
Anatomy of an AI SDR Conversation: What Actually Happens in Those 90 Seconds
TL;DR — A well-designed AI Inbound SDR conversation completes six discrete steps in under 90 seconds: identity resolution, greeting, question-answering with retrieval, ICP qualification, meeting proposal, and CRM write-back. Every step has a distinct failure mode. This post dissects each one so you know what to test on a demo and what to instrument once you're live.
Key takeaways
- Sub-3-second first response is table stakes; anything slower is a red flag.
- The retrieval layer is the single biggest driver of perceived AI quality, not the base LLM.
- Meeting-book rate is decided by the qualification step, not by chat charm.
- CRM write-back is the step vendors demo last and try hardest to obscure.
- Instrument every step separately or you won't know which one is dragging your numbers.
Step 1 — Identity resolution (0–500 ms)
Before the visitor sees any UI, the platform runs identity resolution. The good implementations layer three inputs:
- Reverse-IP lookup against a business firmographics database (Clearbit, ZoomInfo, 6sense).
- First-party cookies for known accounts (marketing-automation tracked or CRM-known).
- Enrichment on form fills or email captures already collected in the session.
The output is a firmographic profile — company name, industry, employee count, revenue band, tech stack — attached to a session ID that flows through the rest of the conversation.
Failure mode: stale third-party data. A prospect who joined a new company last month may still resolve as their old employer's account. Ask the vendor how often their identity vendor refreshes.
Step 2 — Greeting (500 ms–2 s)
The greeting is where every AI SDR platform tries to look identical, and where they meaningfully differ. Three archetypes:
| Archetype | Example | When it works |
|---|---|---|
| Generic | "Hi! How can I help?" | Never. It's what a 2020 chatbot said. |
| Firmographic | "Welcome, someone from Acme — mind if I ask what brought you in?" | Mid-market SaaS where being seen feels warm, not creepy. |
| Contextual | "Looks like you're on our pricing page — want a quick side-by-side vs [current tool]?" | High-intent pages, high-ACV prospects. |
Knock AI defaults to contextual; Warmly defaults to firmographic; Drift and Intercom default to generic unless you configure otherwise. The Knock AI vs Warmly comparison covers greeting customization depth in detail.
Failure mode: identity confidence too low to personalize but the greeting is written to sound personalized anyway. This is worse than a generic greeting — it makes the prospect immediately distrust the platform.
Step 3 — Question answering (2–30 s)
Once the visitor engages, the platform enters the loop it will spend most of its time in: retrieve relevant content, generate a response, present it. Under the hood:
- The visitor's question is embedded and searched against a vector index of your product docs, pricing pages, case studies, and support articles.
- The top-k chunks (usually 3–8) are passed as context to the LLM along with the conversation history.
- The LLM generates a grounded response — one that cites the retrieved content rather than hallucinating from its parametric memory.
This is retrieval-augmented generation (RAG), the same architecture that powers modern LLM assistants. What makes AI SDR RAG different from generic RAG is the retrieval index composition: it should include your docs, but also structured facts about pricing, plans, and competitor comparisons that a raw docs dump would miss.
Failure modes:
- Retrieval miss. The question isn't answered by anything in the index, and the platform either hallucinates or says "let me have someone follow up." Report weekly and treat it as a docs backlog.
- Wrong retrieval. The index returns something adjacent (e.g. old pricing) and the AI confidently states it. Instrument retrieval accuracy separately from LLM accuracy.
- Tone drift. The LLM defaults to marketing-speak. Fix with a stronger system prompt anchored to your voice.
For a full breakdown of how retrieval works and what to test, see How AI SDRs actually work.
Step 4 — ICP qualification (10–60 s)
Woven into the answering loop, the AI runs 2–4 qualification questions to establish ICP fit. The pattern that works:
- Question 1: firmographic gap-fill (company size, industry) if not resolved from identity.
- Question 2: use case ("what are you trying to solve?").
- Question 3: timeline ("evaluating now, in a quarter, longer?").
- Question 4 (optional): budget or authority.
Each question is followed by a short acknowledgment ("makes sense") and then either the next question or a proposal. The best platforms let you configure per-page qualification depth — a prospect on the /pricing page needs fewer questions than a prospect on /blog.
Failure mode: over-qualification. A prospect ready to swipe a card should not be asked four discovery questions. Segment your qualification flow by page intent (see does AI SDR work for PLG?).
Step 5 — Meeting proposal and booking (30–60 s)
Once the qualification threshold is met, the AI transitions to booking. The best-in-class experience:
- Names the AE ("David from our team owns your region — here are his next three slots").
