For buyers watching AI reshape software.
The seat count disappeared: Microsoft stopped reporting Copilot seats the week Anthropic's S-1 showed 80% of its revenue is metered
Week 44 of 2026 · October 31, 2026
Big read
The application layer's unit of account changed this week, in public, in three filings and four price lists. Microsoft, which invented the enterprise seat, reported its first quarter without a paid M365 Copilot seat count (the CFO confirmed the omission when asked; UC Today flagged it first on October 28 in 'Seat Count Goes Missing' and drew the seats-to-meters conclusion on October 30) and replaced it with usage: Copilot daily users up about 50% quarter on quarter, customers consuming credits through usage-based billing up 75% quarter on quarter, 50M-plus agents registered in Agent 365, and a Q2 guide for M365 commercial cloud growth 'primarily' from ARPU (average revenue per user, the price of the seat rather than the number of them) with seat growth slowing to 5%. Anthropic's S-1, the first audited-adjacent view of a frontier lab's customer base, says consumption was 80% of first-half revenue, that 11,900-plus customers now pay $100K or more a year (from 950-plus a year earlier) with net revenue retention around 900% (existing customers spending nine times what they spent a year ago, a figure a seat model would reach only by selling nine times the seats), and, the number a CIO should model, that Claude Enterprise users without Claude Code spent under $75 a month on average while the top 5% of Claude Code users spent about 20 times that (a comparison across two populations, not against an all-user average; TechCrunch and Models at Work drew the 20x reading first). ServiceNow's CEO told investors token economics can lift revenue per user about fivefold, raised the AI ACV (annual contract value) target to $1.7B for the second time this year, reported agentic consumption in production up 13x year to date, and launched Flow, a self-serve consumption-billed service desk. Alphabet said it has sold more than 100 million knowledge-work seats and that 2,600-plus customers consumed 100 billion tokens or more in a year, which is the seat and the meter in one sentence.
What a buyer should do differently is read the cliff, not the list. Every price list published this week has a meter with a hard edge: Zendesk's employee-service agents include 10 or 15 automated resolutions per agent per month and then pause until more are bought at $0.99 each; Salesforce's Model Wallet sets hard budget limits with an Agent Kill Switch now GA and Cost Optimization Policies that Salesforce says cut token usage by up to 75% in its own test scenarios; Gemini Enterprise ships spend caps and Smart Routing with its new agent (Google says 90% of queries already use Auto mode, which is the routing decision taken out of the user's hands); OpenAI's Terra reaches ChatGPT Work and Codex (not Chat) under plan allowances that 'cover more tasks' than the prior model, which is a task meter inside a seat. The pattern is budget enforcement moving from the finance team's spreadsheet into the product, and Agent Techniques carries the same movement at the harness level (Claude Code's pooled group budgets, Strands' harness-cost study). The counterbrief's open question, whether seats-to-meters hides or reveals adoption, has a house answer: it reveals it and imports its risk. Consumption ties vendor revenue to work actually done, which is the honest metric, and it imports seasonality and utilisation risk into vendors that now carry fixed take-or-pay compute (contracts that bill for the reserved capacity whether or not it is used); Anthropic's S-1 says so in its own seasonality risk factor, recording that customer spend fell about 20% in the two weeks ending January 2, 2026 against the two weeks ending December 19, 2025, and Microsoft's dropped metric (the last disclosed count, in W31, was above 30 million paid seats) is the first time Microsoft has dropped a Copilot metric since it began reporting one. For a board, the 20x spend concentration is the budget: agent-heavy seats set the bill, average seats do not, and a renewal negotiated on seats is a renewal negotiated on the wrong variable.
