Most website visitors leave without contact, and inbound leads often wait hours for a reply while the interest cools. An AI sales assistant engages visitors immediately, answers product questions from approved content, asks a few qualifying questions, books a meeting and passes the lead with a summary to the right rep. Typical benefit: faster speed-to-lead, 24/7 coverage and less time reps spend on unfit leads, as long as it is disclosed as AI and hands over cleanly.
Problem & value
- Today: web forms, static chatbots with decision trees, and SDRs who triage inbound leads by hand. Replies come late, qualification is inconsistent, out-of-hours visitors are lost, and reps re-ask questions the prospect already answered.
- AI value: free-form dialogue instead of rigid forms, instant response in many languages, consistent application of qualification criteria, structured data written to the CRM, and meeting booking inside the conversation.
- Evidence (vendor-reported, treat with care): HubSpot markets its Customer Agent as able to qualify prospects, book meetings or transfer to a rep [2]; Salesforce's Agibank story describes an agent that passes the full conversation history and an interest assessment to the rep so no re-qualification is needed [1]; ElevenLabs reports an internal AI SDR whose qualification decisions need no human intervention in 78% of cases [7]. No independent, cross-vendor benchmark exists; measure on your own funnel.
Variants
- B2B SaaS / services: qualification (company size, use case, timeline, budget, decision role), demo booking, routing by territory.
- Manufacturing / industrial: product-finder and configuration questions, RFQ intake, distributor routing.
- Real estate, automotive, education: viewing, test-drive or enrolment enquiries, often via WhatsApp or voice.
- Financial services and insurance: lead capture and appointment booking only; advice stays with licensed humans (sector rules apply).
- Voice variant: inbound call qualification, see CONV-002; outbound calls are a different legal situation, see CONV-008.
- Private use: not applicable; consumer-side shopping assistance is CONV-006.
How it works
- visitor message (web widget / WhatsApp / email / voice)
- AI disclosure + consent/privacy notice
- intent routing (product question | pricing | support | sales | spam)
- grounded answer (RAG over product docs, pricing, case studies)
- qualification dialogue (slot filling against ICP/BANT schema)
- scoring + routing rules (territory, segment, language, account owner)
- action: book meeting via calendar API | create/update CRM lead | handoff live to rep
- summary + transcript to CRM; follow-up email draft; consent logged
- Dialogue and grounding: an LLM with retrieval over approved content; prices and commitments come from tools or fixed text, not from model memory.
- Qualification: slot filling into a schema (company, role, need, timeline, budget); an LLM or classifier proposes a score, deterministic rules decide routing. Existing CRM data (account owner, open deals) is looked up through tool calls.
- Human-in-the-loop: live handoff on request, on high-value accounts, on negative sentiment or off-script questions; reps own quotes, discounts and contracts. Escalation to a person must always be possible.
Data preparation
Typical sources are product sheets, price lists, case studies and FAQ pages (website, CMS, PDFs on file shares, sometimes only in sales reps' heads), plus CRM records and qualification criteria (ICP, BANT) that are often undocumented. Typical gaps: outdated or contradictory price and product statements, PDFs as scans, duplicate CRM contacts, missing lead-stage definitions and no consent records.
Preparation steps: (1) pick one authoritative source per topic and retire old versions; (2) convert PDFs/scans to clean text (OCR where needed); (3) write down the qualification questions, scoring rules and handover criteria in plain language; (4) deduplicate and tag CRM data; (5) collect 30-50 real prospect questions with approved answers as test set. Tools: the vendor's own crawler for Path A, Paperless-ngx or a similar OCR tool for scans, a spreadsheet for the question set. Rough share of effort: about 30-40 percent, mostly curation rather than technology.
Useful PREP use cases: PREP-013 (authoritative knowledge base), PREP-016 (gold-standard examples), PREP-018 (test sets), PREP-025 (purpose and legal basis for reusing CRM data). See also GUIDE-003.
Implementation paths
Which path for whom
- SMB — A: mature EU-hostable chatbot/lead tools exist at SMB prices and the qualification process is standard.
- Enterprise — B: the assistant must plug into the own CRM, routing rules and product data, and it is a core sales channel where volume makes seat prices high.
- Developer — B: cloud LLM APIs plus RAG and a CRM connector give full control over qualification logic without operating models.
