ACIS ResearchAI Civilization Investment Research中文

AI SOFTWARE & ENTERPRISE AI WEEKLY · 008 · 2026.09.29

Microsoft Puts Agents on Usage-Based Billing as Enterprise AI Moves Beyond Seats

AI Software Weekly | Issue 008 | 29 September 2026

This week's most important change is not a smarter agent. It is an agent with its own runtime, identity, permissions, cost and billing unit. The monetization architecture is becoming clearer; revenue, margin and free-cash-flow proof still await the next reporting cycle.

Agent RuntimeUsage-Based BillingFinOps

Public research | Full access

Companies covered: Microsoft (Nasdaq: MSFT) | Amazon (Nasdaq: AMZN) | Alphabet (Nasdaq: GOOG/GOOGL) | Salesforce (NYSE: CRM) | ServiceNow (NYSE: NOW) | Palantir (Nasdaq: PLTR) | Meta (Nasdaq: META)

THE 10-SECOND VIEW

Microsoft placed long-running capabilities including Cowork, Code and Autopilot on usage-based billing and brought Managed Runtime into cost management. Enterprise AI is expanding its unit of value from user seats to agent work and resource consumption, but that is not yet financial proof.

The one-minute brief

On 25 September Microsoft announced a new Copilot: Home and Code are scheduled to begin Frontier rollout in the coming weeks, Autopilot is set to expand in private preview at month-end, and Managed Runtime entered public preview. Microsoft said long-running capabilities including Cowork, Code and Autopilot will use usage-based billing, while FinOps for AI brings model, runtime and agent spending into enterprise budget control. In announcing AgentCore Runtime enhancements, AWS said AgentCore had been used by thousands of teams for production agents; its multi-account Gateway and MCP reference architecture shows how identity, authorization and data isolation enter deployment. Google is pushing multi-step agentic work into browsers and endpoints. Meta formally launched Enterprise Platform, bringing models, agents, infrastructure and enterprise applications into one competitive boundary. ACIS's update is precise: monetization architecture has strengthened, but no new full-quarter result proves that revenue, gross margin and free cash flow are accelerating together.

Browser voice · tap to play
Audio transcript

This week's most important enterprise AI change is not a smarter agent. It is an agent becoming a separately operated, metered and budgeted unit of work. Microsoft announced Home, Code and Autopilot and put long-running capabilities including Cowork, Code and Autopilot on usage-based billing; some capabilities remain scheduled for rollout or in preview. FinOps for AI brings model, runtime and agent costs into enterprise management. In announcing AgentCore Runtime enhancements, AWS said AgentCore had been used by thousands of teams for production agents. Google is pushing multi-step work into browsers and endpoints, while Meta formally launched Enterprise Platform. Competition is shifting from AI features toward runtime, permissions, context and platform control. The boundary matters: product announcements are not revenue recognition, and usage is not gross profit. ACIS's update is that monetization architecture has strengthened while financial proof remains unchanged. The next tests are paid production deployments, persistent task use, contract and revenue growth, incremental gross margin, and free cash flow after capital spending.

Usage-based billing

Microsoft puts long-running agent capabilities on UBB

Production runtime

AWS says thousands of teams run production agents

Financial proof unchanged

No new full-quarter revenue and cash-flow evidence

01 · CONTINUITY

Public view comparison | 22 → 29 September 2026

Prior view

Enterprise AI was expanding from selling seats to selling work; Google and Salesforce supplied runtime, state and work-unit signals.

New evidence

Microsoft puts long-running agents on UBB and FinOps; AWS deepens runtime, identity and cross-account governance; Meta enters enterprise platforms.

Updated view

The charging unit and control layer are clearer, strengthening commercialization direction; revenue, margin and free cash flow remain the proof gates.

This is a thesis-strengthening update, not a financial-proof upgrade.

