PUBLIC RESEARCH · WEEKLY
Enterprise AI Is Shifting from Selling Seats to Selling Work, but Cash-Flow Proof Still Lags
AI Software Weekly | Issue 007 | 22 September 2026
The next layer of AI software value may extend beyond assistant seats into the compute, storage, enterprise context, workflow, governance and transaction services consumed whenever an agent performs work. This week's product evidence makes that path clearer, but launches are not revenue recognition. Investment value still depends on whether usage converts into net revenue, gross profit and free cash flow.
Listed companies covered: Microsoft (Nasdaq: MSFT) | Alphabet (Nasdaq: GOOG/GOOGL) | Amazon (Nasdaq: AMZN) | Salesforce (NYSE: CRM) | ServiceNow (NYSE: NOW) | Palantir (Nasdaq: PLTR)
The next layer of AI software value may extend beyond assistant seats into the compute, storage, enterprise context, workflow, governance and transaction services consumed whenever an agent performs work. This week's product evidence makes that path clearer, but launches are not revenue recognition. Investment value still depends on whether usage converts into net revenue, gross profit and free cash flow.
The one-minute brief
Enterprise AI is moving from giving employees an assistant to placing agents inside operating systems. The clearest incremental evidence this week came from Google Cloud and Salesforce. Google is building execution, isolation, suspend-and-resume and persistent state into a more complete infrastructure layer. Salesforce is measuring completed tasks through Agentic Work Units while continuing to disclose Agentforce recurring revenue and usage. The boundaries matter: Google has not disclosed independent Agent Substrate revenue or scaled production customers, and Salesforce broadened its Agentforce ARR definition this quarter. The latest reported results from Microsoft, AWS, ServiceNow and Palantir remain the commercialization baseline, not new earnings this week. The AI software thesis remains intact, but markets will increasingly separate companies with products, companies with usage and companies that retain cash.
Audio transcript
The key AI software development this week is not another list of agents. Companies are beginning to measure completed work and turn runtime, storage, permissions and governance into explicit products. Salesforce now reports both Agentforce revenue and task usage, while Google Cloud is making agent runtime and persistent state more concrete infrastructure. Yet a product launch is not revenue recognition, and usage is not profit. The latest results from Microsoft, AWS, ServiceNow and Palantir continue to support the enterprise AI commercialization thesis, but they are financial baselines rather than new weekly disclosures. ACIS concludes that enterprise AI is expanding from selling seats to selling work; stock returns still depend on incremental revenue, gross margin and free cash flow, and stronger fundamentals do not remove valuation risk.
Runtime, storage, permissions and governance become explicit products
Salesforce-reported Agentforce ARR; definition broadened this quarter
Usage must still pass through to revenue, gross profit and free cash flow
01 · RESEARCH PRIORITY
The view first: where should research go deeper?
| Company | Research priority | Why now | Proof required | First rejection risk |
|---|---|---|---|---|
| Salesforce | Deepen research | It discloses both Agentforce revenue and task usage, putting task-based economics closest to a measurable test. | Standalone revenue, renewals, expansion, cRPO, organic growth and incremental gross margin. | Broader definitions or AI revenue that merely replaces seat revenue. |
| Alphabet | Deepen research | Google Cloud is productizing agent runtime, sandboxing and persistent state as infrastructure. | Paying production customers, billing, backlog conversion and free cash flow after capital spending. | The product stays in evaluation or creates infrastructure cost without durable profit. |
| Microsoft | Commercial benchmark | Identity, office workflow, enterprise data and Azure form the broadest cross-distribution system. | Paid Copilot use, Azure AI revenue, RPO conversion, cloud margin and free cash flow. | Agent revenue remains undisclosed while capital spending and depreciation dilute returns. |
| ServiceNow | Continue validation | More capable agents increase the value of permissions, approvals, audit and rollback. | AI ACV converting into revenue while cRPO growth and margins hold. | AI expansion fails to offset seat pressure or implementation cost erodes margin. |
| Amazon | Watch cash conversion | Agent tasks can increase AWS consumption, but the software signal is diluted by group capital spending. | AgentCore and Bedrock use, AWS margins and free cash flow after capital expenditure. | Compute investment keeps outrunning revenue and price competition reduces returns. |
| Palantir | Commercial benchmark | valuation gated | Production adoption and commercial growth are strong, but a high-quality company is not automatically an attractive stock. | Customer broadening, commercial revenue, cash quality and per-share value after stock compensation. | Growth normalization, concentration or valuation compression. |
Research priority is not a buy, sell or sizing recommendation. It ranks the next questions worth answering.