The second thing the week settled is where the enterprise value is being concentrated: in the agentic surface, not the chat one. OpenAI disabled new custom GPT creation for Enterprise workspaces and pointed customers to Plugin Creator, and Terra ships to Work and Codex only; ServiceNow and Salesforce each exposed their platforms to other vendors' agents over MCP (Model Context Protocol, the open standard for connecting agents to tools), ServiceNow as a plugin in OpenAI's marketplace and Salesforce through hosted MCP servers to Claude, Slack and Claude Code; and Amazon Quick moved into Outlook, Word, PowerPoint and Excel with named customers (Amazon's Q3 release and earnings page name Principal, GoDaddy, Autodesk, BMW, Mondelez, 3M, Southwest and DXC among Quick's users), putting AWS's work companion inside Microsoft's own applications. Each incumbent is becoming a headless back end to someone else's agent surface while trying to be the surface itself, and the Flex Credit multiplier Salesforce will charge for those headless interactions (Flex Credits are Salesforce's prepaid usage currency; the multiplier sets how many a call consumes) is still 'TBA' with 30 days' notice promised, which is the price the whole pattern waits on. Sierra's disclosure that it crossed $400M of ARR (annual recurring revenue), adding $150M in one quarter with almost half the Fortune 50 as customers and a 'pay for outcomes, not tokens' model, is the company-reported marker that the resolution-priced category is real; the per-resolution prices now published across the field run from HubSpot's $0.50 to Salesforce Help Agent's $2, and Zendesk's $0.99 employee-service overage sits in the middle (its customer-service agents list $1.50 committed and $2.00 pay-as-you-go). Microsoft's offer of voluntary exits to about 60,000 employees (CNBC, citing a person familiar; TechCrunch and Bloomberg report data-centre, sales and marketing staff excluded, and CNBC adds engineering; grade 4 on this publication's 1-5 evidence scale, where 5 is a filing or audited disclosure, 4 a primary disclosure or multi-outlet report, and 3 a single source) is the same substitution seen from the cost side, in Bloomberg's framing of headcount traded for compute, inside the vendor that just stopped counting seats.
Vertical movements
2026-10-26 · OpenAI · Engineering · Frontier Lab · Hybrid
GPT-6.1 Terra in Codex and ChatGPT Work
The release notes say the same plan allowance covers more tasks with Terra than with GPT-5.6 Terra and 'substantially more' than with GPT-6.1 Sol or GPT-6 Astra, and OpenAI reports about nine times lower cost per task than Claude Opus 5.5 at max effort on Agents' Last Exam (vendor-run). The distribution choice is the signal: a model built for computer use goes to the surfaces where work is delegated, and the plan allowance is a task meter inside the seat. UAE inference residency is unsupported at launch. The Model Pulse carries the model; Agent Techniques carries the Codex CLI releases that added Terra to the Vertex and Bedrock catalogues.
2026-10-26 · Zendesk · Hr · Incumbent Saas · Hybrid
Employee-service AI agents with Advanced Action Flows
The adoption figure (6 million executions, an estimated 500,000 hours saved, both Zendesk's) and the confirmed GA window are new against last week's pricing item. The commercial structure is the one to model: Growth and Professional plans include 10 automated resolutions per agent per month, Enterprise and Enterprise Plus 15, and the agent stops working when the allowance is exhausted until more are bought at $0.99 each. Twenty-plus permission-aware knowledge connectors (Google Drive, SharePoint, Confluence, Notion, Slack, Teams, Box) and Slack and Teams surfaces put it directly against ServiceNow EmployeeWorks (which reached $100M of ACV in under three quarters) and Microsoft's employee-service Copilot pitch, each with a different meter.
2026-10-27 · Salesforce · Operations · Incumbent Saas · Usage Based
Agent Fabric: FinOps Dashboard, Model Wallet, Cost Optimization Policies, Agent Kill Switch, Agent Evals
The release moves budget enforcement from the finance team's spreadsheet into the product: Model Wallet sets hard limits, the Agent Kill Switch is GA now with triggered kill switches due January 2027, and the Cost Optimization Policies (semantic caching, payload compression, response trimming) carry a vendor-tested 'up to 75%' token reduction that a buyer should verify on its own traffic. Agent Evals and API Evals (beta November, GA January) add a verification layer. The October 28 Headless Toolkit shows Agentforce exposed through hosted MCP servers to Claude, Slack and Claude Code, but the Flex Credit multiplier for Headless Platform Interactions remains 'TBA' with 30 days' notice promised before metering starts, so the price of the cross-platform pattern is still unknown.
