💼 Sales & Revenue
AI across the revenue cycle from lead identification to closed deal. The most consistently mature domain: three-quarters of practices are good practice, including lead scoring, pipeline forecasting, and conversation intelligence. CRM copilots are mainstream. The few leading-edge practices involve autonomous prospecting and deal-coaching agents. Momentum is moderate — most gains are incremental rather than transformative.
The Headline
The vendor that told you to stop hiring salespeople has withdrawn the claim. The setup that works is AI plus a person, and it only works on clean customer data.
The Picture
Nearly every sales organization now uses AI. About one in five can show a measurable result. This fortnight the big surveys confirmed the shape from the top down: McKinsey finds only 6% of companies are AI high performers, with the share reporting any earnings impact stuck at 37% for a year despite record spending, and Deloitte puts the failure rate for AI agent pilots (software that acts on its own without being prompted) at 89%. Everyone buys from the same short list of vendors, so software is not what separates leaders from the pack. Two things are: the state of the customer records the AI reads, and whether the company redesigned how its sales team works or bolted a tool onto the old process. The window to catch up is open, because the leaders' advantage is organizational rather than technical. It is also the harder kind to copy.
This Fortnight
The loudest "replace your sales team" vendor changed its pitch. Artisan retired its $2 million "Stop Hiring Humans" campaign and now sells its product as a co-worker. The numbers explain the retreat: in head-to-head tests, human sales development reps generated 2.6 times more revenue than AI-only outreach, and nearly half of automated outreach programs hit an email-deliverability wall within 90 days. If anyone in your building is modeling headcount savings from autonomous outreach, revisit the model.
Salesforce moved its CRM inside Claude, and Siemens showed what disciplined deployment looks like. The integration is live with about 7,000 sellers at GitLab, Siemens, and Legora. Siemens now has AI engaging every inbound lead across 18,000 sellers, with a hard rule that agents may quote only technically valid upgrades. The lesson is the constraint, not the scale: deployments that work give the AI a narrow job and a fence around it.
Call transcription now costs less than half a cent a minute. Microsoft launched transcription at $0.004 a minute, undercutting specialist vendors, and a vendor teardown found seven of ten tools sold as "conversation intelligence" are commodity transcription rebranded at $20 to $50 per seat per month. Before your next renewal, ask your vendor to show you what it does beyond transcribing.
78% of executives have acted on AI recommendations they later suspected were wrong. Validity's survey of 500 marketers traced the problem to data quality: only 21% rate their customer records well prepared for AI. If your CRM has no named owner accountable for its accuracy, that is the cheapest fix available to you this quarter.
Pricing algorithms can form a cartel without anyone agreeing to one. A field study of German gas stations found independent pricing algorithms learned to coordinate and lifted margins 38% with no human agreement; Amazon's Project Nessie reportedly generated $1 billion in excess profit the same way. If you run automated pricing, your general counsel should know how it responds to competitors' moves.
Coming Up
Microsoft re-indexes partner incentives to AI adoption on October 1. Its cloud reseller program ties a 19.5% uplift directly to AI adoption and co-selling. If you buy Microsoft through a partner, expect AI pushed harder in every conversation; if you are a partner, the revenue model changes in two weeks.
Financial regulators are writing AI into the examination agenda. FINRA, the US brokerage regulator, added its first standalone AI section to its annual oversight report, warning that autonomous agents need explicit controls against hallucination (when an AI tool confidently makes things up). Any regulated firm using AI to quote, advise, or draft client material should be able to show its controls before the next exam cycle.
The platforms are absorbing your point tools at renewal. HubSpot's fall release ships a self-updating CRM and a prospecting agent watching 40 buying signals; Zoom launched a bundled revenue suite; Salesforce is now inside Claude. Inventory your sales stack before the next renewal and ask which line items your core platform now covers.
What's Hard About This
Vendors publish returns; nobody publishes accuracy. A critical review found no independent accuracy test of deal-risk scoring from any major vendor, and the largest published return figure, 481%, traces to the vendor's own report. You will have to measure results yourself, against a frozen control group, because no one will do it for you.
AI made the first draft free and left the approval queue untouched. Nearly half of sales content teams name approvals, not writing, as their biggest delay, and one 120-person software company lost $1.4 million of quarterly pipeline to an eleven-day approval queue. Speeding up the front of a broken process delivers the mess sooner.