02 · EVIDENCE MAP
What do this week's product signals and the financial baseline actually prove?
| Company | Confirmed evidence | ACIS read | Still unproven |
|---|---|---|---|
| Salesforce | Its latest results reported more than $1.5bn of Agentforce ARR, nearly $3.9bn for Agentforce plus Data 360, and 7bn cumulative Agentic Work Units. | The economic unit of enterprise software may extend from seats to completed tasks. | The ARR definition broadened; Work Units are neither revenue nor profit, and unit economics still need proof. |
| Google Cloud | On 15 September, Agent Substrate became available on GKE with sandboxing, suspend-and-resume and persistent state; production support remains allowlisted. | Agent runtime is becoming a purchasable infrastructure layer. | Independent revenue, paying-customer count and scaled production usage remain undisclosed. |
| Microsoft and AWS | The latest results continue to show strong cloud demand; Microsoft 365 Copilot exceeded 30m paid seats and AWS said its AI revenue run rate exceeded $25bn. | Enterprise AI already drives real cloud consumption, but these are financial baselines rather than new weekly disclosures. | Broad cloud data does not isolate agent revenue, margins or returns on invested capital. |
| ServiceNow | Its latest results reported more than $1bn in AI ACV and a ninefold increase in agentic deployments over nine months. | Workflow and governance already have commercial demand. | ServiceNow did not disclose production-only deployment counts or separately recognized AI revenue. |
| Palantir | Second-quarter revenue rose 93%, U.S. commercial revenue 149%, and adjusted free cash flow reached $1.22bn. | The financial data strengthen the case for AI inside production decision systems. | The figures do not prove every customer project is beyond a pilot, while stock returns remain valuation-sensitive. |
Product events from this week are separated from earlier earnings baselines.
03 · INDUSTRY ANALYSIS
From software seats to agent systems
01 | Seats Will Persist, but Billing Units Are Expanding
Enterprise software has traditionally charged by user, seat and annual subscription. Agents can add tasks, sessions, inference, compute, storage and transactions as billing units. That does not make seat revenue disappear. Investors need to compare incremental usage revenue with any seat cannibalization. The model improves only if the former exceeds the latter and covers inference and support costs.
02 | Value Is Spreading Across Seven System Layers
ACIS maps the enterprise-agent value chain across model, runtime, state and storage, enterprise context, workflow and governance, observability and security, and transactions. This is an analytical framework, not a standard accounting classification. It helps distinguish revenue from model calls, cloud consumption, workflow expansion, data services and transaction economics rather than treating all “agent” revenue as equal.
03 | Stronger Fundamentals Do Not Remove Valuation Risk
On 21 September, U.S. Treasury data put the 10-year yield near 4.96% and the 30-year yield near 5.29%. Higher discount rates reduce the present value of distant cash flows and punish execution misses. Industry commercialization and stock valuation require separate judgments: a product direction can be right while shareholder returns disappoint if cash is too distant, investment too heavy or expectations already too high.
04 · INVESTMENT IMPLICATIONS
Where can value accrue?
Platform and distribution
Microsoft, Alphabet and Amazon control cloud, identity, data and enterprise access. The research question is moving from whether demand exists to incremental revenue, capital efficiency, gross margin and free cash flow.
Workflow and enterprise context
Salesforce, ServiceNow and Palantir sit closer to operating workflows. Product momentum becomes shareholder value only when paid deployment, renewal, contract growth and cash conversion improve together.
Observability, security and storage
More agents create more calls, failures, permissions, costs and leakage risks. Monitoring, security and persistent storage may become picks-and-shovels layers, while facing bundled cloud features and price competition.
Company thesis versus stock thesis
Palantir and other firms can combine strong operating evidence with a demanding valuation. A commercial benchmark is not the same as an obvious expectations gap.
05 · PROOF STILL NEEDED
Five gates before usage becomes durable value
- Agent usage becomes repeatable revenue rather than temporary trials or marketing.
- Usage pricing covers inference, storage, support and security while improving margins.
- New task revenue exceeds any loss from seat cannibalization.
- Agents remain reliable, auditable and recoverable inside high-permission workflows.