2026-10-29 · Amazon Web Services · Security · Incumbent Saas · Unknown
AWS Continuum
Continuum is the first hyperscaler security package to ship inside the three coding agents at once rather than as a standalone console, which places remediation where the code is written. Named production users at GA distinguish it from Claude Security (public beta for Enterprise, per Anthropic's S-1) and from the ServiceNow-OpenAI cyber solution (early access January 2027, GA Q2 2027). It lands in the same week Google restricts Gemini 4 Carbon to vetted cyber defenders, so the security vertical is where the frontier arrives first and where the application layer is being bundled fastest; the scorecard moves Continuum to challenger.
2026-10-28 · Google · Finance · Frontier Lab · Hybrid
Gemini Enterprise agent with financial-services and legal specialisations, spend caps and Smart Routing
Google Cloud's revenue more than doubled to $30.7B (the Weekly carries the capex and margin story), and Kurian's seat and token figures are the seat-plus-consumption model stated as a single product: the Gemini Enterprise cohort that used it in Q3 2025 more than quadrupled spend year on year, and about 75 customers consumed 10 trillion tokens or more. The new agent's financial-services and legal specialisations compete with Anthropic's seven vertical packages and with Harvey and Legora at the task level, and its spend caps are the same hard-edge metering Zendesk and Salesforce shipped the same week. ServiceNow and Workday are now integrating against Gemini Enterprise as a headless front end.
Sources Alphabet Q3 2026 call transcript; SEC EDGAR (Alphabet 8-K Ex. 99.1)
2026-10-28 · ServiceNow · Hr · Incumbent Saas · Hybrid
EmployeeWorks and Flow
The $100M ACV (annual contract value) figure is the fastest product ramp ServiceNow has reported and lands in the same week Zendesk priced the same category at $0.99 per resolution, so the employee-service vertical now has an incumbent seat-plus-consumption pitch and a resolution-metered one side by side. Flow is the incumbent conceding that a consumption-billed, self-serve desk is where AI-native entrants compete. The professional-services reclassification (about 25 bps, or basis points, hundredths of a percentage point, in Q3 and 50 bps in Q4) is forward-deployed engineering becoming a revenue line, the same move Sierra makes with Ghostwriter. Shares fell about 17% after hours on an 18.5% cRPO (current remaining performance obligations, contracted revenue due inside twelve months) guide; one reading, the house's, is that the market is pricing the meter's volatility rather than the product's adoption, and the plainer one is a whole-company growth guide that missed.
Sources ServiceNow Newsroom (Q3 results)
2026-10-26 · Anthropic · Other · Frontier Lab · Usage Based
Seven vertical solutions (Financial Services, Financial Advisors, Healthcare, Life Sciences, Legal, Education, Government) plus Claude Security in public beta
The filing is the first regulatory-grade account of a frontier lab's packaging strategy: 250 model versions, platform features and applications launched in nine months, more than 50% of $100K-plus customers on both Claude Platform and a first-party app (from about 30% at the end of 2025), and Cowork outpacing Claude Code's adoption curve at the equivalent post-preview point. The vertical packages compete with Harvey, Abridge and the incumbents' specialised agents; the multi-product figure is the retention mechanism behind the ~900% net revenue retention. The Weekly carries the financials and compute commitments; Agent Techniques carries the Claude Code harness economics.
