Choosing What To Build

Choosing what to build

Deploying an agent to do rule-based lookup is hiring a neurosurgeon to take your blood pressure.

---

The mistake this prevents

The most expensive error in these programmes is picking the wrong tool for the job, and it goes wrong in both directions.

An agent doing deterministic lookup is slow, expensive, non-repeatable and impossible to audit, for work a script did correctly. A rules engine handed an unstructured appeal letter fails spectacularly and expensively, because there was never a rule to write.

Neither failure is a technology failure. Both are classification failures, and both are cheap to avoid with three questions.

This page assumes you already have a candidate. Finding the work is how candidates are found and scored; this is how you decide what shape the answer takes.

Three gates

Every candidate passes through them in order, and the first "no" ends the sequence.

	The question	Yes	No
Gate 1 · Structure	Is the process fully structured, with deterministic rules?	Go to gate 2	Skip to the model track. There is no rule to encode
Gate 2 · Judgement	Does it require human-like judgement or interpretation?	Go to gate 3	Traditional automation. Do not put a model on it
Gate 3 · Complexity	Does it need multi-step reasoning, memory, or tools across systems?	An agentic system	A single model step in a workflow

Gate 2 is the one that saves the most money and gets skipped the most. A structured process that needs no judgement is an automation problem that has been solved for twenty years, and putting a model on it converts a deterministic, auditable, cheap process into a probabilistic, expensive one.

Gate 3 is the one people get wrong in the other direction. Most work that clears gate 2 does not need an agent. It needs one model call inside an ordinary workflow, which is faster to build, cheaper to run and vastly easier to reason about when it breaks.

The selection matrix

Once the gates give you a track, this is what you are committing to.

	Traditional automation	Model in a workflow	Agentic system
Input	Structured, predictable	Unstructured or variable	Unstructured, multi-source
Decision logic	Deterministic rules	Pattern-based, probabilistic	Reasoning and planning
Memory	None, or session only	Single-session context	Persistent and cross-session
Tool use	One system	One model, one or two tools	Multi-tool orchestration
Human oversight	Exception handling	Review before action	In the loop by design
Setup time	Weeks	A month or two	A quarter or more
Risk profile	Low to medium	Medium	Medium to high
Governance	Standard change control	Tier review	Tier review plus a risk charter
Example, finance	Payment reconciliation	Drafting a dispute response	End-to-end dispute resolution across three systems
Example, logistics	Order status lookup	Interpreting a damage report	Exception resolution across carrier, warehouse and customer

Read the setup-time row as a commitment, not an estimate. The gap between the first column and the last is not a scheduling difference, it is a difference in what you are agreeing to operate afterwards. Nothing in the last column is ever finished, because it needs a registry entry, a drift check and a rung on the ladder for as long as it runs.

The oversight row is the honest one. Traditional automation is supervised by exception. An agentic system is supervised by design, which means the human cost does not go to zero. It changes shape.

The determinism axis

The gates are a decision procedure. The axis underneath them is worth naming separately, because it is how experienced people make the same call in ten seconds.

High determinism: clear rules, fixed outcomes, one right answer. Put it on rails. You are building a train, and a train that leaves the rails is a disaster rather than a variation.

Low determinism: nuance, empathy, negotiation, judgement. You are flying a helicopter. It goes where a train cannot, and it needs a pilot the whole time.

Most real work sits in the middle, and that is exactly where the mistakes are made, because a process that is eighty percent deterministic looks deterministic in a workshop.

The practical move for the middle: split it. Put the deterministic eighty percent on rails, and route the remainder to the model track with the human oversight the last twenty percent actually needs. That is almost always cheaper and more auditable than treating the whole process as low-determinism, and it is the shape most successful systems here have taken.

How this relates to the rest

	Answers	Where
Opportunity score	Is there work here worth automating at all?	Finding the work
Three gates	What kind of thing should we build?	This page
BXT	Should this candidate be funded ahead of the others?	The framework spine
Three proof gates	Have we proved it before scaling it?	The three proof gates

Running these out of order is the common failure. Funding a candidate before classifying it means committing a budget shaped like automation to a problem shaped like an agent, and the overrun is discovered at gate 3 of the proof model rather than in the first conversation.

_{Adapted from the AI CoE and Agent Factory Playbook, based on The Augmented Enterprise framework. The split-the-process recommendation and both example rows are the author's own.}

_{Back to Home · Finding the work · The three proof gates}