- Independent software vendors defend their layer as model companies and cloud platforms move into applications.
06 · NEXT 90 DAYS
What would strengthen—or weaken—the thesis?
| Company or layer | Evidence needed | Warning signal |
|---|---|---|
| Salesforce | Standalone Agentforce revenue, paying customers, renewals, expansion, cRPO and incremental margin. | Definitions broaden while organic growth and profit fail to improve. |
| Google Cloud | Production customers, actual billing and storage attach revenue for Agent Substrate. | A prolonged evaluation phase without scaled production use. |
| Microsoft | Copilot usage, Azure AI revenue, contract conversion, cloud margin and free cash flow. | AI capital expenditure grows faster than revenue and cash returns. |
| AWS | AgentCore and Bedrock consumption, customer adoption, AWS margins and free cash flow. | Usage rises while competition and investment depress returns. |
| ServiceNow | AI ACV converts into revenue while cRPO growth and margins hold. | AI expansion fails to offset seat pressure or higher implementation costs. |
| Palantir | Customer broadening, U.S. commercial growth, cash quality and per-share value. | Rapid deceleration, rising concentration or valuation compression. |
07 · PRINCIPAL RISKS
What can interrupt cash conversion?
Valuation
A correct commercialization thesis does not make every valuation attractive.
Seat cannibalization
Agents may reduce traditional software seats, leaving the net revenue effect uncertain.
Margin pressure
Inference, storage and support can make AI revenue less profitable than traditional software.
Platform encroachment
Model providers and cloud platforms continue to move into applications, governance and development tools.
Reliability
Permission errors, hallucinations, unintended actions, outages and data leakage can slow production adoption.
Metric definitions
ARR, ACV, Work Units and deployment definitions differ and are not directly comparable.
FAQ
Key questions
Has AI software entered a broad rally?
The evidence does not support that conclusion. Commercialization continues, but high rates and elevated expectations reduce tolerance for misses; selective dispersion is more plausible.
Why do Agentic Work Units matter?
They are closer to completed tasks than tokens, but remain a company-defined usage metric that must still connect to pricing, gross profit and cash flow.
Why could stronger models help ServiceNow?
More capable models increase the need for permissions, approval, audit, rollback and accountability. ServiceNow's role is to govern how work is executed.
Why is Google's new product not automatically a revenue catalyst?
A launch proves product direction, not scaled adoption. Production customers, billing and profit contribution remain undisclosed.
Which metrics matter most now?
Track paid production deployment, durable usage, revenue or contract growth, gross margin and free cash flow. Launch counts are only early signals.
KEY TERMS
Reader reference
Agent runtime — The computing environment where an agent executes code, calls tools and completes tasks.
Persistent state and storage — Files, memory and task progress retained so an agent can resume work.
Enterprise context — Company data, permissions, customer records and business rules used for correct execution.
Workflow — The connected process across approvals, orders, refunds, service or other tasks.
Governance — Controls for permissions, audit, compliance, rollback and accountability.
Observability — The ability to trace agent calls, cost, errors, latency and execution paths.
ARR — Annual recurring revenue implied by the current subscription or contract base.
ACV — Annual contract value; it is not the same as revenue recognized in the current period.
RPO and cRPO — Contracted revenue not yet recognized; cRPO generally covers the portion expected within 12 months.
Free cash flow — Operating cash flow after capital expenditure.
SOURCES & CONTINUITY
Primary sources
Salesforce | Q2 FY2027 results | 26 August 2026 ↗
Salesforce | Koa reasoning model | 15 September 2026 ↗
Salesforce | AIforce product update | 15 September 2026 ↗
Google Cloud | Agent Substrate available on GKE | 15 September 2026 ↗
Alphabet | Second-quarter 2026 results ↗
Microsoft | FY2026 fourth-quarter results | 29 July 2026 ↗
ServiceNow | Second-quarter 2026 results | 22 July 2026 ↗
Amazon | Second-quarter 2026 results | 30 July 2026 ↗
Palantir | Second-quarter 2026 results and SEC filing | 3 August 2026 ↗
Related ACIS research
This report uses company releases, regulatory filings and U.S. Treasury data available through 22 September 2026. Product events this week are separated from earlier financial baselines. Company-defined usage measures, contract value and run rates are not current-period revenue or profit. For research and education only; not personalized investment advice, a security recommendation, an offer or a solicitation.