2026-10-29 · Amazon Web Services · Research · Incumbent Saas · Unknown
Amazon Quick in Outlook, Word, PowerPoint and Excel
A hyperscaler without a productivity suite placing its agent inside the incumbent's applications is the headless pattern run against Microsoft itself: Quick becomes the surface, M365 becomes the document store. Bedrock Managed Agents (powered by OpenAI) entered public preview the same day, so the same customer can run an OpenAI-powered agent inside AWS inside Word. Named customers at expansion, rather than at launch, are the datapoint; pricing is not disclosed.
Incumbent responses
2026-10-28 · Microsoft
Microsoft 365 Copilot and Agent 365
The vendor that defined the enterprise seat chose not to show the denominator the week consumption became its growth story; GitHub Copilot's usage-based revenue drove the beat, agents open more than a third of GitHub pull requests, and M365 commercial seat growth slowed to 5%. For a buyer the dropped metric is a negotiating fact: Microsoft is now selling credits and ARPU, and a renewal priced on seats is priced on the variable Microsoft stopped reporting. Pricing model defended: seat plus credits (hybrid).
Sources Microsoft Investor Relations; FY27 Q1 call transcript
2026-10-28 · ServiceNow
OpenAI partnership expansion: MCP plugin, cyber models, Codex in AI Workflow Factory
The incumbent becomes a headless back end to OpenAI's surfaces (the MCP plugin) while embedding OpenAI's models and agents in its own (cyber, Workflow Factory), which is the two-sided posture every platform incumbent adopted this week. The employee rollout of Codex is a reference deployment ServiceNow will sell against; the cyber solution's January early access puts it a quarter behind AWS Continuum's GA. Pricing model defended: subscription plus consumption-billed agentic SKUs (hybrid).
2026-10-28 · Salesforce
Headless Toolkit (Pronto sample app) and Headless Platform Interaction pricing
Salesforce has shipped the pattern before the price. A buyer wiring Agentforce into Claude Code today is accruing a bill at a multiplier that does not yet exist, and the 30-day notice is the only protection; last week's watchlist item stays open, and the Weekly's synthesis treats the multiplier as the test of whether platform incumbents can capture a toll on cross-platform agent traffic. Pricing model defended: prepaid usage credits (Flex Credits) under a per-user licence (hybrid).
Sources Salesforce Developers Blog; Salesforce Help (AIforce Impact)
2026-10-26 · OpenAI
ChatGPT Enterprise custom GPTs and Plugin Creator
OpenAI is retiring the chat-era packaging unit (the custom GPT) and concentrating enterprise value in the agentic surfaces, which is the frontier lab's version of the incumbents' seat-to-meter move: the thing you build and pay for is a plugin an agent calls, not a chat persona. Enterprises with custom GPT estates have a migration to schedule and a retirement date to find. Pricing model defended: seat with a task allowance (hybrid).
Sources OpenAI Help Center (Enterprise and Edu release notes)
2026-10-26 · Microsoft
Voluntary exit program
The exclusion list is the signal: in the quarter Microsoft spent $45.2B on capex with 75% on short-lived compute assets and told investors its model is shifting to consumption, the staff it will not let go are the ones who run the data centres. Bloomberg's framing, headcount traded for compute, fits a vendor that stopped counting seats the same week, and the December decision date puts the result inside the next earnings cycle. Pricing model defended: the same seat-plus-credits hybrid, from the cost side.
Sources CNBC (person familiar); TechCrunch (Bloomberg reporting)
Startup signals
2026-10-30 · Support · $400M+ ARR (company-reported); last valued at $15.8B in a $950M Series E, May 2026
Sierra
The largest disclosed revenue scale for an outcome-priced agent company ('pay for outcomes, not tokens': loans refinanced, claims paid, sales closed), with no gross margin, retention or outcome-revenue mix disclosed, so every figure is Sierra's. It outpaces Decagon (reported above $100M annualised in August) and validates the resolution-priced CX category the incumbents now price at $0.50-$2.00 per resolution. The scorecard holds Sierra as the support leader; the open question is the margin on an outcome when the model bill is a meter.