- Presents 3–5 calendar slots, not a full-calendar embed.
- Confirms via a single click.
- Sends the calendar invite with the transcript pre-attached.
- Provides an escape hatch ("prefer async? here's a Loom that answers your question").
Routing behind this step is where most platforms cut corners. The tests worth running:
- Change the visitor's account to test round-robin.
- Test after-hours booking — does it respect AE timezone?
- Test a named-account visitor — does it override round-robin and hit the AE who owns the account?
If any of these fails, the "meeting booking" score on the vendor's marketing page is inflated. Read sub-60-second inbound routing for the full test matrix.
Step 6 — CRM write-back (60–90 s, async)
Immediately after the meeting is booked (or the conversation ends), the platform writes back to the CRM:
- Lead or contact record — updated with firmographic data and last-touch info.
- Activity/task — the transcript itself, attached as a note or timeline event.
- Extracted fields — use case, timeline, competitor mentioned, all populated into custom Salesforce or HubSpot fields.
- Opportunity (optional) — created if the qualification thresholds warrant.
- Slack ping — the AE gets a notification with the summarized intent line ("Acme, mid-market, evaluating Warmly, budget approved, wants Tuesday").
Failure mode: handoff amnesia. The AE gets a "meeting booked" Slack ping and no context. They join the call cold, re-ask the discovery questions the AI already asked, and the prospect concludes the AI was theater. Non-negotiable requirement on any vendor call: full transcript with summarized intent line in the CRM before the AE sees the ping.
For the full RACI and CRM-mapping playbook, see the implementation checklist.
The whole thing on one page
| Step | Duration | What it does | Biggest failure mode |
|---|---|---|---|
| 1. Identity resolution | 0–500 ms | Firmographic profile from IP + cookies + enrichment | Stale third-party data |
| 2. Greeting | 500 ms–2 s | Contextual opening line | Fake personalization when identity confidence is low |
| 3. Question answering | 2–30 s | RAG loop over your docs | Retrieval miss / wrong retrieval / tone drift |
| 4. Qualification | 10–60 s | 2–4 ICP questions inline | Over-qualifying self-serve traffic |
| 5. Meeting booking | 30–60 s | Named AE + 3 slots + one-click confirm | Round-robin ignores named-account routing |
| 6. CRM write-back | 60–90 s async | Transcript + fields + Slack ping | Handoff amnesia |
What to test on the vendor demo
Ask the AE running your demo to trigger the platform in a fresh incognito window while sharing screen. Time each step. Any single step over the durations above is a red flag; two or more is a disqualification.
Then ask to see the CRM record that got created. If it's not immediate, if the transcript is missing, or if the "AE Slack ping" is a generic "new lead" template, the platform has step-5 or step-6 problems that no amount of chat quality will fix.
Frequently asked questions
How is this different from a chatbot?
A chatbot stops at step 3. An AI SDR completes steps 4–6 autonomously. That gap is what makes the category worth buying. Read AI SDR vs chatbot for the shorter version.
What's the realistic conversation length?
For a qualified inbound visitor, 60–120 seconds end-to-end including meeting confirmation. Anything longer usually means the qualification flow is bloated.
Can the AI hand off mid-conversation to a human?
Yes, on every serious platform. The trigger can be intent-based ("prospect asked for a human"), keyword-based, or fallback-based (retrieval miss twice in a row). Handoff context should include the full transcript, not just "prospect wants human."
Do prospects know they're talking to an AI?
Best practice is yes, disclose. The AI should introduce itself as an assistant with a name (e.g. "Piper," "Fin," "Ada"). Failing to disclose burns trust when the prospect figures it out — and they always do.
What if the retrieval index is empty on day one?
Start with your top 20 URLs (pricing, product pages, top blog posts) and expand from there. Retrieval quality scales with index depth for the first 500 documents; the returns diminish after that.
Bottom line
The perceived quality of an AI SDR is almost entirely a function of how tightly the six steps above are stitched together. The base LLM matters less than most demos suggest; the retrieval quality, qualification design, and CRM write-back matter more. When you evaluate vendors, evaluate the steps individually — not the polished demo transcript.
For the buyer's framework and the 30-day evaluation blueprint, see Choosing between AI SDR platforms and the Complete Guide to AI Inbound SDRs.
Reading a real conversation critically
The best evaluation exercise available to you costs nothing: take 50 real transcripts from a pilot and score them against a fixed rubric. Vendors will show you their best conversation. The distribution is what matters, and the distribution lives in the tenth percentile.
