AI inside business functions

Original research edition

This is the supplied September 2026 report, including its caveats and source links. Its prices, vendor events and verification claims have not all been independently rechecked for this preview. Consult the original correction log alongside the chapters.

AI Across Business Functions: A Buyer's Field Guide

Research compiled September 15, 2026. All pricing verified against vendor pricing pages or 2026 sources unless flagged in the "Could Not Verify" section at the end. Prices are US list, per month, and generally reflect annual-commitment rates where a vendor offers both.

A useful frame for everything below is the distinction between AI-enhanced and AI-native products. An AI-enhanced product is an existing system of record — a CRM, a helpdesk, a general ledger — that has had models bolted onto its workflows. Its advantage is that it already owns your data and your users' habits. An AI-native product was designed on the assumption that a model does the work, and it usually prices accordingly: not per seat, but per outcome. The single most important commercial development of 2025–2026 is that outcome pricing has escaped customer support, where it started, and spread into CRM (Salesforce), marketing (HubSpot), and even bookkeeping (Digits). That shift matters more to a buyer's budget than any individual feature.

A. CRM and Sales

Salesforce (Agentforce) is the AI-enhanced incumbent, and in 2026 it finally has a coherent price. Salesforce abandoned its original $2-per-conversation model in favor of Flex Credits, which bill per action rather than per conversation. The published examples on Salesforce's pricing page are concrete: a case-management interaction of three actions costs 60 credits, or $0.30; field-service appointment scheduling at six actions costs $0.60; a single employee-onboarding action costs $0.10. Seat-based options sit alongside it — Agentforce add-ons for Sales, Service and Field Service run $125/user/month with unmetered internal usage, Industries add-ons $150/user/month, and the bundled Agentforce 1 Editions start at $550/user/month including 1M Flex Credits and 2.5M Data 360 Credits annually. There is also a $5/user/month Agentforce User License for company-wide deployment that requires Flex Credits on top. Strengths: it sits on the CRM data that most large enterprises already run their revenue on, and the per-action model at least makes cost forecastable. Weaknesses: the total cost of a real deployment is the product of three variables (seats, credits, Data Cloud consumption), and Salesforce implementation costs remain substantial. Best for: enterprises already deep in Salesforce, particularly service organizations with high, repetitive case volume.

HubSpot (Breeze) is the mid-market equivalent and has made the same move. HubSpot's core platform remains cheap at the entry point — a Starter Customer Platform seat is $7/month (promotional, from a $20 list) — but AI is metered through HubSpot Credits. Reported rates put extra credits at roughly $10 per 1,000 (about a cent each), with a Customer Agent resolution at 50 credits (~~$0.50), a Prospecting Agent lead recommendation at 100 credits (~~$1.00), and an AI-generated article at 1,000 credits (~$10). Breeze Assistant and standard enrichment are free. Be warned: HubSpot's published allowances per tier and what partners report in the field do not match, so any buyer should check Settings → Usage & Limits rather than trusting a number from a blog. Best for: SMB and mid-market go-to-market teams who want marketing, sales and support in one place and will tolerate credit accounting.

Clay is the most genuinely AI-native tool in this category and has become the default for outbound research. It is a spreadsheet-shaped orchestration layer that waterfalls across dozens of data providers and uses AI to research and write. Current pricing: Free (100 data credits, 500 actions/month), Launch at $167/month monthly or from $54/month annually (3,000 data credits, 15,000 actions), Growth at $446/month monthly or from $185/month annually (6,000 credits, 40,000 actions), and custom Enterprise. Strengths: nothing else matches its flexibility for building bespoke enrichment and personalization. Weaknesses: the credit math is genuinely hard to predict, it requires an operator who enjoys building systems, and it produces the volume of mediocre personalized email that is currently killing cold outbound. Best for: sales-ops and growth engineers at companies with a dedicated outbound motion.

Attio is the AI-native CRM itself — a fast, flexible data model aimed squarely at startups fleeing Salesforce. Pricing is transparent: Free, Plus at $35/user/month annually ($44 monthly), Pro at $79/user/month annually ($99 monthly), and custom Enterprise. AI is bundled through seat credits — 100/user/month on Free rising to 1,000 on Pro and 2,500 on Enterprise — with call intelligence, auto-labeling and auto-summaries at Pro and above. Best for: venture-backed startups and investment firms who want a CRM that feels like Notion rather than an ERP.