Sources Sierra (company blog); Sacra
2026-10-27 · Finance · $1.7B + up to $400M earn-out
hyperexponential (acquired by Guidewire)
A vertical-SaaS incumbent paying $1.7B for an agentic-underwriting layer is the insurance vertical's version of the week's pattern: the incumbent buys the AI decision layer rather than building it, and the earn-out structure prices the uncertainty about how fast the AI revenue converts. The finance scorecard row notes it; the Weekly's capital-flow table carries M&A in text only (grade 4).
Sources Guidewire IR
2026-10-30 · Commerce · 50+ enterprises (company-reported)
Sierra (Poppy open standard)
A standard led by the agent vendor with the platform and payments incumbents as members is a bid to own the protocol layer of agentic commerce before Amazon's Seller Assistant or a retailer-led protocol does; the commerce scorecard row carries it as the challenger to Amazon and notes the overlap with W43's Personal Agent Protocol partner list. The membership count is Sierra's and the standard's specification is not in the packet.
Sources Sierra (company blog)
Pricing shifts
2026-10-28 · Seat Based → Hybrid
Microsoft
Paid M365 Copilot seat count dropped from disclosure; growth reported through daily users, usage-billed customers (+75% QoQ) and registered agents, with Q2 M365 commercial cloud growth guided 'primarily' from ARPU as seat growth slows to 5%. The seat still exists (E7 at 'millions of paid seats') but the vendor now reports the meter.
Sources Microsoft Investor Relations; FY27 Q1 call transcript
2026-10-02 · Seat Based → Hybrid
Zendesk
Employee-service agents priced as a per-agent seat with 10 (Growth, Professional) or 15 (Enterprise, Enterprise Plus) included automated resolutions per month, a $0.99 overage per resolution, and a hard stop: the agent pauses when the allowance is exhausted until more resolutions are purchased. The pricing was published on October 2 (W40) and is dated to that announcement; the October 26 GA release confirmed it unchanged, with GA rolling November 3 to December 1.
Sources Zendesk Help (GA announcement)
2026-10-28 · Seat Based → Hybrid
ServiceNow
Management frames a seat-plus-consumption model in which token economics can lift revenue per user about fivefold (CEO), launches Flow as a self-serve consumption-billed service desk, and reclassifies 25-50 bps of subscription revenue to professional services for forward-deployed AI deployment; agentic consumption in production is up 13x year to date.
Sources ServiceNow Newsroom (Q3 results)
Vertical scorecard
As of 2026-10-31
| Vertical | Leader | Challenger | Read |
|---|---|---|---|
| Commerce | Amazon Seller Assistant | Sierra-led Poppy standard partners (Meta, Shopify, Stripe, Walmart) | Held. The challenger is restated as the Poppy open standard Sierra disclosed this week with 50-plus enterprises, which overlaps W43's Personal Agent Protocol retail partner list; the specification is not public and the membership count is Sierra's. Amazon's Q3 did not update Seller Assistant; Bedrock Managed Agents (OpenAI-powered) in preview is the hyperscaler-side event. |
| Engineering | Codex (on GPT-6.1 Terra and Sol) | Claude Code (on Opus 5.5) | Held on distribution: Terra ships to Codex on day one, ServiceNow is rolling Codex to all employees and into AI Workflow Factory in 2027, and AWS Continuum embeds in Codex at GA. The challenger's model moves to Opus 5.5, which leads Artificial Analysis's Coding Agent Index with Claude Code at 61.7 (The Model Pulse's coding row), and Anthropic's S-1 shows Claude Code's top 5% of users spending about 20x what Enterprise users without Claude Code spend (two populations, not an average), the strongest monetisation datapoint either product has published. Agents now open more than a third of GitHub pull requests (Microsoft). |