Bought buying signals decay in days and are now nearly free. Engagement signals lose most of their value within 24 to 72 hours, and account-level scoring often cannot tell a rep who to call. The signal you own, how customers actually use your product, remains the most reliable predictor of a renewal or an expansion.
Practices in this Domain (16)
| PRACTICE | TIER | TREND |
|---|---|---|
| Account intelligence — contact mapping & briefing | LEADING EDGE | — Steady |
| Account-based marketing signal identification | GOOD PRACTICE | — Steady |
| Conversation intelligence — transcription & analysis | GOOD PRACTICE | ↘ Slowing |
| CRM data management & enrichment | GOOD PRACTICE | — Steady |
| Deal intelligence — risk assessment & win/loss analysis | GOOD PRACTICE | — Steady |
| Lead scoring & ideal customer profiling | GOOD PRACTICE | — Steady |
| Partner & channel sales support | BLEEDING EDGE | — Steady |
| Price optimisation & dynamic pricing | GOOD PRACTICE | ■ Blocked |
| Prospecting & outreach personalisation | GOOD PRACTICE | — Steady |
| Quote & contract generation | GOOD PRACTICE | — Steady |
| Real-time conversation guidance | GOOD PRACTICE | — Steady |
| Revenue intelligence & expansion signals | GOOD PRACTICE | — Steady |
| Sales content — proposals, battle cards & objection handling | GOOD PRACTICE | — Steady |
| Sales enablement content generation | LEADING EDGE | — Steady |
| Sales forecasting & pipeline analysis | GOOD PRACTICE | — Steady |
| Territory & account planning | GOOD PRACTICE | — Steady |
Read the full technical briefing (2,453 words) →
Where AI Stands in Sales & Revenue
Sales remains the most thoroughly AI-instrumented function in the enterprise, and the past two weeks have made its central paradox harder to argue with rather than easier. A fortnight ago the domain acquired a clean readiness number: Salesloft's survey of 500 US revenue leaders found AI use at effectively 100% and production-ready AI with measurable outcomes at 20.6%, with a UK sample of 406 returning 28.3%. This cycle the macro studies lined up behind it. McKinsey's global survey of 1,719 organisations found only 6% qualify as AI high performers and the share attributing any earnings impact to AI stuck at 37%, unchanged from 2025 despite record spending. Deloitte and Teradata put the agent pilot-to-production failure rate at 89%. Gartner's RevOps benchmark found 70% of generative AI sales pilots never reach production. A Harris Poll for Collibra found 76% of organisations hit critical roadblocks moving from pilot to production in the past year, with 72% naming an unaligned or poor data foundation as the cause. Six independent samples, one shape: near-universal adoption, a fifth to a quarter of it doing verifiable work, and a large, expensive middle where licences are paid and returns are not booked. The sixteen practices tracked here span the full revenue cycle, and thirteen of them are mainstream capabilities with generally available tooling. The technology question is closed; the readiness question is not.
What distinguishes this domain is that its failure mode has been diagnosed with unusual precision and has now migrated from the operations floor to the boardroom. The diagnosis is data. Validity's survey of 500 marketers found that 78% of C-suite executives have acted on AI recommendations they later suspected were wrong because of underlying data quality, that only 21% rate their CRM data "very well prepared" for AI, and that 62% of organisations have lost revenue directly to poor records. Salesloft's respondents named manual CRM updates as their single biggest operational bottleneck, cited by 37.6%. Integrate and Demand Metric's study of 245 organisations found high-growth companies four times more likely to have advanced data governance and three times more likely to report strong lead-acceptance rates, which is the first time this cycle the correlation between governance and growth has been quantified rather than asserted. Two new failure modes sharpened the picture: entity-resolution errors that route updates to the wrong customer with high confidence and no alert, and BCG's warning that business-built AI agents operating without governance fragment customer records rather than enrich them. The platforms have noticed. HubSpot's Fall '26 release ships a self-updating CRM that captures calls, emails and meetings autonomously and scores record completeness, an explicit attempt to engineer the data problem away rather than wait for organisational discipline to arrive. Clay's climb past $100M in annual recurring revenue with more than 17,000 customers, including Anthropic, OpenAI and Google, shows how much money buyers will spend on the enrichment layer alone.
