Skip to main content

Navigating Ambiguity

5 min read

Tpm

Your job: turn 'we need better analytics' into something buildable. AI needs that output. You produce it.

Tech Lead

Stakeholders say different things. You synthesize. AI can't do the synthesis—it wasn't in the three meetings.

Solutions Eng

The customer said 'integration' but meant 5 different things. You tease that out. AI gets clean input from you.

Navigating Ambiguity

TL;DR

  • AI excels where truth is explicit—syntax rules are stable, errors aren't debatable. When ambiguity dominates, "the anchor disappears." No ground truth, unbounded search, models hesitate.
  • AI agents struggle to distinguish well-specified from underspecified instructions. Given vague requirements, they often make unwarranted assumptions rather than asking clarifying questions—leading to suboptimal outcomes and wasted compute.
  • SpecFix (automated description repair) improves code generation by up to 30.9% on modified descriptions. Translation: clarity pays. You provide that clarity. AI gets the clarified input and does better. You're upstream.

AI is great when the problem is well-defined. "Write a function that does X." Clean input, clean output. Real work is messier. Humans retain an advantage: a lifetime of unclear facts. AI's logic requires true premises. Your life is saturated with social ambiguity—tone, implication, uncertainty. AI isn't. Someone has to clean it first. That someone is you.

Why Ambiguity Breaks AI

The Anchor Disappears

  • In typical coding tasks, AI has a confident anchor. Syntax rules are stable. In ambiguous environments—noisy perception, unclear requirements, social context—that anchor disappears. Example: historical census forms with smudges, ink bleed, cropped scans. OCR produces estimates, not ground truth. Without ground truth, no anchor. The search space becomes unbounded. Models hesitate.
  • Even minor imperfections in task descriptions cause significant performance drops. Contradictory descriptions produce numerous logical errors. Larger models are more resilient but not immune.

AI Assumes Clarity

  • "Build a dashboard." — What metrics? For whom? Real-time or batch? AI will guess. Its guess might be wrong.
  • "Improve performance." — Latency? Throughput? Perceived speed? AI doesn't know. It optimizes something. Maybe the wrong thing.
  • "Make it scalable." — 10 users or 10 million? AI assumes. You know.

Contradictory Inputs

  • Product says: "We need it fast." Engineering says: "We need it right." Sales says: "We need it by quarter end." AI can't resolve that. You broker.
  • "It should be simple and feature-rich." — Trade-off. AI might give you both and create a mess. You decide what "simple" means in practice.

Missing Information

  • "The customer wants better reporting." — What reports? What's "better"? Who's the audience? AI fills in with generic. You discover the real requirements.
  • "We need to migrate off the legacy system." — Why? What's the trigger? What's in scope? AI can't run the discovery workshop. You do. When AI agents can interact with users to resolve underspecified inputs, they achieve significant performance gains. You're the human in that loop.

Your Value: The Clarification Layer

You Ask the Questions

  • "When you say X, do you mean A or B?" — Disambiguation. AI gets the answer. You get the question.
  • "What's out of scope?" — Boundaries. AI doesn't know what not to build. You define it.
  • "What does success look like?" — Criteria. AI optimizes for something. You define what.

You Synthesize

  • Three stakeholders, three opinions. You produce one coherent spec. AI works from that. You created it.
  • "We need A, B, and C—but we only have budget for two." Prioritization. AI can't do it. You do.

You Iterate

  • First version of requirements is always wrong. You learn that in review. You update. AI gets better input in round 2. You're the feedback loop. SpecFix proves it: repair the description, get 30%+ better code. You're the repair mechanism before the model runs.

How to Use This as a Moat

  1. Own the discovery. Before you prompt AI, do the messy work: interviews, whiteboarding, "what does that actually mean?" The better your input, the better AI's output. And the input is your skill.
  2. Document assumptions. When you give AI a task, write down what you're assuming. "We're optimizing for X, not Y." "Out of scope: Z." That doc is the contract. AI executes. You own the contract.
  3. Treat AI as a tool for the clarified problem. Once you've reduced ambiguity, AI accelerates. Before that, it amplifies confusion. Order matters.

Quick Check

Product says 'We need it fast.' Engineering says 'We need it right.' Sales says 'By quarter end.' What can AI do with that?

You get 'Build a dashboard.' You build something. It's wrong. 'We wanted X, not Y.' Rework. Or you spend weeks in meetings before you touch code.

Click "Clarified Then AI" to see the difference →

Do This Next

  1. Take one vague ask from your backlog (or invent one). Write down 5 clarifying questions you'd ask before you could build it. That's the ambiguity-navigation skill. Practice it.
  2. Before your next AI prompt, add one explicit constraint. "Assume we're on Postgres." "Assume we have 2 months." See how it changes the output. You're adding the anchor AI lacks. Control the ambiguity.