| Healthcare | Abridge | Suki | Held; no healthcare package shipped in the window. Anthropic's S-1 lists Healthcare and Life Sciences among its seven vertical solutions, which are the frontier-lab entrants the row watches; neither has a disclosed customer count. |
| Support | Sierra | Salesforce Agentforce | Held and widened. Sierra's $400M-plus ARR, $150M added in Q3 and almost half the Fortune 50 (all company-reported) is the largest outcome-priced scale disclosed; Agentforce shipped the control plane (Model Wallet, Kill Switch, Agent Evals) rather than adoption numbers. Zendesk's $0.99-per-resolution GA and ServiceNow's Flow enter the row's price field, which now runs $0.50 (HubSpot) to $2.00 (Salesforce Help Agent) per resolution. |
| Sales | ZoomInfo | Salesforce Agentforce with Docusign IAM | Held; no sales-specific package in the window. ServiceNow reported CRM AI net-new ACV tripling quarter on quarter and year on year, which puts a third incumbent in the row's field without a product event to place it. |
| Security | Codex Security Cloud | AWS Continuum | Challenger changes. Continuum reached GA with Capital One, IBM, Infor and Robinhood named and is being embedded into Claude Code, Codex and Kiro, which outranks Claude Security (public beta) on both availability and distribution. The vertical is where the frontier arrives first this week (Gemini 4 Carbon to vetted defenders only; The Model Pulse) and where the ServiceNow-OpenAI cyber solution (early access January 2027) will enter next. |
| Legal | Harvey | Legora | Held. Two frontier-lab legal specialisations entered the field in one week (Gemini Enterprise's legal agent; Anthropic's Legal vertical solution) without customer disclosures; the row changes when either names a law firm or a legal department at scale. |
| Research | Claude Dashboards | Gemini agent | Held, with a third entrant: Amazon Quick's expansion into Outlook, Word, PowerPoint and Excel with seven named customers (Principal, GoDaddy, Autodesk, BMW, 3M, Southwest, Mondelez) puts AWS's work companion inside Microsoft's applications. Alphabet's 100M-plus knowledge-work seats is the distribution figure the Gemini agent rides. |
| Finance | SAP Joule | Workday agentic finance | Held. The week's finance events were adjacent to the row: Guidewire's $1.7B acquisition of hyperexponential (agentic underwriting), Gemini Enterprise's financial-services specialisation, and Anthropic's Financial Services and Financial Advisors packages. None displaces the ERP incumbents on finance-function agents yet. |
Architecture watch
The meter replaces the seat as the reported unit, and the vendor stops showing the denominator
Four vendors reported the same quarter in the same new unit (UC Today drew the seats-to-meters frame for Microsoft on October 30; this publication extends it across the four). Consumption is the honest metric because it ties revenue to work done, and it is the riskier one because it imports utilisation and seasonality into businesses that now carry fixed take-or-pay compute: Anthropic's S-1 seasonality risk factor records customer spend falling about 20% in the two weeks ending January 2, 2026 against the two weeks ending December 19, 2025, and ServiceNow lost 17% after hours on a guide that one reading, the house's, takes as the market pricing the meter's volatility (a whole-company cRPO guide, contracted remaining performance obligations due within twelve months, not an AI-consumption print, is the other). For a buyer the operating consequence is that spend concentration, not seat count, sets the bill (Claude Code's top 5% at ~20x what Enterprise users without Claude Code spend), and a renewal negotiated on seats is negotiated on a variable the vendor has stopped reporting.