Against that backdrop, the vendor side spent the fortnight consolidating. Anthropic and Salesforce launched Salesforce in Claude with 37 prebuilt sales skills, live with roughly 7,000 sellers at GitLab, Siemens and Legora; Agentforce Coworker passed 100,000 activated users in 35 days; Siemens now engages every inbound lead across 18,000 sellers and 132 countries through agents that may quote only technically valid upgrades. Zoom shipped a Revenue OS bundling buyer intelligence, conversation intelligence and forecasting, with Barracuda Networks reporting 15% more deals. Outreach reported a twelvefold rise in AI credit consumption and 480% AI revenue growth. Microsoft priced transcription at $0.004 a minute. The direction is unambiguous: the CRM and collaboration incumbents are absorbing the point-solution layer, and the standalone signal, enrichment and transcription vendors are being repriced towards zero. Fifteen of sixteen practices are holding position or losing momentum. The one that continues to advance, revenue intelligence and expansion signals, does so because expansion signals are first-party and already inside the product; even there, the sharpest new evidence is a 180-person SaaS company lifting net revenue retention 18 points and win rate 24 points through a redesign of how its revenue team works around agents, not through a new platform.
What's New, 2026-09-05 to 2026-09-19
The replacement thesis closed this fortnight, and it closed on the vendors' own terms. Artisan's founder retired the $2M "Stop Hiring Humans" campaign and repositioned the product as a co-worker, citing structural barriers including legal limits on cold calling, in-person selling and relationship dynamics that a bot cannot navigate. The production data explains why. Of the 41% of enterprise B2B teams that had deployed an AI sales development rep by Q1 2026, only 22% completed a full replacement; in head-to-head testing, human SDRs generated 2.6 times more revenue and achieved 71% meeting show rates against 52% for AI-only outreach. Empiraa's field analysis added the failure mechanism: 86% of teams report positive first-year returns, but 47% hit a domain-reputation wall within 90 days when spam complaints cross the 0.3% threshold that triggers inbox filtering at Gmail, Outlook and Yahoo, and 21% never recover placement. Personalisation depth, not volume, is what survives: Gong Labs' analysis of 30,000 sales emails found activity-based personalisation triples reply rates, and Hunter.io's 31-million-email study found two custom attributes lift replies 56%. On the guidance side, the human-on-the-call configuration kept producing: Amotions documented a verified $160K monthly sales lift at House of Hearing from real-time objection handling, and OpenAI's GPT-Live-1 API brought full-duplex voice with 80% fewer false interruptions in early deployments. Against that, Centific's testing of five leading models across more than 500 production-like sales-operations tasks returned a 12.7% agent success rate, a number that should sit beside every autonomy claim in this domain.
The second theme is consolidation and commoditisation of the input layers. Microsoft's MAI-Transcribe-2 arrived at $0.004 a minute with 97.8% accuracy, undercutting AssemblyAI at $0.0062 and Deepgram at $0.0059, while its Teams meeting-insights API delivers summaries and action items natively without third-party transcription. A vendor teardown found only three of ten platforms marketed as conversation intelligence deliver genuine deal-risk analysis; the other seven rebrand commodity transcription at $20 to $50 a seat per month. In intent data, G2 began routing behavioural signals through Bombora and Intentsify embedded persona-level scoring into Clay workflows, extending the modular-stack trend that HubSpot's free Bombora bundle started, and independent benchmarking put 6sense and Demandbase at feature parity. Quoting produced the fortnight's cleanest production evidence: Atira, freshly funded with $17.5M, has fifteen paying industrial customers with Robel saving 95 hours per quotation; Talkulate cut a US reseller's quote cycle from one to two days to fifteen minutes while lifting first-pass accuracy from 76% to 100%. The regulators arrived at the same time. FINRA added its first standalone generative AI section to its annual oversight report, warning that autonomous agents in financial workflows require novel controls against hallucination; a Saturn Report found 57% error rates in AI financial guidance rising to 88% on complex queries; and a Crux Digits analysis showed a single 4% mispriced job per month cancels a year of automation savings. In pricing, NYU SPS research across 58,000 hotel properties found more than half using AI but fewer than 10% achieving meaningful impact, while Fortune's account of algorithmic "ghost cartels" documented Amazon's Project Nessie generating $1B in excess profit and German petrol stations gaining 38% margin from independent algorithms that learned to coordinate without any human agreement. Delta scaled Fetcherr from 3% to 20% of its domestic network; Mews delivered a 13% revenue-per-square-metre lift across 6,000 hotels. Finally, the measurement gap widened: a critical review found no independent accuracy tests of deal-risk scoring across any major vendor, with the largest published ROI claim of 481% tracing to vendor self-report; a forecasting benchmark set 15% deviation as the defensible floor and found 79% of B2B organisations still missing by 10% or more; and aggregated 2026 statistics show win rates down 28% cumulatively over two years with 36% of forecast deals slipping. No practice changed its assessment this cycle. In a domain this mature, stability is the finding.