Gong owns revenue intelligence: it records, transcribes and analyzes sales conversations, then surfaces deal risk and coaching signal. Gong does not publish prices, and every number in circulation is a third-party estimate — commonly cited figures are a mandatory annual platform fee scaling from roughly $5,000 (1–20 users) to $25,000–$50,000 (100+ users), plus $1,200–$1,920 per user per year depending on tier. Treat those as directional only. Strengths: the data asset is real and the coaching use case has the clearest ROI story in the category. Weaknesses: expensive, annual-contract-only, and narrow — it is conversation intelligence, not a sales engagement platform.

Sales engagement (briefly). Apollo.io bundles a B2B contact database with sequencing at $49/seat/month annually (Basic), $79 (Professional) and $119 (Organization, 3-seat minimum), with a usable free tier; credits do not roll over. Outreach is the enterprise-grade sequencing and forecasting platform and, like Gong, does not publish pricing. The honest assessment of this sub-category in 2026 is that AI has made sending email trivially cheap and therefore made email itself much less effective; buyers should weight deliverability and data quality over generation features.

B. Customer Support

This is where AI has produced the clearest, most measurable business results, and where pricing innovation started.

Intercom (Fin) is the category's commercial benchmark. Intercom seats run $29/seat/month (Essential), $85 (Advanced) and $132 (Expert), with Fin included on all of them, and Fin itself is billed from $0.99 per outcome. The word "outcome" is doing work there — Intercom moved from "resolution" to "outcome" language, and its pricing page does not define the billable trigger, which is a fair criticism. Strengths: Fin is the best-documented AI support agent, it works on top of an existing helpdesk (including Zendesk and Salesforce), and $0.99 against a typical $4–$8 human ticket cost is an easy CFO conversation. Weaknesses: resolution rates depend almost entirely on the quality of your knowledge base, and the outcome meter makes a bad month expensive.

Zendesk AI is the AI-enhanced incumbent response. Suite pricing is $55/agent/month (Team), $115 (Professional) annually, with Enterprise quoted. AI agents are included in every plan but billed on Automated Resolutions, and Copilot — the assist tool for human agents — is a $50/agent/month add-on. Zendesk does not publish per-resolution rates; third-party analyses converge on roughly $1.50 per resolution on committed volume and about $2.00 pay-as-you-go, with only a small included allotment (on the order of 5–15 per agent-month). Treat those as unverified. Strengths: enormous installed base, mature ticketing, and AI that requires no migration. Weaknesses: the stacked meters — seat, Copilot add-on, and resolutions — make Zendesk one of the harder support bills to forecast, and the AI agent is generally considered a step behind Fin, Sierra and Decagon.

Sierra and Decagon are the AI-native challengers, and their 2026 numbers are the strongest evidence that this category is real. Sierra, founded by Bret Taylor, raised $950M in May 2026 at a $15.8B valuation, with ARR moving from $100M in November 2025 to $150M in February 2026, and claims over 40% of the Fortune 50 as customers. It charges per resolved interaction — commonly reported around $1.50 — with enterprise contracts typically starting near $150,000 annually. Decagon is the closer competitor: **$100M annualized revenue as of July 2026**, a $4.5B valuation, $481M raised, and customers including Deutsche Telekom, Avis Budget Group, Notion and Duolingo. It offers both per-conversation and per-resolution models. Both do voice as well as chat and email.

The resolution-pricing trend deserves a paragraph on its own, because it changes what buyers should measure. Under seat pricing, a support leader's job was headcount planning. Under resolution pricing, the number that matters is deflection rate at acceptable quality — and vendors have a structural incentive to classify marginal interactions as resolutions. Practical advice for buyers: negotiate the definition of a billable resolution in the contract, insist on a CSAT floor tied to AI-handled conversations, demand a shadow-mode pilot on historical tickets before committing, and model the downside case where volume doubles. Bret Taylor's own framing in 2026 was that the ramp-up phase is expensive before returns materialize, which is an unusually candid thing for a vendor to say and worth taking seriously.

C. Marketing

Jasper was the first breakout AI writing company and has narrowed to survive. It now sells itself as "the AI purpose-built for marketing," emphasizing brand-voice training, governance and distribution into the tools marketers already use. Pricing is $59/seat/month annually ($69 monthly) for Pro, with custom Business pricing. The strategic problem is unavoidable: general-purpose models write well enough for most marketing copy at $20/month, so Jasper's defensibility rests entirely on brand memory, workflow and team controls rather than output quality. Best for: marketing teams of 5–50 who need consistent brand voice across many contributors and want an auditable system rather than a chat window.

