60 seconds to first alert: the onboarding research that rejected everything obvious

Most B2B SaaS onboarding ends the same way: a blank dashboard. The user followed all the steps, confirmed they understood, clicked "Done" — now they're looking at empty charts wondering if the product works. That's churn. They just haven't clicked cancel yet.

Canarist monitors supply chain risks — tariffs, port closures, supplier failures — for SMBs without a dedicated procurement team. These users arrive after a shock event. They have 5 minutes and zero patience for a 12-field form. I designed the onboarding to prevent that blank dashboard. No user interviews — research first: 8 competitive platforms analyzed, 40+ SaaS onboarding teardowns, behavioral framework research, and one empirical LLM test that killed the most obvious approach. Solo, in 3 days, shipping features in parallel.

The problem

B2B onboarding splits at two extremes. Enterprise tools (Everstream, Resilinc) use CSM-guided setup — weeks of implementation, zero self-serve. The "simple" tools use form wizards: 4 steps, 12 fields, a blank dashboard at the end.

The target: TTV ≤ 60 seconds. First personalized alert within one minute of registration. Everything else is a distraction.

Two approaches we rejected

Approach 1: AI-autocomplete form (rejected on security + trust)

The first idea: use a work email to enrich company data via Apollo/Clearbit, pre-fill the form, ask the user to confirm. Fewer keystrokes, same data.

Rejected on two grounds. Security: Canarist's users handle confidential procurement documents — asking them to connect data enrichment APIs at signup creates a trust barrier that kills conversion before it starts. Trust: Auto-filled data can be silently wrong. A user who confirms incorrect supplier data and gets irrelevant alerts will blame the product, not the enrichment API.

Approach 2: AI-driven step-by-step extraction (rejected on error cascade)

The second idea: run an LLM at each onboarding step to extract structured data from free-text input. Step 1: extract industry. Step 2: extract suppliers. Step 3: validate. Sequential AI decisions, guided UX.

Rejected because chaining AI decisions compounds errors. A misclassification at step 1 contaminates every downstream extraction. The system optimizes for flow completion, not data quality.

The test that confirmed this: I ran extraction quality tests on gpt-4o-mini with 6 inputs — two complete company descriptions, two partial, two deliberately vague. Per-step extraction on partial context produced consistent misclassifications and hallucinated supplier names. Running the same extraction on the full transcript: all classifications correct, zero hallucinations. LLM extraction is reliable only when run once on the full transcript — not per-step. Confidence scores drop and hallucination risk rises when the model operates on partial context.

What the research showed

Two parallel research tracks informed the final decision.

Competitive analysis (competitor-onboarding.md) of 8 platforms — Everstream, Resilinc, Prewave, Craft.co, and others — surfaced one pattern worth keeping. Everstream's onboarding opens with a risk lens selector: you choose logistics, financial, or geopolitical focus. One choice, and every downstream view — dashboards, alerts, reports — adapts to that lens. The structural principle: a single axis of personalization that cascades downstream, instead of twelve independent fields. In Canarist, that axis is the supply chain description itself. Once extracted, it drives alert filtering, scenario prioritization, and the simulation shown post-onboarding.

Adjacent SaaS research (adjacent-saas-patterns.md) across HubSpot, Pipedrive, Monday, Semrush, and Vanta revealed a consistent pattern: suggest-and-confirm beats open-ended input. Every best-in-class onboarding shows the user 3–5 options and asks "which apply?" instead of asking them to type from scratch.

The synthesis (onboarding-methodology.md) reduced the research to one principle: minimize ability requirements, maximize trigger resonance (Fogg Behavior Model). The user arrives after a shock event — a tariff surprise, a port closure. Motivation is already high. The job is not to persuade — it's to not get in the way.

The decision

Replace the form wizard with a 3-turn conversational AI agent.

One LLM call on the full conversation transcript extracts the structured JSON. No per-step inference. No enrichment API at signup. No 12-field form.

The blank dashboard problem is solved differently — not by faster setup, but by simulation: immediately after onboarding, the user sees a hyperrealistic crisis scenario tailored to their suppliers, demonstrating what Canarist would have warned them about 72 hours before it happened.

Trade-offs of the chosen approach

The conversational model has its own failure modes worth naming.

A user who types "I import stuff from China" gives the LLM too little context to build a useful profile. Turn 2 handles targeted clarification — but only on completeness gaps, not quality. Incomplete is recoverable. Wrong isn't. The system is designed to prefer a sparse-but-accurate profile over a complete-but-hallucinated one.

A second risk: users who abandon mid-conversation. The 3-turn format is intentionally short, but "tell me about your supply chain" is a more open-ended ask than "select your industry." The phrasing of the opening message was iterated specifically to minimize this friction — the goal is to make Turn 1 feel like a conversation opener, not a form field in disguise.

The numbers

• 3 days — research to decision to spec, while shipping features in parallel
• 12 → 3 inputs: the reduction from first draft to shipped design
• 1 LLM call instead of per-step inference — eliminates error cascade, reduces cost
• ≤ 60 seconds TTV target from registration to first personalized alert

Canarist is currently in closed alpha. Quantitative retention data is pending — the architecture is in production, measuring starts next.

What this means

Research doesn't just validate decisions. It kills the wrong ones before they cost anything.

Two approaches were designed, prototyped in spec, and discarded based on evidence — not stakeholder preference, not gut feel. The conversational model wasn't the obvious choice. It was the choice left standing after the others failed their tests.

That's the constraint I work within: every design decision needs a reason that survives a question. If you're building something where onboarding is the first impression — that conversation is worth having early.