Answers

The questions you'd ask in the first call.

Hedged where the truth is hedged. Ranges where a point value would be a lie. And a No section, because that's also an answer.

A — Investment tier

Questions about money.

Automations with a line item: what it costs, what it returns, and how honest the numbers are.

“Will an AI agent actually pay for itself?”

Usually — when the workflow is repetitive, measurable and high-volume. A qualified lead flow of 200+/month or 6+ hours of weekly manual reporting clears the bar. Below that, the honest answer is often a $40 tool and a better template.

Typical payback3 — 6months
Build cost$8k — $45k

“What does it cost to run after you leave?”

Hosting and model spend land between $80–600/month depending on volume; most builds sit near the low end. You get the metering dashboard with the handover — the bill is never a surprise we walk away from.

Run cost$80 — 600/mo
Owner lock-in$0
B — Operations tier

Questions about the daily grind.

Inbox triage, approvals, reporting — the work that happens whether or not anyone notices it.

“Can my inbox stop being a to-do list?”

Mostly. Classification, drafts and routine replies are reliable today; ambiguous messages get summarised and queued, not guessed at. Expect 60–80% of volume handled without you, with a review queue for the rest.

Handled untouched60 — 80%
Review queue20 — 40%

“Can approvals move without me chasing people?”

Yes, for rule-bounded approvals. Large exceptions keep a human gate by design — we will not automate judgement your regulator attributes to you. The chasing itself is fully automatable, and it's usually 90% of the pain.

Approval cycle3.2d — 0.5d
Chase messages−90%

“Can the weekly numbers assemble and check themselves?”

Mostly. Assembly and reconciliation are fully automatable; variance commentary stays human-reviewed with suggested drafts. First working version in 2–4 weeks, stable by 6–8.

Assembly effort6h — 0.5h/wk
Stable by6 — 8wk
C — Infrastructure tier

Questions about the systems underneath.

Knowledge bases, retrieval, operational software — the layer everything else depends on.

“Can answers cite our docs instead of guessing?”

Yes, for documented behaviour. Undocumented edge cases return “no grounding found” rather than an improvisation — refusal is a feature. Recall against your corpus needs an eval pass first.

Cited answers88 — 94%
Eval setup1 — 2wk

“Can you replace our spreadsheet-and-prayers CRM?”

Yes, and in that order: we map what the spreadsheet actually does — including the tabs only one person understands — then build the app around it. Data migration is part of the build; parallel-run typically lasts 2–4 weeks.

Parallel run2 — 4wk
Migration errors<1%

“What happens when the model gets it wrong?”

It will, and the system is designed for that day: every action traces to its inputs, evals run continuously, and bounded permissions cap the blast radius. You see the failure in a log, not in a customer's inbox.

Actions traced100%
Blast radiusScopedby role
D — Questions we say no to

Three requests, three refusals.

  • “Build us a chatbot that sounds human.” — If the goal is to pass as human, that's a deception product. No.
  • “Automate the compliance sign-off.” — The signature carries legal weight. We automate everything around it; the signature stays yours.
  • “Just do a small pilot for the board.” — A pilot with no ship criteria is theatre. Bring a metric or bring patience.