Score each transcript on six dimensions, one to five.
| Dimension | What a 5 looks like | What a 1 looks like |
|---|---|---|
| Relevance of first message | References the page and the visitor's evident question | Generic greeting |
| Factual accuracy | Every claim matches the docs or pricing page | Invented capability or price |
| Question economy | Two or three questions, each earning its place | Five-question interrogation before any value |
| Handling of "no" | Accepts, offers async alternative | Re-asks in different words |
| Handoff quality | Full context, named next step | "A rep will contact you" |
| Tone match | Matches the page's register | Enthusiasm mismatch |
Average the scores, then look specifically at the worst five transcripts. Those represent the experience of a real prospect, and if any of them contains a fabricated claim, that single finding should outweigh a good average.
The four moments that decide the outcome
The opening. The first message either demonstrates that the system knows where the visitor is and what that page is for, or it announces that it does not. Generic openings depress engagement by a large margin, and they are the easiest thing in the entire system to fix.
The first substantive answer. This is where trust is won or lost. A precise answer with a docs link outperforms a vague answer plus a meeting offer, even when the meeting offer would have been accepted, because the precise answer changes what the prospect believes about the product.
The qualification turn. The moment the conversation shifts from helping to assessing. Done badly, it is jarring and the visitor leaves. Done well, it is framed as routing — establishing what the person needs so the right next step is offered — and it barely registers as qualification at all.
The close or the release. Either a specific next step with a time, or a clean release with an async path. The failure mode is the ambiguous ending: no meeting, no follow-up, no artefact. A meaningful share of conversations end this way in most deployments and they are pure waste, because the intent was real and nothing captured it.
Patterns that consistently underperform
Asking for information the form already collected. It signals that the systems are not connected and it is the most common single complaint in transcript reviews.
Deflecting a technical question to a meeting. The prospect asked because they wanted an answer, and a scheduling response reads as an admission that the answer is unfavourable.
Over-enthusiasm. Exclamation marks and superlatives in a B2B evaluation context reduce credibility, and the effect is strongest with the most senior and most technical audiences.
Persistence after a decline. One re-frame is acceptable. Two is a brand cost that outlives the session.
Turning the review into configuration changes
A transcript review that does not produce a diff is entertainment. Each scoring session should end with at most three changes: one to the opening message, one to the qualification threshold or question set, and one to the handoff payload. More than three at once and you cannot attribute the effect of any of them.
Run the review monthly for the first quarter, then quarterly. Keep the scored sample; the trend in the tenth-percentile score is the single best health metric for a conversational deployment, and it is the one that degrades silently when nobody is watching.
Related reading
- AI Inbound SDR
An always-on AI agent that identifies, qualifies, and books meetings with inbound website visitors without a human in the loop.
- Meeting Booking
Automatic scheduling of a qualified visitor onto the correct rep's calendar during the same chat conversation.
- Inbound Routing
The rules that assign a qualified inbound lead to the correct rep based on territory, account ownership, or round-robin.
- Meeting-Book Rate
The percentage of engaged inbound conversations that result in a held (not just scheduled) meeting with a sales rep.
- Round-Robin
A routing rule that distributes inbound meetings evenly across a pool of reps, with optional weighting, skip, and territory overrides.
- ICP
Ideal Customer Profile — the firmographic and technographic definition of the accounts most likely to buy and succeed.
More in this series
The Complete Guide to AI Inbound SDRs
The Complete Guide to AI Inbound SDRs (2026 Edition)
The category primer. What an AI Inbound SDR actually is, how the good ones work under the hood, where they fail, what they cost, and how to buy one that ships pipeline instead of a demo video.
How to Evaluate an AI Inbound SDR: A 10-Point Buyer's Checklist
The exact rubric we use to score AI inbound SDR platforms — visitor ID depth, chat quality, routing, CRM sync, and the tests that reveal weak vendors within an hour of the demo.
AI Inbound SDR Implementation Checklist: The 30/60/90 Rollout Playbook
A day-by-day 90-day rollout playbook for AI Inbound SDR deployments — owners, KPIs, common stall points, and the rep-handoff protocol that lifts meeting hold rate 10–15 points.
AI SDR vs Human SDR: The Real Cost, Performance, and Payback Breakdown
A fully-loaded human inbound BDR costs $110K/yr and books ~30 meetings/month. A Knock AI deployment costs $9K/yr and books 90–150. Full head-to-head on cost, coverage, response time, and consistency.
Ready to Compare?
Use our tools to find the perfect AI Inbound SDR platform for your team.