what each is for, when to reach for it, and how they fit together into a solution you can honestly ship.

They compose: a typical deployment is a RAG system, sometimes wrapped in an agent that also calls the customer’s tools, gated by an eval that proves it clears the success bar. Understanding each — and, crucially, when not to reach for the fancier one — is the applied-AI half of the modern role.

Retrieval-augmented generation is the workhorse. The customer has a body of knowledge — policies, manuals, tickets, contracts — and wants the model to answer using it, accurately and with citations. RAG does this without retraining: chunk the documents, embed them, retrieve the chunks most relevant to a question, and hand them to the model as grounding so the answer comes from the customer’s truth rather than the model’s guesswork.

RAG is usually the first thing to try because it’s fast to stand up, grounds the model in verifiable sources, and updates the moment the documents do. The trade-offs versus other approaches are the subject of the RAG vs fine-tuning handbook.

When the task is more than “answer a question” — when it involves several steps, decisions, and actions in the customer’s systems — you reach for an agent: a model that can call tools, observe the results, and decide what to do next. An agent might read a ticket, look up a record, draft a response, and file it — the kind of multi-step workflow that used to need a human at every hop.

Agents are powerful and easy to over-reach with. The FDE discipline is to use the least autonomy that solves the problem: a fixed workflow with one model call beats a free-roaming agent when the steps are known, because it’s more predictable and easier to debug. And for any consequential action — sending an email, moving money, changing a record — keep a human in the loop. Two site handbooks go deep here: agent patterns for the how, and agents vs workflows for the crucial when-not-to.

An AI system without evals is a system you can’t honestly ship, because “it looks good in the demo” doesn’t survive contact with real inputs. An eval is a set of representative examples plus a way to score the output — a test suite for behavior that’s fuzzy rather than exact. For an FDE, it’s the bridge back to discovery: the success metric everyone agreed on becomes the eval you measure against.

Evals turn AI from a vibe into engineering. They’re deep enough to warrant their own study — agent evals and evals in CI — and worth building a reflex for on the interactive challenges.

The meta-skill is restraint: reach for the simplest pattern that clears the bar. A well-crafted prompt beats RAG when the knowledge fits in context; RAG beats fine-tuning when the need is knowledge, not behavior; a fixed workflow beats an agent when the steps are known. Complexity is a cost you pay in reliability and debuggability, so add it only when a simpler pattern demonstrably falls short.

And ship it honestly. Most AI solutions aren’t perfect, and that’s fine if the workflow is designed so mistakes are caught, not blindly acted on. Define the good-enough bar in discovery, keep a human in the loop for consequential decisions, be upfront about failure modes, and — because the model reads untrusted customer content — treat every input as potentially adversarial (the security handbook covers prompt injection). An FDE who ships an honest 85% that safely handles the other 15% delivers real value; one who oversells a fragile 95% loses the customer the first time it breaks. That honesty — about what the system can and can’t do — is the last piece of the whole FDE craft.