Key Tensions
Replacement is dead; the co-worker model won on the evidence and now on the marketing. Artisan's retreat from "Stop Hiring Humans" follows the numbers: human SDRs out-earn AI-only outreach 2.6 to 1, hybrid lead qualification hits 87% accuracy against 76% human-only, and nearly half of autonomous outreach programmes hit a deliverability wall inside 90 days. The autonomy that survives is narrow and governance-gated: Siemens' agents engage every lead but may quote only technically valid upgrades, and Centific's 12.7% task success rate for general-purpose sales agents shows why. Buyers should treat any pitch premised on removing the human from the loop as a claim the vendors themselves have stopped making.
Data governance has become a C-suite liability, not a RevOps chore. Seventy-eight percent of executives have acted on AI recommendations they later doubted because of data quality; only 21% rate their CRM data well prepared; 62% of organisations have lost revenue to bad records; and manual CRM maintenance is the top operational bottleneck for 37.6% of revenue teams. The correlation now runs the other way too: high-growth companies are four times more likely to have advanced governance. Platforms are trying to automate the problem away with self-updating CRMs and auto-capture, but automated capture of an undocumented process captures the mess faster, and ungoverned business-built agents fragment records rather than repair them.
Vendors publish ROI; nobody publishes accuracy. No independent accuracy test of deal-risk scoring exists across any major vendor; the 481% ROI figure is self-reported; Aviso's 98% accuracy claim is unaudited; Gong's theme clustering cannot infer causality without frozen closed-lost cohorts; and only three of ten marketed conversation-intelligence tools do deal-risk analysis at all. The ICML study putting unaided agent task completion at 30.3% and Centific's 12.7% are the only independent numbers on the table, and both are bad. The practical consequence is that buyers must run their own frozen-cohort measurement, because nothing in the vendor literature substitutes for it.
The input layers are commoditising to zero, and value is migrating one layer up. Transcription now costs $0.004 a minute from Microsoft; Teams generates meeting insights natively; HubSpot gives away intent data that cost $25,000 standalone; G2, Bombora, Intentsify and Clay have made the intent stack modular; 6sense and Demandbase are at feature parity. What still commands a premium is the layer above the feed: activation within the 24-to-72-hour half-life of an engagement signal, compound-signal verification before outreach, and the ability to answer "who do I call?" when account-level scoring has destroyed contact-level granularity. Every conversation-intelligence and intent-data renewal in the next twelve months should be renegotiated on that basis.
The first draft is free; the approval is not, and the 6% who win redesigned the workflow rather than the tool. Knak found 47% of teams naming approvals and sign-offs, not generation, as their biggest delay; Pipedrive found only 12.1% of professionals rate AI-written material as matching human quality and 44% never use it for communications. A 120-person SaaS company lost $1.4M of quarterly pipeline to an eleven-day median approval queue. Microsoft's own transformation, disclosed this fortnight, delivered a 20% higher close rate and 9.4% more revenue per account manager through workflow redesign and change management, which is precisely the pattern McKinsey finds separates its 6% of high performers from everyone else. The tooling is the same; the organisations are not.