Copy.ai made the pivot the question anticipates, and it is complete. It now positions as a GTM AI platform — an "AI OS" for sales, marketing and operations workflows — rather than a copywriting tool. Pricing reflects the reposition dramatically: a $24–$29/month Chat tier for individuals, then a jump to Growth at $1,000/month (75 seats, 20K workflow credits), Expansion at $2,000/month (150 seats, 45K credits), and Scale at $3,000/month (200 seats, 75K credits), all annually. This is now an enterprise workflow-automation purchase, not a writing tool, and buyers evaluating it against Jasper are comparing different categories.

HubSpot AI (Breeze) is covered above; for marketers specifically, the calculation is that content generation costs real credits (~$10 per AI-generated article) while the assistant layer is free. If you are already a HubSpot customer, the marginal cost of trying Breeze is low; if you are not, Breeze is not a reason to migrate.

AdCreative.ai represents the generative-creative niche: it produces ad creative variants scored against predicted performance. Pricing runs Starter from $20/month annually ($39 monthly, 10 credits), Professional $125/month annually ($249 monthly, 50 credits), and Ultimate $500/month annually ($999 monthly, 100 credits), with credit sub-tiers within each. Strengths: genuinely fast volume creative for performance marketing, with platform integrations. Weaknesses: output is recognizably templated, the credit system is confusing, and the underlying capability is increasingly available directly from Meta's and Google's own ad tools for free.

SEO tools have quietly become AI-visibility tools, which is the most interesting shift in this category. Surfer now sells an AI Search Analytics plan at $82/month dedicated purely to tracking brand visibility across ChatGPT, Gemini, AI Overviews and Perplexity, alongside its content plans (Discovery $49, Standard $99, Pro $182, Peace of Mind $299, Enterprise from $999). Clearscope has gone further: its plans are now denominated in tracked prompts — Essentials $129/month (50 tracked prompts, 50 pages) and Business $399/month (300 prompts, 300 pages) — with brand visibility monitoring in every tier. A content-optimization company repricing its product around prompt tracking is a strong signal about where the demand went.

The GEO/AEO trend is therefore not hype; it is a budget line that already exists. Purpose-built entrants like Profound sell Starter at $99/month (ChatGPT only, 50 prompts), Growth at $399/month (three answer engines, 100 prompts) and custom Enterprise (up to nine engines). The measurable case for this spend is weaker than vendors suggest, though. Nobody has published a credible, independent link between AI-assistant citation share and revenue, the underlying answer engines change their retrieval behavior without notice, and much of what these tools sell is monitoring rather than influence. The defensible reason to buy one in 2026 is diagnostic: if your brand is being misdescribed or omitted in the assistants your buyers use, you want to know. Paying for that knowledge at $99–$399/month is reasonable; paying five figures for "GEO optimization" services is not yet supported by evidence.

D. Finance and Administration

Ramp is the standout, and its trajectory is the clearest proof of the "AI applied to a boring workflow beats AI applied to a chat window" thesis. Ramp raised $750M at a $44B valuation in June 2026. Its pricing is unusual for the AI era in being mostly free: Ramp Free at $0/user/month with basic accounting rules and AI reporting; Ramp Plus at $15/user/month plus a platform fee, which adds AI-driven expense reviews and policy insights, auto-coded line items, and AP-agent approval recommendations; and custom Enterprise with Workday and Oracle Fusion integrations. Ramp can price this way because it earns interchange on card spend — which means the AI is genuinely a customer-acquisition feature rather than a revenue line. Strengths: the AI does unglamorous, verifiable work (coding transactions, flagging out-of-policy spend) where correctness is checkable. Weaknesses: you are adopting a card and spend platform, not just a tool, and the free tier's economics depend on you routing spend through Ramp.

QuickBooks has rebranded its AI layer as Intuit Intelligence, described as a "connected intelligence layer" with access to human experts, rolled out in August 2026. QuickBooks Online pricing (list, with a common 50%-off-three-months promotion) is Simple Start $38, Essentials $85, Plus $140, Advanced $340 per month. AI features are bundled rather than sold separately, but they are tiered: conversational business intelligence and task automation across all plans, proactive alerts at Plus and Advanced, and presentation-ready management reports and unlimited custom KPIs only at Advanced. Note that "Intuit Assist" appears to have been folded into or renamed as Intuit Intelligence; Intuit's own product-update page does not mention the older name, so the relationship between the two brands is not something I could establish definitively.