02 · EVIDENCE MAP

Four new signals and two continuity checks

What this week's evidence proves—and does not prove
CompanyConfirmed evidenceACIS readStill unproven
MicrosoftThe new Copilot was announced on 25 September; Home/Code are due to begin rollout in the coming weeks, Autopilot to expand in private preview at month-end, and Managed Runtime entered public preview; Cowork, Code and Autopilot use UBB.The agent is becoming a persistent, separately metered and budgeted work unit rather than only a feature inside a seat.An announcement and billing architecture do not disclose revenue, gross margin or free cash flow.
AWSAWS announced AgentCore Runtime enhancements on 18 September, said AgentCore had been used by thousands of teams for production agents, and published a multi-account Gateway and MCP reference architecture on 24 September.Runtime, identity, gateway, authorization and observability are becoming enterprise-agent infrastructure.“Thousands of teams” is company-reported; customer revenue, utilization and incremental profit were not disclosed.
GoogleGoogle Intelligent Endpoints puts multi-step workflows, enterprise Skills, browser security and data-loss prevention into one endpoint system.Enterprise agents are moving from cloud chat windows into browsers, devices and existing applications.This week's material provides no standalone incremental revenue, paying-customer or profit disclosure.
MetaMeta Enterprise Platform will bring Muse, Business Agent, Muse API and Muse Code to enterprises and developers under a new platform leader.Model and consumer-platform companies continue moving down into enterprise applications.Strategic entry does not prove customer adoption, pricing, distribution or enterprise revenue.
SalesforceAgentic Work Units remain a framework for measuring completed work rather than tokens alone.Task usage is a better starting point for connecting product activity to business outcomes.The main evidence was included last issue; AWUs still need to connect to billing, margin and cash flow.
ServiceNowOn 23 September ServiceNow said INRY deployed EmployeeWorks in six weeks; that customer reported a 60% reduction in median self-service submission time and a 97% request-completion rate.This is a quantified customer case for workflow and an AI front door in production.One customer case does not replace company-wide AI ACV, revenue, margin and renewal data.
PalantirAt this issue's cutoff, the latest quarterly baseline remained the Q2 2026 business update published on 3 August.Palantir remains a comparison point for production AI and commercial conversion.This is a dated financial baseline, not new weekly operating evidence or a valuation or sizing conclusion.

03 · INDUSTRY ANALYSIS

From seat subscriptions to metered work

01 | The charging unit is changing

Traditional SaaS is largely Users × Seats × Annual Subscription. The agent era adds Agents × Tasks × Sessions × Runtime × Compute × Storage × Transactions. Microsoft UBB is this week's clearest mainstream-platform signal: long-running capabilities that remain in rollout or preview have already been placed inside a measurable, authorized and budgeted commercial architecture. The investment question is not whether usage rises, but whether new usage creates high-quality net revenue and gross profit.

02 | Runtime is becoming a product layer

When agents move from seconds-long chat turns to hours-long or ambient work, enterprises need isolation, durable state, identity, permissions, recovery, audit and cost controls. AWS AgentCore and Microsoft Managed Runtime reinforce that direction. A larger runtime layer can expand cloud consumption, but it also makes competition and cost more visible; platforms still need to prove unit economics.

03 | Enterprise context and governance remain the moat

As model capability diffuses, the harder assets to copy are enterprise data, permissions, customer records, business rules and cross-system workflows. Salesforce, ServiceNow and Palantir remain strongest in context, workflow and governance, while Google and Microsoft add distribution, identity and productivity entry points. Winners must combine these elements into systems that can be deployed, audited and recovered.

04 | Meta turns feature competition into platform control

Meta Enterprise Platform combines models, agents, APIs, coding tools and enterprise products as a new business pillar. It will not immediately erase incumbent moats, but it raises pressure on middle-layer software with an AI interface and little proprietary context or workflow control. Defensibility will depend increasingly on context, process depth, distribution and cash generation—not model wrapping.

05 | Monetization architecture strengthened; financial proof unchanged

This week's evidence answers how agents may be charged, where they run and how they are governed. It does not answer how much revenue grew, whether margins improved or whether cash rose after capital spending. ACIS therefore raises confidence in the monetization architecture, not the financial conclusion. Product release is still not revenue recognition.

04 · AGENT ECONOMY STACK

Eight layers that must work together

Agent-economy stack and proof points
LayerRoleProof point
ModelReasoning capability and model choiceQuality, latency and model cost
Runtime / SandboxLong-running, isolated executionDuration, reliability and margin
State / Memory / StorageFiles, memory and task continuityRetention, storage revenue and cost
Enterprise ContextCompany data and business rulesPermission coverage and production use
Workflow / GovernanceApprovals, audit, rollback and accountabilityDeployment depth and expansion
Observability / SecurityCalls, errors, cost and leakageIncidents, SLA and attach revenue
Identity / PermissionsWho can authorize which actionsLeast privilege and accountability
Billing / FinOpsMetering tasks, models and resourcesTake rate, margin and cash flow

05 · INVESTMENT IMPLICATIONS

Where can durable value accrue?