Path A — Buy (recommended for most SMBs)
- CRM-native agents: HubSpot Breeze Customer Agent and Prospecting Agent (the latter in beta per HubSpot's pages) [2], Salesforce Agentforce with SDR/Sales agents [1], Microsoft Dynamics 365 Sales/Copilot Studio, Zoho Zia, Pipedrive.
- Conversational marketing platforms: Drift (Salesloft), Intercom Fin, Qualified, Conversica, Ada, Landbot (Spain, EU), Userlike (Germany, EU), Tidio (Poland, EU), Botpress, Voiceflow, Cognigy (Germany, EU; part of NICE), Moveworks-style enterprise assistants for internal use.
- Voice: ElevenLabs Agents, PolyAI, Retell, Vapi; EU hosting and data residency must be checked per vendor.
- Choose A when: you already run a supported CRM, volumes are moderate, and the DPA, EU data residency and sub-processor list are acceptable.
Path B — Build on cloud APIs
- Models: Claude, GPT, Gemini or Mistral models (Mistral is a French provider) via EU-region endpoints (Azure OpenAI EU, Google Vertex EU, AWS Bedrock Frankfurt/Paris, Mistral La Plateforme); tool calling and structured outputs for slot filling.
- Orchestration: LangGraph, Pydantic AI or the vendor agent SDKs; n8n/Make for CRM and calendar integration; Model Context Protocol servers for HubSpot, Salesforce or Google Calendar.
- Retrieval: pgvector, Qdrant (Germany) or Azure AI Search with a multilingual embedding model.
- Guardrails: prompt-injection and PII filters, output checks for price or contractual claims.
- Choose B when: you need custom qualification logic, several channels or tight integration.
Path C — Self-host (open source)
- Models: Mistral Small/Medium open weights, Qwen3, Llama or Gemma family, served via vLLM or Ollama; embeddings such as bge-m3 or multilingual-e5; Qdrant or pgvector.
- Bot frameworks: Rasa, Botpress (open core), Chatwoot as inbox, n8n for workflows.
- Hardware: HW2 handles a mid-sized chat load with a ~30B quantised model; add voice (Whisper-class ASR, open TTS) and HW3 is more comfortable.
- Choose C when: strict data residency, or the chat contains sensitive business information.
Path D — Custom training
Rarely needed. Fine-tuning a small model on past transcripts (tone, qualification outcomes) can cut cost at high volume; a lead-scoring model on CRM history is a separate tabular task (PRED-005). Needs hundreds to thousands of labelled conversations with won/lost outcomes.
State of the art (2026)
- Agentic SDRs: vendors now ship agents that qualify, score, route and book in one flow, configured by instructions plus CRM actions rather than hand-built dialogue trees [1][2].
- Voice and chat parity: realtime speech models make the same qualification logic available by phone (CONV-002).
- Limits and failure modes: hallucinated prices, features or promises; qualification that favours leads fitting the scoring model while mishandling atypical ones (e.g. a senior buyer on a private email address, a critique raised about Agentforce's SDR agent [8]); prompt injection from visitors; over-eager data collection; irritation when the bot blocks access to a human; and inconsistent behaviour across languages. Evaluate with replayed real conversations and adversarial test sets before launch.
Shared resources & combinations
- Same stack as CONV-001 (support chatbot): knowledge base, RAG, widget, handoff; one bot often does both and routes by intent. Also CONV-026 (in-product copilot) and CONV-006 (shopping assistant).
- Downstream: PRED-005 lead scoring, AGENT-011 CRM enrichment and hygiene, AGENT-007 outbound prospecting agents (a different legal basis, see below).
- Infrastructure: a shared LLM gateway, vector index and guardrail layer (HW2/HW3) can serve all conversational use cases.
Effort, cost & KPIs
- Path A: days to a few weeks; priced per seat, per conversation or per resolved/qualified lead (HubSpot lists usage-based prices [2]; verify current pricing).
- Path B: 1–3 months including CRM integration, evaluation and consent flows.
- Path C: 2–4 months plus operations; GPU server and monitoring.
- KPIs: speed-to-lead, visitor-to-lead and lead-to-meeting conversion, MQL-to-SQL rate, share of qualified leads accepted by sales, handoff rate, answer accuracy on audited samples, customer satisfaction, complaints, cost per qualified lead.