Puzzle is the AI-native startup ledger and publishes clean pricing: Starter $25/month annually ($30 monthly), Core $60 ($72), Complete $100 ($120), Scale $300 ($360), with AI credits as the metered resource — 25 lifetime credits on the lower tiers, 100/month on Complete and 300/month on Scale. Complete adds AI-powered reconciliations, accuracy review and insights. Best for: seed-to-Series-B startups who want real-time books rather than a monthly bookkeeper.

Digits made the most interesting pricing move in the category. In April 2026 it announced outcome-based pricing for accounting firms: it charges only for clients where 95% or more of transactions are "zero-touch" — neither created nor edited by a human accountant before close. CEO Jeff Seibert's framing was "if Digits does the work, we should get paid; if it doesn't, we shouldn't." The company cited a top-400 firm moving from roughly 75% to 98% automated transaction handling during 2025. Exact dollar rates were not disclosed. This is the resolution-pricing model from customer support arriving in accounting, and it is a template other vertical vendors will copy.

AP automation is the least glamorous and most reliably valuable application. BILL prices at $49/user/month (Essentials), $65 (Team) and $89 (Corporate), with a W-9 Agent for automatic vendor tax-form collection available from Essentials and an Invoice Coding Agent for multi-line bill coding from Team upward. The pattern across all of these products is worth naming: the finance AI that works is narrow, checkable, and embedded in an existing approval workflow. Nobody is asking a model to close the books unsupervised, and the vendors that promise that should be treated skeptically.

E. Human Resources

Recruiting AI consolidated. The headline event is that Workday completed its acquisition of Paradox on October 1, 2025 for a reported ~$1B. Paradox's conversational assistant Olivia — text-based screening, scheduling, application collection and onboarding for high-volume hourly hiring — now ships as the Workday Paradox Candidate Experience Agent. Paradox never published pricing and still doesn't; third-party estimates suggest roughly $30,000–$95,000 annually for mid-size deployments plus $15,000–$35,000 in setup, which should be treated as unverified. The acquisition narrows the market: Paradox's natural buyer is now a Workday customer, and standalone frontline-hiring AI has one fewer independent option.

HireVue is the cautionary tale the education debate has an analogue of. In March 2025 the ACLU of Colorado filed complaints with the Colorado Civil Rights Division and the EEOC alleging that Intuit's use of HireVue's AI video interview platform discriminated against a deaf and Indigenous employee — that automated speech recognition performed worse for deaf speakers and Indigenous English dialects, that a captioning accommodation was denied, and that a promotion was declined on "communication style." Both companies deny the allegations; HireVue's CEO called the complaint "entirely without merit" and said Intuit did not use HireVue's AI assessment tool. Regardless of outcome, the compliance environment has hardened considerably — NYC Local Law 144, Illinois, Colorado and Ohio all now impose obligations on automated employment decision tools. Buyers of AI hiring software in 2026 should assume they carry the legal exposure, not the vendor, and should demand bias audit documentation as a contract condition.

Rippling and Workday are the platform players. Rippling's published entry point is $8 per employee per month for its Unity platform plus a small flat monthly fee, but realistic all-in costs land around $20–$35 per employee per month once payroll (~~$6 PEPM), benefits administration (~~$4 PEPM), recruiting (~$4) and IT management ($5–$20) are added; median verified annual contracts sit near $40,000. Rippling's AI story is comparatively thin — its strength is the unified employee record across HR, IT and finance, which is a data advantage that could matter later more than any current feature.

Lattice is the clearest and cheapest AI-in-HR purchase to evaluate, and it publishes everything: Foundations at $13/seat/month (1:1s, feedback, performance, goals, analytics, AI Agent and integrations); modular Performance $10, Goals & OKRs $8, Engagement $4; add-ons for Compensation (+$6) and Grow (+$4); annual billing with a $4,000 minimum. The AI Agent is included rather than metered — described as acting on your data, answering employee questions and coaching managers in the flow of work. Best for: 50–1,000-person companies that want performance and engagement infrastructure and will get incidental AI value; not a reason to buy on its own.

F. Education

Khanmigo remains the most defensible consumer AI-education purchase on price alone: free for teachers in 44+ countries, $4/month or $44/year for learners and parents, custom for districts. It is deliberately Socratic — it withholds answers and asks questions — and that design choice turns out to be the one the evidence supports.

Duolingo Max sits at the top of Duolingo's ladder: Super at ~$12.99/month or $83.99–$95.99/year (family $119.99/year), and Max at ~$29.99/month or ~$168/year (family $239.99/year). Max adds Explain My Answer, Roleplay and Video Call with an AI character, though AI features remain limited to a subset of languages. The honest read is that Max's AI features are pleasant rather than transformative, and Duolingo's core engagement mechanics do most of the work.