Top 10 Evidence Items
Revenue teams all use AI, but only 20.6% call it production-ready (adoption-metric) — the Salesloft survey of 500 US revenue leaders is the cleanest statement of this fortnight's central paradox: universal adoption sitting on top of a fifth of verified production value, with manual CRM updates named the top bottleneck. https://www.marketscale.com/industries/marketing-tech/100-of-revenue-leaders-say-they-use-ai-but-most-cant-prove-it-pays-off
Enterprise hits and misses: bad data gives Marketing an AI hangover (industry-report) — Validity's finding that 78% of C-suite executives have acted on AI recommendations they later suspected were wrong, and that 62% of organisations have lost revenue to bad records, is the evidence that turned data quality from a RevOps chore into a boardroom liability. https://diginomica.com/enterprise-hits-and-misses-bad-data-gives-marketing-ai-hangover-jensen-huang-gives-linkedin-agi
How CIOs Can Govern AI Agents Built by Business Users (industry-report) — BCG's warning that ungoverned, business-built agents fragment customer records rather than enrich them is the sharpest articulation of why platforms racing to auto-capture CRM data (HubSpot's Fall '26 release) are treating a symptom, not the disease. https://www.bcg.com/publications/2026/cios-govern-business-built-ai-agents
Harris Poll for Collibra: 72% of tech decision-makers feel AI initiatives are falling short (adoption-metric) — an independent sixth sample lining up behind Salesloft, McKinsey, Deloitte/Teradata and Gartner: 76% hit critical pilot-to-production roadblocks, and 72% name an unaligned or poor data foundation as the cause. https://www.prnewswire.com/news-releases/new-survey-from-collibra-by-the-harris-poll-finds-72-of-tech-decision-makers-feel-ai-initiatives-today-are-falling-short-302879680.html
The ghost cartel — your pricing algorithm may have stopped competing without your knowledge (opinion/investigative) — Amazon's Project Nessie generating $1B in excess profit and German petrol stations gaining 38% margin from algorithms that learned to coordinate without human agreement is the domain's most uncomfortable deployment story, and a preview of the antitrust exposure dynamic pricing is walking into. https://fortune.com/2026/09/12/ghost-cartel-algorithms-antitrust-law/ (403 on automated check — Fortune paywall; page renders for readers)
Why AI agents struggle in enterprise sales operations (case-study) — Centific's testing of five leading models across 500+ production-like sales-operations tasks returning a 12.7% agent success rate is the number that should sit beside every autonomy claim in this domain, and it does, in the Key Tensions above. https://www.centific.com/blog/why-ai-agents-struggle-in-enterprise-sales-operations
Best revenue intelligence platforms for deal insights (opinion) — the meta-analysis tracing the industry's largest published ROI claim (481%) to vendor self-report, with no independent accuracy test of deal-risk scoring existing across any major platform, is the clearest documentation of the measurement gap the domain has yet to close. https://www.attention.com/blog-posts/best-revenue-intelligence-platforms-for-deal-insights
As Regulators Warn of AI Hallucinations in Financial Guidance — FINRA Adds AI Section (industry-report) — FINRA's first standalone generative AI section in its annual oversight report, requiring novel controls against hallucination in autonomous financial workflows, marks the moment quote and contract generation moved from an efficiency play to a compliance obligation. https://www.morningstar.com/news/pr-newswire/20260915sf47636/as-regulators-warn-of-ai-hallucinations-in-financial-guidance-financial-finesses-closed-loop-ai-earns-fresh-industry-recognition (403 on automated check — press-release syndication site returns a soft paywall to bots; content is publicly indexed)
Siemens Engages 100% of Its Inbound Leads with Agentforce (case-study) — 18,000 sellers and 2,800+ weekly leads across 132 countries is the domain's largest verified autonomous-qualification deployment, and its governance gate — agents may quote only technically valid upgrades — is exactly the narrow, constrained autonomy that survived this fortnight's replacement-thesis collapse. https://www.salesforce.com/uk/news/press-releases/2026/09/15/siemens-agentforce-redefine-industrial-sales-service/?bc=OTH
AI SDRs in 2026: What Is Working and What Is Quietly Failing (adoption-metric) — Empiraa's field data that 86% of teams see positive first-year returns but 47% hit a domain-reputation wall within 90 days, with 21% never recovering placement, is the mechanism behind the fortnight's broader retreat from full-autonomy outreach. https://www.empiraa.com/blog/ai-sdr-what-works-2026