AI × Sales
How to test lead scoring, rep assistance, forecasts, outbound, pricing, and conversational qualification while keeping people, consent, and consequential decisions under accountable control.
Lead Scoring & Qualification
A score is a fallible prioritisation signal, not a fact about a person or permission to contact them.
Opaque prioritisation
Definitions differ by rep → missing CRM fields are mistaken for low intent → protected or proxy traits can influence treatment → people may be contacted without current eligibility.
Decision support with review
Define the decision and lawful inputs → exclude sensitive/proxy features → test by relevant groups and missingness → show reasons and uncertainty → let a trained person override → check consent and suppression before any contact.
Baseline worksheet: record current lead volume and observed win rate. It does not predict who will buy or an improvement.
Evidence to collect
Sales Copilots
Assist with notes and drafts, but preserve recording consent, source context, human judgment, and CRM accountability.
Fragmented records
Notes vary by rep → recording notice and retention are inconsistent → summaries omit uncertainty → CRM writes are hard to attribute.
Consented draft assistance
Provide required recording notice and consent → restrict retention/access → produce source-linked draft notes → show uncertainty → rep corrects and approves → CRM write uses least privilege, exact-field approval, idempotency, and an audit log.
Baseline worksheet: total the team's current reported admin time. It does not assume that time disappears or becomes selling time.
Pilot evidence to collect
Sales Forecasting
A forecast is a dated estimate with uncertainty, not a single “real number.”
Uncalibrated point estimates
Stage definitions drift → missing activity is treated as negative evidence → overrides are undocumented → finance receives a point estimate without assumptions or uncertainty.
Versioned scenarios
Define outcomes and cut-off date → validate data completeness → backtest out of time → report calibrated ranges and scenario assumptions → capture human overrides and reasons → monitor drift and decision impact.
Baseline worksheet: expose current pipeline and management's stated accuracy assumption. It does not translate an accuracy percentage into revenue.
Pilot evidence to collect
AI SDRs & Outbound
Use assistance to research and draft only for eligible, relevant outreach; do not automate unsolicited bulk contact.
Uncontrolled prospecting
Purchased or scraped data has unclear provenance → relevance is assumed → suppression is stale → volume harms people, brand, and sender reputation.
Eligible, reviewed outreach
Document lawful purpose and source → minimise data → check consent, suppression, region, and channel rules → draft a truthful accessible message → human approves exact audience/content/volume/window → cap frequency → handle opt-out and escalation immediately.
Baseline worksheet: calculate today's reported SDR cost, meeting count, and simple cost per meeting. It does not estimate an AI replacement.
Pilot evidence to collect
Pricing & Negotiations
Use analysis to draft a recommendation; an authorised person owns price, fairness, margin, and customer impact.
Inconsistent discretion
Discount reasons are missing → margin and contractual constraints are checked late → similar customers can receive unexplained treatment.
Bounded recommendation
Use approved non-sensitive features → show comparables and uncertainty → test fairness and contractual constraints deterministically → create a PAUSED quote draft → authorised owner approves exact price, terms, customer, and expiry.
Baseline worksheet: expose the gross value of current discounts. It is not margin leakage or recoverable revenue.
Pilot evidence to collect
Conversational AI (WhatsApp & Chat)
Offer a clearly disclosed, accessible assistant on an opted-in channel and make human help easy to reach.
Inconsistent intake
Notice, consent, sensitive questions, accessibility, retention, and handoff quality vary by conversation.
Disclosed, limited assistance
Disclose automation → confirm channel eligibility → collect only necessary fields → avoid sensitive inference → show accessible choices → escalate uncertainty, distress, complaints, regulated topics, or human requests → require approval before booking or CRM writes.
Baseline worksheet: total current inquiry volume and reported handling time. It does not estimate automation, time saving, or meetings.
Pilot evidence to collect
Reference pattern: a controlled sales workflow.
Start read-only and test on representative and adversarial cases. Separate tenants and use least-privilege service identities. Models may produce explanations and drafts; they cannot grant consent or approve actions. Every contact, CRM update, meeting, quote, or price write stays PAUSED until deterministic policy validates eligibility, suppression, caps, and an unexpired signed approval bound to the exact recipient/resource, content, values, and time window. Use idempotency, complete audit logs, human escalation, stop controls, and tested rollback.
Vendor discovery links — verify before use
These links identify products mentioned in the earlier lesson. They do not prove an outcome. Locate precise, dated official documentation, terms, security/privacy material, and case methodology before relying on a feature or claim.
- 6sense Revenue AI — 6sense.com
- Demandbase One — demandbase.com
- Madkudu — madkudu.com
- Gong — gong.io
- Dooly.ai — dooly.ai
- Clari — clari.com
- BoostUp.ai — boostup.ai
- Aviso — aviso.com
- 11x.ai — 11x.ai
- Artisan — artisan.co
- Clay — clay.com
- PROS Holdings — pros.com
- Zilliant — zilliant.com
- DealHub — dealhub.io
- Salesloft — salesloft.com
- Haptik — haptik.ai
- Air.ai — air.ai