ChatGPT and Claude in education. OpenAI's consumer ladder as of September 2026 is Free, Go at $8/month, Plus at $20, Pro at $100 or $200 depending on the usage multiplier, Business at $25/seat/month ($20 annually) with a Business Premium at $125 ($100 annually) added in August 2026, and custom Enterprise. Anthropic's is Free, Pro at $17/month annually ($20 monthly), Max from $100, Team at $20/seat/month annually ($25 monthly, 2–150 seats) with a Premium seat at $100 ($125 monthly), and Enterprise at $20/seat/month plus usage-based API costs, billed annually. Claude for Education is sold as a university-wide plan with undisclosed pricing; named partners include Northeastern, LSE, Dartmouth, the University of Virginia, Syracuse, the University of San Francisco, Champlain, Northumbria and the University of Pittsburgh. Its differentiating feature is Learning Mode, which asks questions rather than supplying answers — the same pedagogical bet Khanmigo makes. ChatGPT Edu's specific institutional pricing is likewise not published.

Quizlet is now the low-stakes case: Quizlet Plus at $19.99/year, with AI features (Magic Notes, Q-Chat) that the upgrade page does not itemize clearly. At that price the question is not whether the AI is good but whether flashcards remain a useful primitive when a model will generate them on demand for free.

The evidence debate is the most important thing in this section, and it has genuinely sharpened. Three findings define the current state:

First, well-designed AI tutoring can produce large gains. A Harvard randomized crossover trial with 194 students, published in Scientific Reports, found AI-tutored students scored a median 4.5 versus 3.5 for in-class active learning (p < 10⁻⁸), roughly 0.63 standard deviations by linear regression and 0.73–1.3 SD by quantile regression, in less time (median 49 minutes versus 60). But the authors' own limitations matter: the content sat at the understanding/applying/analyzing levels of Bloom's taxonomy, success depended on expert-crafted question-specific prompts, and the system used pre-written answers rather than live LLM reasoning specifically to avoid hallucination. That is not "give students ChatGPT."

Second, unrestricted chatbot use appears to harm retention. A randomized controlled trial published November 2025 with 120 undergraduates found the ChatGPT group scored 57.5% versus 68.5% on a surprise test 45 days later (Cohen's d = 0.68) — consistent with cognitive offloading theory and the "desirable difficulties" literature.

Third, and most sobering, Stanford's SCALE Initiative reviewed more than 800 papers on AI in K-12 and found only about 20 high-quality causal studies. Its conclusions: students often perform better on tasks with AI but results are mixed when assessed without it; tools with pedagogical guardrails outperform general-purpose chatbots; there is limited evidence AI reduces teacher lesson-prep time without quality loss; and there are effectively no rigorous studies of student AI use in U.S. K-12 classrooms, nor meaningful research on equity, wellness or long-term outcomes.

The practical synthesis for a buyer or a parent: pay for tools that make the student do the work (Khanmigo, Learning Mode), be skeptical of tools that produce answers, and treat any vendor citing "studies show" without naming a causal design as marketing.

G. Personal Organization and Personal Assistants

This is the weakest category in the entire report, and the honest finding is that the consumer personal-AI-assistant market has not produced a durable winner.

Siri and Apple Intelligence. The status as of this writing is dramatic and recent: after roughly two years of delays, Apple shipped "Siri AI" on September 14, 2026 alongside iOS 27, iPadOS 27, macOS 27, watchOS 27 and visionOS 27 — in beta. The new assistant offers personal-context understanding (searching your messages, emails and photos), onscreen awareness, systemwide app actions, and multi-turn conversation. Critically, Apple confirmed the models were "custom-built in collaboration with Google and its Gemini models" — worth stating precisely, because this is an Apple Foundation Model trained with Google, running on Apple's own infrastructure, not Gemini answering queries directly. The newest on-device models are limited to iPhone Air and iPhone 17 Pro and later. Availability is limited: iPhone 16 and later plus iPhone 15 Pro/Pro Max, M1-or-later iPads and Macs, Apple Watch Series 9+, and Vision Pro; English at launch with French, Japanese, Korean, Portuguese and Spanish the following month; not available in the EU or China at launch for regulatory reasons. Apple stated no upfront price but noted expanded access to certain features "will be available for a fee in the future" — the first explicit signal that Apple intends to charge for AI. It is far too early to assess quality; the two-year delay, the shipped-in-beta label, and the reliance on a competitor's models are all facts a buyer should weigh before upgrading hardware for it.