- Examples
- Microsoft (no paid Copilot seat count; usage-billed customers +75% QoQ), Anthropic S-1 (consumption 80% of revenue; ~900% net revenue retention), ServiceNow (seat-plus-consumption; ~5x revenue per user claim; Flow), Alphabet (100M+ seats and 2,600+ customers past 100B tokens)
Sources Microsoft Investor Relations; SEC EDGAR (Anthropic S-1); ServiceNow Newsroom
Hard budget enforcement moves into the product: caps, cliffs and kill switches
Every price list published this week carries an enforcement mechanism, not just a rate: Salesforce's wallet stops the agent at a budget, Zendesk's agent stops at an allowance, Google's agent routes by cost under a cap, and Anthropic's harness enforces pooled budgets across a group (Agent Techniques carries the harness controls). The pattern is a response to the meter: once revenue is consumption, the vendor needs the buyer to trust the ceiling, and the buyer needs the ceiling to be a product feature rather than a finance process. The 'up to 75%' token reduction Salesforce reports from its optimisation policies is a vendor test figure and should be measured on the buyer's own traffic before it is budgeted.
- Examples
- Salesforce Model Wallet hard limits and Agent Kill Switch (GA), Zendesk (agent pauses at 10 or 15 resolutions until more are bought at $0.99), Gemini Enterprise spend caps and Smart Routing, Claude Code pooled group budgets (Agent Techniques)
Sources Salesforce Newsroom; Zendesk Help; Alphabet call transcript
Every incumbent is a headless back end and a front end at once, and the toll is unpriced
Salesforce exposes its agents to Anthropic's surfaces, ServiceNow exposes itself to OpenAI's, Google positions Gemini Enterprise as the front end other incumbents integrate against, and Amazon puts its companion inside Microsoft's documents. Each is trying to be the surface the user talks to while accepting that it will often be the service another agent calls. The economics of that posture rest on one unpublished number: the Flex Credit multiplier Salesforce will charge for a Headless Platform Interaction, promised with 30 days' notice. Until it exists, an architect wiring cross-platform agents is accruing a bill at an unknown rate, and the synthesis in The AI Stack Weekly treats the multiplier as the first test of whether a platform incumbent can toll agent traffic it does not originate.
- Examples
- Salesforce Agentforce via hosted MCP to Claude, Slack and Claude Code (HPI multiplier TBA), ServiceNow MCP plugin in OpenAI's marketplace; Codex in AI Workflow Factory, Gemini Enterprise as a headless front end for ServiceNow and Workday, Amazon Quick inside Outlook, Word, PowerPoint and Excel
Sources Salesforce Developers Blog; ServiceNow Newsroom; About Amazon
Forward-deployed engineering becomes a revenue line, and vertical packages become the lab's product
Three ways to close the gap between a model and a deployed workflow appeared in one week: the incumbent books deployment as professional services and tells investors so (ServiceNow), the startup productises it as a named tool and a time-to-live claim (Sierra), and the frontier lab packages it as vertical solutions listed in a securities filing (Anthropic), while a vertical-SaaS incumbent buys the layer outright (Guidewire). The common thread is that the last mile is now priced and reported rather than given away, which is what happens when the meter, not the seat, carries the revenue: the vendor's incentive is to get the agent to production fast, because production is where consumption starts.
- Examples
- ServiceNow (25-50 bps of subscription reclassified to AI deployment services), Sierra Ghostwriter (a $40B retailer live in 27 days), Anthropic (seven vertical solutions; 250 launches in nine months), Guidewire-hyperexponential ($1.7B for an agentic-underwriting layer)
Sources ServiceNow Newsroom; Sierra; SEC EDGAR (Anthropic S-1); Guidewire IR
Watchlist
November 3 to December 1
Zendesk employee-service agents GA rollout
The first resolution-metered employee-service product with a hard pause at the allowance reaches general availability; the first customer reports of agents pausing mid-month, or of overage bills, will show whether a $0.99 cliff is a budget control or a churn driver.
30 days before metering begins
Salesforce Headless Platform Interaction Flex Credit multiplier
Still 'TBA' after two weeks of shipping the pattern; the multiplier is the price of cross-platform agent traffic and the Weekly's synthesis treats it as the test of whether platform incumbents can toll agents they do not originate.
