Translating a claim

Lesson 4 of 4 in The Words People Use — Agent, Agentic, Copilot, Assistant.

Time to use it. Below are phrases lifted from the shape of real product pages and job postings. Sort each one by what is actually being described — and note the third bucket carefully.

“Cannot tell from this sentence” is a correct answer, and it will often be the right one. That is the skill, not a failure of the exercise. Most marketing sentences are genuinely underdetermined: they were written to be attractive to several kinds of buyer at once, which means they cannot commit to a control-flow claim. Reaching “cannot tell” quickly, and knowing exactly which follow-up resolves it, beats guessing confidently.

Interactive sorting exercise: Sort each claim by what it tells you about control flow. Remember: “cannot tell” is a legitimate verdict, and here it is the most common one.

Notice the pattern in the cards that did resolve. None of them resolved because of a noun. They resolved because someone described behaviour on a specific input — identical steps every run, fourteen enumerated branches, a new hypothesis after an unfamiliar failure. Whenever a claim is decidable, it is decidable for that reason.

So in a live conversation you do not need a vocabulary argument. You need three follow-ups, in this order: what happens on an input nobody anticipated? what can it do without asking? what stops it? The first locates control flow, the second sizes the blast radius, the third tells you whether anyone has thought about the run ending badly. Two more are held in reserve for when an answer stalls — can you show me the graph? and walk me through one concrete run — and the tree below wires all five together. Walk a real conversation through it.

Decode a vendor claim in three questions — plus two for when an answer stalls

Interactive decision tree — outcomes:

  • Workflow-shaped — whatever it is called

    The path is fixed by humans, so you can regression-test the graph and bound behaviour by construction. Your remaining work is per-step quality (each model-powered node needs its own evals) and brittleness when upstream systems change. If the sales material said “agentic,” that is fine — just plan for this system, not that word.

  • Model-directed, but reads only until a human approves

    Control flow is the model’s; write authority is not. This is a good launch posture — the blast radius is capped by the gate rather than by hope. Two checks before you relax: does the reviewer see enough to genuinely say no, and is anyone tracking review time and rejection rate? A gate that always says yes has quietly stopped being a gate.

  • Model-directed with real authority, and a containment story

    The honest configuration for serious work: the model chooses steps, it can act, and somebody has named the limits and the stop control. Now buy the evidence rather than the design — ask for pass rates over many runs on a representative task set, ask to see a trace, and ask who reads the logs on a Tuesday.

  • Model-directed, write authority, no stopping story

    This is the combination that produces incidents: the model picks the path, the path can change production, and nothing has been named that ends a bad run. Not a reason to walk away — a reason to make containment a requirement before pilot. Start by shrinking authority to reads plus a gate, then earn the rest back with measured pass rates.

  • Verdict: unknown — and that is your finding

    You did not fail to get an answer; you learned that nobody on the other side can describe their own control flow, tool inventory, or stopping conditions. Those are the same facts they would need to test the system or respond to an incident. Ask again in writing, to an engineer, about one concrete run — and treat a second non-answer as a real risk signal rather than a communication problem.

Interactive flashcard deck.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.