AI journaling and life-organizer apps (Rosebud, Mindsera, Reflection, Day One's AI features) are a real but small category, typically $8–$15/month, offering guided reflection, pattern detection across entries and mood tracking. I could not verify current pricing for the leading apps against their own pricing pages, so I am declining to quote figures. The category's honest assessment: these are wellness products whose value is mostly the habit, not the AI, and the privacy trade — handing a model your most private writing — deserves more scrutiny than the category gives it.

The Rabbit and Humane postmortem is the most useful thing in this section. Humane's AI Pin died in February 2025 when HP acquired the company's assets for $116M and shut the device down, stranding customers. Rabbit survived by retreating from hardware: the r1 received a substantial RabbitOS 2 rebuild in September 2025, and the company launched a second product, "rabbit intern," a software agent, in June 2025 — effectively conceding that the software was the product and the $199 device was a distribution mistake. The lessons generalize: a dedicated AI device must do something a phone cannot, latency and battery are product-defining rather than engineering details, demos are not products, and buying first-generation AI hardware means accepting that the company may not outlive the warranty. Those lessons apply directly to the next section.

H. Mobile and Wearable Assistants

Phone-level AI is now a three-way default fight, and the striking thing is how little of it is paid. Apple just shipped Siri AI (above) on Gemini models, free for now with paid tiers signaled. Samsung has confirmed that Galaxy AI features remain free — with caveats about which features and for how long — and Galaxy S26 devices ship with deeper Gemini integration. Google Pixel remains the reference implementation, with Gemini Nano on-device and Gemini in the cloud. For a consumer, the practical consequence is that phone AI is not a purchase decision in 2026; it is a byproduct of which handset you already own. Nobody should buy a phone for its assistant this year.

Smart glasses are the one wearable category with genuine commercial traction. EssilorLuxottica reported that Meta AI glasses sales more than tripled, with the company crediting the Ray-Ban Meta line for a meaningful share of its revenue growth; Ray-Ban Meta frames sell at $299, the Oakley Meta variant at $499, and the display-equipped Meta Ray-Ban Display launched at $800. Meta's stated target is 10 million glasses annually by the end of 2026. Strengths: glasses solve the form-factor problem Humane could not — they are socially legible, they already have a reason to be on your face, and camera-plus-audio is the right sensor set for an ambient assistant. Weaknesses: battery life, the ongoing bystander-privacy problem, and the fact that the display model is a first-generation product at a premium price.

Earbuds and pendants are more contested. Limitless, the best-known always-on pendant, was acquired by Meta in December 2025 and stopped selling the Pendant on December 5, 2025; Meta committed to supporting existing customers for at least a year, made the Unlimited Plan free for them, provided data export, and withdrew from the EU, UK, Brazil, China, South Korea, Turkey and Israel with user data in those markets deleted on December 19, 2025. Sources disagree about whether the device is still shipping in any form; the acquisition is well documented, continued availability is not. Bee, acquired by Amazon in July 2025, sells at roughly $49.99 with optional premium around $12/month and is the cheapest entry point. Plaud NotePin at roughly $169 plus subscription takes a different and arguably wiser tack: it is a press-to-record device rather than always-listening, which sidesteps the consent problem entirely. That distinction — always-on versus deliberate capture — is the single most important thing for a buyer to understand in this category, legally and socially.

OpenAI hardware remains unshipped but is now officially acknowledged. OpenAI's Chris Lehane said publicly at Davos in January 2026 that the first consumer device would launch in the second half of 2026, with reporting pointing to AI earbuds (internal codename "Sweetpea") using a 2nm Samsung Exynos chip, and a first-year sales target of 40–50 million units. Separately, a February 2026 report indicated the Jony Ive-led hardware had slipped to no earlier than end of February 2027, and OpenAI confirmed it had abandoned the "io" name after a trademark dispute. These two threads may describe different products or a shifting schedule; as of September 15, 2026 nothing has shipped, and the 40–50 million first-year target should be read as ambition rather than forecast.

I. Enterprise AI Platforms and Vertical AI

The horizontal platforms have converged on remarkably similar pricing, which tells you the market is now competing on integration rather than capability. Microsoft 365 Copilot is $30/user/month for enterprise (annual), with a cheaper SMB route — Copilot as a Business add-on from $18/user/month annually, or bundled into Microsoft 365 Business Standard with Copilot at $23.50 and Business Premium with Copilot at $32. Gemini Enterprise ships in editions reported at Business $21/user/month (capped around 300 users), Standard $30 annual / $35 monthly, Plus $50 annual / $60 monthly, plus a Frontline tier requiring at least 150 Standard or Plus seats; Google does not publish these on a single canonical page, so treat the edition names and caps as well-sourced but not vendor-confirmed. Claude Enterprise is $20/seat/month plus usage-based API costs, billed annually, with SSO, SCIM, audit logs and a HIPAA-ready option. ChatGPT Business is $25/seat/month ($20 annually) with Enterprise quoted.