Week of November 9-20
Anthropic roadshow and any amended S-1 customer metrics
The 11,900-plus $100K customers, ~900% net revenue retention and 20x Claude Code spend concentration are as of September 30; an amended filing or roadshow disclosure updating them to October would be the first in-flight read on whether consumption held after the Index v5 and Terra price moves.
Early December
Microsoft voluntary-exit decisions (exits December 31)
About 60,000 employees are eligible with data-centre staff excluded; the acceptance rate (the April program saw more than 30% of ~8,750) is the measure of how much headcount Microsoft trades for compute inside the quarter it stopped counting Copilot seats.
January 2027
Salesforce Agent Evals GA and triggered Kill Switch; ServiceNow-OpenAI cyber early access
The verification layer of the control plane reaches GA and the ServiceNow-OpenAI security solution opens, a quarter behind AWS Continuum; the security scorecard row re-ranks on named customers.
December quarter
Sierra margin or outcome-mix disclosure
$400M of ARR with no gross margin, retention or outcome-revenue mix disclosed leaves the outcome-priced category's unit economics unverified; any disclosure, or a competitor's (Decagon), is the datapoint that decides whether 'pay for outcomes' survives a metered model bill.
Changelog
- Scorecard: security challenger moves from Claude Security to AWS Continuum on GA with named production customers and embedding in three coding agents; the engineering challenger's model is restated to Opus 5.5; the commerce challenger is restated as the Poppy standard partners with the overlap to W43's Personal Agent Protocol list noted. All other rows held, with the week's adjacent events recorded in the notes.
- Pricing shifts: three recorded (Microsoft, Zendesk, ServiceNow), each from seat_based to hybrid, on the evidence that the seat persists while the vendor reports or enforces the meter; the Zendesk shift is dated to the October 2 pricing announcement, with the October 26 GA release noted as confirming it. Each incumbent entry closes with the pricing model it is defending, since the schema carries no tag for it. Credits and attributions: UC Today for the seat-count observation; TechCrunch and Models at Work for the 20x Claude Code reading, which is stated as a comparison between the top 5% of Claude Code users and Enterprise users without Claude Code rather than against an average; Bloomberg for the headcount-for-compute framing of the Microsoft buyouts, with the buyout figures sourced to CNBC (a person familiar) and the exclusion list to TechCrunch via Bloomberg, CNBC adding engineering; Amazon's own Q3 release and earnings page for Quick's named customers. The Anthropic December spend figure is attributed directly to the S-1's seasonality risk factor with its window stated and described as customer spend; the 'first time' claim is narrowed to Microsoft's own Copilot reporting; the ServiceNow -17% reading is presented as one reading rather than the fact; Zendesk's $0.99 is identified as the employee-service overage against its $1.50 / $2.00 customer-service list; Gemini Enterprise's 90% Auto-mode figure is Google's. Vendor-tested figures (Salesforce's 'up to 75%' token reduction, Zendesk's 6 million executions and 500,000 hours, Sierra's ARR and customer counts, OpenAI's cost-per-task comparison) are labelled at every use.
- Grades: Microsoft's buyout program (CNBC, TechCrunch citing Bloomberg) and Sierra's disclosure are carried at grade 4; Guidewire-hyperexponential at grade 4 from the acquirer's release. No grade-3 claim enters the scorecard.
- Where each fact lives: this issue owns the S-1's customer metrics, Microsoft's seat-to-meter change, Zendesk's and Salesforce's pricing and controls, ServiceNow's print and partnership, Alphabet's seat and token figures, Amazon Quick and Continuum, and Sierra. The S-1's financials and compute commitments, the hyperscaler capex and the Anthropic lease structures live in The AI Stack Weekly; Terra's model specification and Index v5 live in The Model Pulse; Codex CLI releases, Claude Code budget controls, AgentCore adoption and the Strands harness study live in Agent Techniques. Each appears here as one clause and a pointer.