Cloud platforms

Runtime, storage, networking and model calls can expand consumption, but capital intensity and price competition determine returns.

Enterprise applications

Companies with context, workflow and distribution are better positioned to convert task use into retention, expansion and high-quality revenue.

Control layer

Identity, security, observability and FinOps move from back-office capabilities to requirements for production adoption.

Investment judgment

A company can be right on product direction without its stock offering a mispriced opportunity; valuation and cash conversion remain separate tests.

06 · NEXT 90 DAYS

What would strengthen—or weaken—the view?

Commercialization checks for the next 90 days
CompanyNext proofWarning signal
MicrosoftUBB adoption, Copilot Runtime consumption, budget controls and Azure / Copilot revenue conversion.Capex, depreciation and inference cost outgrow incremental revenue.
AWSAgentCore production customers, runtime utilization, cloud consumption and AWS margin.Usage rises while price competition erodes unit returns.
GoogleEndpoint workflows, Gemini Enterprise production customers, actual billing and Cloud backlog.Use remains in testing or security cost rises.
SalesforceAWUs, Agentforce revenue, renewals, expansion, cRPO and incremental margin.Task usage grows but only replaces seat revenue.
ServiceNow / PalantirProduction deployments, contract conversion, commercial growth and free-cash-flow quality.Valuation expands faster than durable cash conversion.
MetaFirst enterprise customers, pricing, channels, governance and enterprise revenue.The platform remains a product portfolio without production adoption.

07 · PRINCIPAL RISKS

What can interrupt cash conversion?

Valuation and discount rates

The right commercialization direction does not immunize high-multiple software from higher rates or lower expectations.

Seat cannibalization

Agent-usage revenue may replace rather than fully add to legacy seat revenue.

Cost creep

Long-running agents raise inference, storage, tool and support costs; gross margin may not rise with usage.

Platform encroachment

Model, cloud and consumer platforms continue moving into enterprise applications.

Governance and reliability

Permission errors, unintended actions, data leakage, outages and weak accountability can delay production adoption.

FAQ

Five questions readers should ask

Has enterprise AI moved from subscriptions to usage billing?

Not wholesale. Seat subscriptions remain, while long-running agents are adding new billing units based on tasks, sessions, runtime and resources.

Why is Microsoft's UBB an important signal?

It places Cowork, Code and Autopilot inside a measurable and budgetable commercial system, treating agents as distinct work units.

Does AWS's 'thousands of teams' prove revenue?

No. It is an AWS adoption measure that supports broader production use, but economics still require customer revenue, utilization, margin and cash flow.

Will Meta immediately reshape enterprise AI competition?

Not immediately, but it increases pressure on middle-layer software. The real test is customers, pricing, channels, governance and revenue.

Which metrics matter most now?

Paid production deployments, persistent task use, billing and contract growth, incremental gross margin, and free cash flow after capital spending.

KEY TERMS

Reader reference

Usage-based billing (UBB)
Charging for tasks, sessions, models or resources rather than only user seats.
Managed Runtime
A platform-managed environment for execution, isolation, scaling and recovery.
FinOps for AI
The discipline of linking AI budgets, costs and usage to business value.
Enterprise context
Company data, permissions, customer records and business rules that guide agent action.
MCP
A protocol through which models or agents discover and invoke external tools and data.
Observability
The ability to trace calls, cost, errors, latency and execution paths.
Incremental gross profit
New AI revenue after direct inference, storage, support and related costs.

SOURCES & CONTINUITY

Primary sources

Related ACIS research

This report covers AI software and enterprise AI and excludes AI biotech. It uses company materials available through 29 September 2026. AWS's 'thousands of teams' is company-reported; some Microsoft capabilities remain scheduled for rollout or in preview; Microsoft, Google and Meta product or platform announcements are not revenue recognition. Salesforce task measurement is continuity evidence and is not treated as new financial proof this week. No internal workflow, publication status or position information is used. For research and education only; not personalized investment advice, a security recommendation, an offer or a solicitation.