The adoption evidence for Copilot is the most-scrutinized number in enterprise software, and it is genuinely ambiguous. Microsoft reported 30 million paid Microsoft 365 Copilot seats in its FY26 Q4 (quarter ended June 2026), up from 20 million the prior quarter — a 10-million-seat quarter, the fastest in the product's history. But three weeks before that announcement, leaked internal communications indicated that fewer than 4.5% of Microsoft's ~450 million commercial Microsoft 365 users held paid Copilot licenses, and only 20–30% of those used it weekly. Both facts can be true. Seats contracted is not revenue recognized, enterprise agreements have reportedly involved significant discounting in competitive displacement scenarios, and Microsoft has not disclosed average revenue per seat or weekly active use alongside the seat count. The correct reading for a buyer: Copilot is being purchased at enormous scale and used at far smaller scale, and any organization deploying it should budget for adoption enablement as a larger line item than the licenses.

Vertical AI is where the returns are clearest, and 2026 produced the numbers to argue it. The thesis is simple: a model that knows a domain's documents, workflows and liability structure beats a general assistant, and the vendor can charge for outcomes because the customer's alternative is billable human hours.

Harvey (legal) is the flagship. In September 2026 it raised $550M at a $15.6B valuation with revenue above $400M, serving 3,000+ organizations — up from about 1,300 earlier in the year — including Latham & Watkins, Microsoft's legal department, 80% of the Am Law 100, and five Fortune 10 companies. It also acquired Guardrails AI, a security startup, and launched Harvey LAB as a benchmark for legal AI agents. The growth from roughly $190M ARR earlier in 2026 to $400M+ by September is the steepest curve in applied AI outside the foundation labs.

Medicine has two distinct winners with different business models. Abridge does ambient clinical documentation — turning the doctor-patient conversation into a note, with a "Linked Evidence" feature that traces each line back to the moment in the conversation that produced it. It reached $100M ARR in 2025, a $5.3B valuation in its June 2025 Series E with a $316M extension in April 2026, and is deployed at Kaiser Permanente (24,600 physicians), Mayo Clinic and 90+ health systems, at roughly $2,500 per clinician per year. OpenEvidence is the more radical model: a clinical-evidence assistant trained on licensed content from NEJM, JAMA and specialty guidelines, given away free to verified clinicians and monetized through pharmaceutical advertising at $70–$1,000+ CPM. It reached $300M annualized revenue by July 2026 (from $150M at end-2025), a $12B valuation in January 2026, roughly 20 million clinical consultations per month, and claims over 40% of U.S. physicians use it daily across 10,000+ hospitals. Its free-to-clinician model bypasses hospital procurement entirely — the single smartest go-to-market decision in vertical AI.

Finance and insurance round out the picture. Rogo raised a $160M Series D led by Kleiner Perkins in April 2026 (over $300M total), serving 35,000+ financial professionals across 250+ institutions including Rothschild & Co, Jefferies, Lazard, Moelis and Nomura; its Felix agent handles deal screening, document generation and data-room management. In insurance, Sixfold raised $30M in January 2026 for AI underwriting — a much earlier-stage market, which is itself informative: verticals adopt in order of how well-documented and how expensive the human labor is, and underwriting sits behind law and medicine on both counts.

What the vertical thesis actually says, stripped of hype: the winners are not the companies with better models — most of them build on the same frontier APIs — but the companies that solved distribution into a regulated profession and took on the liability of being wrong. Abridge's Epic integration, OpenEvidence's free-to-physician model, and Harvey's Am Law penetration are distribution achievements. That is the pattern a buyer should look for when evaluating any vertical AI vendor: not benchmark scores, but whether the product is already inside the workflow of people who do the job.

Sources

Could NOT Verify With a Current Primary Source

These figures appear in the report only where explicitly flagged, or were omitted. Treat each as unconfirmed.

