Skip to module content
🧠AI Academy
0/91
Sign in
Module 03 · ~10 min

Opportunity Selection — Which Workflows, Which Shape

Two axes, one bias to correct, and a shape ladder — replacing 'AI brainstorming' with an actual method.

Reading progress
0/7 · 0%

The big idea

💡Key idea
Good opportunity selection runs three filters in sequence. First: does this workflow have real friction and a closable learning loop — a place where AI output can be checked and corrections fed back in? Second: can the solution be made domain-specific (built around the vocabulary and rules of one narrow function) AND deeply integrated (living inside the process, not beside it as a chat window)? Third: what is the simplest technical shape that serves this workflow — a single model call, a composed pipeline, or a full agent? A candidate that passes all three filters is rare. That's the point. The 5% of AI projects that survive weren't lucky; they were selective.
Quick check
1 question · instant feedback
0/1
  1. Two axes that predict pilot-to-production survival:

Numbers that matter

50%+
of GenAI budgets flow to sales & marketing while measured ROI clusters in back-office ops
Quick check
1 question · instant feedback
0/1
  1. Anthropic's guidance on the shape ladder is:

Deep dive

3/3 open

Replotting the GenAI Divide data reveals a clean pattern. Projects survive the pilot-to-production cliff based on two variables:

**Domain specificity** — does the system actually embed the vocabulary, rules, and edge cases of one narrow business function? Or is it a generic model that happens to be pointed at a new problem?

**Workflow integration** — does the system live *inside* the process, with memory and context, so it shapes decisions in real time? Or does it sit *beside* the process as a chat window people open when curious?

Low on both axes is where the 95% of failing projects live — generic chatbots dropped into processes they don't touch. High on both — deeply integrated, domain-specific, most often delivered by a specialized vendor — is where success rates roughly double.

More than half of all GenAI budgets flow to sales and marketing — the visible, demo-friendly use cases that look impressive in board meetings. Meanwhile, MIT's measured returns cluster somewhere much less glamorous: operations, finance, and compliance. Document processing, claims handling, reconciliation — the workflows firms previously outsourced precisely because they were high-friction and low-glamour.

Your prospects' competitors are over-investing where ROI is thinnest. The contrarian, evidence-backed move: inventory the boring workflows first. Score friction. Pitch accounts payable before pitching the AI sales assistant.

Think of your architecture options as a ladder with three rungs — and a strong prior to stay as low as possible.

**Rung 1: A single augmented model call.** One call to a model, but with retrieval (pull in relevant documents), tools (let it take actions), or memory (give it context from prior interactions) attached. Anthropic's explicit advice: for many production applications, this is genuinely all you need.

**Rung 2: Composable workflow patterns.** Prompt chaining (output of one call feeds the next), routing (classify the input, then dispatch to a specialist), parallelization (split a task or run independent checks and vote), orchestrator–workers (a coordinator breaks work into subtasks dynamically), evaluator–optimizer (generate a draft, critique it, refine it). These patterns handle medium-complexity tasks without full autonomy.

**Rung 3: Autonomous agents.** Only for workflows where the path genuinely cannot be enumerated in advance — and only with the tool-surface limits and budget guards from Volume 1 §1.

The rule: climb only when the rung below demonstrably fails. Agent-first ambition where a routed workflow was the honest shape is a documented failure mode, not a sign of ambition.

Quick check
1 question · instant feedback
0/1
  1. Where does measured ROI actually cluster in the MIT data?

In the field

🔬Worked example
The 2×2 matrix: LOW domain / LOW integration = 'hype experiments' (generic chatbots, home of the 95%). HIGH domain / LOW integration = 'niche point tools' (right vocabulary, no process contact). LOW domain / HIGH integration = 'generalist copilots' (adoption without ROI). HIGH domain / HIGH integration = 'survivors & scalers' — vertical, compliance-aware, learning-capable systems in high-friction workflows. Where the 5% lives.
🚫When not to reach for it
Use-case brainstorming workshops that generate 40 excitement-ranked ideas are exactly what the evidence says fails. Inventory workflows and score friction instead. And do not climb the shape ladder ahead of demonstrated need — Gartner-tracked agentic projects stall at high rates because teams reached for autonomy where a routed workflow was the honest shape.
Quick check
1 question · instant feedback
0/1
  1. A client wants 'an AI agent' for a fully predictable weekly report. Best call:

Pitfalls & takeaways

Failure modes

  • Use-case brainstorming. Workshops that generate 40 excitement-ranked ideas; the evidence says inventory workflows and score friction instead.
  • Glamour bias. Budget follows the demo; ROI was in accounts payable all along.
  • Agent-first ambition. Reaching for autonomy where a routed workflow was the honest shape.
  • Building what should be bought. Ignoring the ~67% vs ~33% deployment gap out of pride or 'not invented here' security posture.
  • One giant bet. A single flagship 'transformation' instead of a portfolio of narrow, instrumented workflows that can fail cheaply and compound individually.

Durable takeaways

  • Two axes: domain specificity × workflow integration. High/high wins.
  • Score boring back-office workflows first — that's where ROI lives.
  • Simplest shape that serves it. Climb only when the rung below fails.
  • Portfolio of narrow bets beats one flagship transformation.

Do the work

📦Artifact to produce
AI Opportunity Radar entry: workflow inventory → 2×2 gate → rubric score → shape decision.

Sources

  • · MIT NANDA, The GenAI Divide
  • · Harvard Business School — competing frames
  • · Anthropic, Building Effective Agents