People & change
Start with management and the sales lead (process owner): they define goals, qualification rules and what the bot may promise. Involve IT early for website, CRM and messaging integration and the DPO before launch for the privacy notice, consent, AVV with the vendor and, for profiling or scoring at scale, a DPIA check. Let the sales reps (end users) co-write the qualification questions and the handover format; they must trust the lead summaries.
Works council: depends. The bot mainly processes prospect data, but if conversation logs or lead statistics are used to assess individual reps, co-determination under §87(1) no. 6 BetrVG applies; §90(1) no. 3 requires timely information anyway when sales processes change with AI. Only relevant in companies with an elected works council.
AI literacy: support staff in understanding limits and hallucinations (AI Act Art. 4, measures to support literacy) and train reps on reviewing handovers. Communicate that the bot takes over first contact, not the sales conversation. Run a pilot on one product line or one channel for 4-6 weeks, with an easy human takeover and weekly review of transcripts. Typical concern: fear of job loss or wrong promises; address with clear scope, disclosure that users talk to AI (Art. 50) and human control over quotes.
Guides: GUIDE-008 (change management), GUIDE-005 (data protection), GUIDE-009 (pilot).
Risks, regulation & compliance
- EU AI Act: limited risk. Art. 50(1) requires that people are informed they interact with an AI system unless obvious from context; Art. 50 applies since 2 August 2026, and the Commission's non-binding Art. 50 guidelines were published in July 2026 and read the "obvious" exception narrowly, so state it at first contact [3][4]. Not Annex III high-risk in the usual case. It can become problematic if the bot infers emotions or vulnerabilities to manipulate behaviour (Art. 5(1)(a),(b) prohibited manipulation and exploitation) or if scoring leaves sales and feeds decisions on credit, insurance pricing or employment (high-risk). Art. 4 AI literacy applies.
- GDPR: names, emails, phone numbers and chat content are personal data. Define the legal basis (consent or legitimate interest, Art. 6; EDPB Opinion 28/2024 stresses a case-by-case balancing test for legitimate interest in AI contexts [5]), give Art. 13 information, honour access, deletion and objection rights, set retention periods, sign DPAs and check transfers. Profiling with legal or similarly significant effect (Art. 22) should not occur; keep qualification as a sales-priority aid.
- ePrivacy / marketing law: chat widgets with tracking need cookie consent (ePrivacy); follow-up marketing emails need consent or the narrow existing-customer exception, and in Germany UWG §7 double opt-in practice applies; cold outreach rules differ by country and apply to AI-driven outreach too.
- Consumer law: no binding quotes or false claims; pricing from authoritative sources; clear identity of the seller.
- Mitigations: disclosure banner, consent checkpoint before capturing contact data, data minimisation (only fields needed), red-team tests, logs of conversations and decisions, easy human escalation, transcripts excluded from model training by contract.
Getting started
- Define the ICP and 4–6 qualification questions with sales, plus the rules for routing and escalation.
- Collect approved content (product pages, pricing logic, FAQs) and decide what the bot must never say.
- Pilot with one channel (website chat) and one tool from Path A or a thin Path B build; wire in calendar and CRM.
- Add the AI disclosure, privacy notice and consent logging before going live.
- Review 50 transcripts weekly for the first month, compare conversion against form leads, then expand.
Sources
- Salesforce — Agibank agentic lead qualification customer story — https://www.salesforce.com/customer-stories/agibank/agentic-lead-qualification/
- HubSpot — Breeze Customer Agent — https://www.hubspot.com/products/chatbots
- Regulation (EU) 2024/1689 (AI Act), Art. 50 and Art. 5 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- McCann FitzGerald — Commission guidelines on Article 50, provider obligations — https://www.mccannfitzgerald.com/knowledge/data-privacy-and-cyber-risk/ai-transparency-european-commissions-guidelines-on-article-50-part-1-provider-obligations
- CMS — EDPB Opinion 28/2024 key takeaways — https://cms.law/en/int/legal-updates/edpb-opinion-28-2024-key-takeaways-on-processing-personal-data-in-the-context-of-ai-models
- HubSpot — Breeze Prospecting Agent (beta) — https://www.hubspot.com/products/sales/ai
- ElevenLabs — How we scaled inbound sales with an AI SDR — https://elevenlabs.io/blog/how-we-scaled-inbound-sales
- Mota AI — Is Agentforce's SDR Agent qualifying the right leads (opinion) — https://motaaidev.substack.com/p/is-agentforces-sdr-agent-qualifying