HubSpot Breeze credit allowances per tier. HubSpot's public pricing page mentions "500 HubSpot Credits" for Starter but does not explain the credit system. Third-party sources report conflicting allowances — one official-rate reading of 5,000/10,000/15,000 credits for Starter/Professional/Enterprise, versus partner reports of 500/3,000/5,000. The per-action credit costs (50 for a resolved conversation, 100 per lead recommendation, 1,000 per article) come from a single secondary source. HubSpot's knowledge-base URL for credits returned 404.
Zendesk's per-Automated-Resolution price. Zendesk confirms AI agents are billed on Automated Resolutions but publishes no rate. The ~$1.50 committed / ~$2.00 pay-as-you-go figures and the 5–15 included resolutions per agent-month are third-party estimates only.
Gong's entire price list. Gong publishes nothing. All platform-fee bands ($5,000–$50,000/year) and per-user ranges ($1,200–$1,920/year) are third-party estimates from a vendor-comparison blog and should not be quoted as fact.
Outreach pricing. Not published; I found no reliable 2026 figure and therefore quoted none.
Sierra's per-resolution price (~$1.50) and ~$150,000 contract minimum. From a VC analysis blog, not Sierra. Sierra's funding ($950M), valuation ($15.8B) and ARR trajectory are confirmed by TechCrunch.
Decagon's pricing. Per-conversation and per-resolution models are described qualitatively; no rates are published. Valuation ($4.5B) and ~$100M annualized revenue come from Sacra, a research firm, not from Decagon.
Digits' actual dollar rates under outcome-based pricing. The 95%-zero-touch qualification threshold is confirmed by CPA Practice Advisor; the price is not disclosed.
Paradox pricing ($30K–$95K annually plus $15K–$35K setup). Third-party estimate. The Workday acquisition (completed Oct 1, 2025) and the ~$1B price are well sourced; the price was reported, not confirmed by Workday.
Rippling's realistic all-in cost ($20–$35 PEPM) and module prices. From a competitor's blog; the $8 PEPM Unity base is more widely cited but I did not confirm it on Rippling's own page.
ChatGPT plan prices. OpenAI's pricing page redirected and rendered without dollar figures. The Free/$8 Go/$20 Plus/$100–$200 Pro/$25 Business/$125 Business Premium structure comes from a September 2026 third-party guide, not from OpenAI directly.
ChatGPT Edu institutional pricing and Claude for Education pricing. Neither is published; both are quote-based.
Gemini Enterprise edition names, prices and seat caps ($21 Business / $30 Standard / $50 Plus, 300-user cap, 150-seat Frontline minimum). Google's cloud.google.com/gemini-enterprise/pricing and gemini.google/enterprise/ both returned 404; the Workspace pricing page confirms Gemini inclusion and an "AI Expanded Access" add-on but shows no prices. All Gemini Enterprise figures are from a Google partner's blog.
Microsoft 365 Copilot's two prices. The enterprise page shows $30/user/month; the Business page shows an $18/user/month annual add-on "down from $21." I could not reconcile whether these are different SKUs, a segment discount, or a price change, so both are reported as-is.
Quizlet's AI feature set at the Plus tier. The upgrade page lists $19.99/year but does not itemize Magic Notes or Q-Chat.
Khanmigo district pricing. Custom, not published.
Limitless Pendant's current availability. Sources conflict: Meta's acquisition and the December 5, 2025 halt to new sales are well documented, but one 2026 roundup lists the Pendant as "shipping, updated" at ~$99 + $19/month Pro. I could not determine whether it is currently purchasable.
Bee and Plaud current prices (~$49.99 + ~$12/month; ~$169 + subscription). From a review roundup, not vendor pages.
OpenAI's hardware timeline. Directly contradictory sourcing: a January 2026 OpenAI executive statement said H2 2026 for earbuds; a February 2026 report said the Ive-led device ships no earlier than end of February 2027. Nothing has shipped as of September 15, 2026.
AI journaling app pricing (Rosebud, Mindsera, Reflection). Rosebud's pricing URL returned 404 and I did not verify the others, so no figures are quoted.
The "Intuit Assist" → "Intuit Intelligence" relationship. Intuit's own August 2026 product-update page describes Intuit Intelligence but never mentions Intuit Assist, so whether this is a rename, a superset, or a parallel brand is unestablished.
EssilorLuxottica's unit figures for Meta glasses. The "more than tripled" growth claim and the 2-million-pairs-since-2023 figure come from a 2025 article; Bloomberg and CNBC reported updated 2026 figures behind paywalls I could not read. The $299 / $499 / $800 device prices are well sourced.
Whether Samsung's "Galaxy AI stays free" commitment has an end date. Reporting notes a "catch" regarding which features and how long; I could not pin down the terms.