Salesforce's investment question is whether cash generated by its established applications can support an AI transition while the company takes on more debt to repurchase shares. The cash generation is observable. The effect of autonomous agents on future license demand is uncertain. Treating every new agent task as economic success would be premature; treating the transition as inevitable destruction is equally unsupported.
Revenue, contracted backlog and ARR answer different questions
Salesforce reported $41.5 billion of FY2026 revenue, up 10%, including $399 million from Informatica. Its current remaining performance obligation, or cRPO, rose 16% as reported and 13% in constant currency, including a four-percentage-point contribution from Informatica. Subtracting that contribution from the constant-currency figure yields approximately 9% growth excluding the acquisition; subtracting it from the reported figure yields approximately 12%. Those bases must not be mixed. FY2026 results and metric definitions.
cRPO is contracted revenue expected to be recognized within the next 12 months. It provides useful visibility but is not the same as current revenue, bookings or a complete measure of demand. Contract timing, duration and currency can move it. The 9% estimate is a reason to ask about underlying momentum; it does not prove that customer acquisition costs are rising exponentially or that individual product lines have permanently stopped growing.
Agentforce reached $800 million of annual recurring revenue. Dividing that by $41.5 billion gives about 1.9%, but this is a scale comparison between an annualized metric at a point in time and revenue recognized over a year. It is not Agentforce's disclosed share of FY2026 revenue. Likewise, the fact that over 60% of Agentforce and Data 360 Q4 bookings came from existing customer expansion can indicate cross-selling strength as well as dependence on the installed base. It does not by itself establish failure to win new customers. Salesforce's Agentforce disclosures.
AWUs measure activity, not proven customer value
The Agentic Work Unit is Salesforce's measure of a discrete task performed by an agent, such as processing a prompt or invoking a tool. That can be more informative than raw token consumption, but it does not show whether a customer issue was resolved correctly, how much human checking remained, or whether the customer earned a satisfactory return. Salesforce's AWU explanation.
The limitation is important without inventing a billing rule. AWUs were introduced as a performance metric; the announcement does not say that every failed reasoning iteration creates an AWU invoice. Agentforce has separate commercial models, including conversations, user subscriptions and Flex Credits. The May 2025 Flex Credits announcement priced 100,000 credits at $500. An investor cannot infer revenue or profit from an AWU total without knowing how activity maps to the actual customer contract.
A useful customer-value test would pair task counts with resolution accuracy, human escalation, total cost and renewal behavior. Repeated attempts could increase costs in some consumption arrangements, but the effect depends on which actions are billable and the applicable terms. The original claim that Salesforce deliberately bills fifty failed reasoning loops is not supported by the disclosed metric.
Seat cannibalization is a scenario to test
Agents could reduce demand for some human-user licenses if customers need fewer users to perform the same workload. They could also expand the amount of work handled, make existing users more productive or create additional subscription and consumption demand. The net outcome depends on adoption, prices, contract terms and customer staffing choices.
That makes the relevant question a revenue bridge: how much license revenue is lost, how much agent revenue is added, and what does serving that work cost? A hypothetical reduction from 5,000 customer representatives to 3,000 is not evidence that Salesforce has already lost 2,000 licenses. It is one downside scenario. The optimistic case needs the same scrutiny: rising activity does not guarantee profitable, durable demand.
The buyback adds real financial obligations
Salesforce priced $25 billion of senior notes in March 2026 with maturities from 2028 through 2066 and coupons ranging from 4.5% to 6.7%. It said the proceeds would fund an accelerated share repurchase. The ASR delivered approximately 103 million shares initially; the final amount would depend on the contractual settlement mechanism. Debt offering; ASR announcement.
This exchange increases fixed interest obligations while reducing the equity base. It deserves scrutiny because cash spent on a repurchase cannot simultaneously fund acquisitions, investment or debt reduction. But the transaction is not evidence that the company is merely offsetting employee awards. The initial share delivery is a real reduction, and the FY2026 weighted-average diluted share count had already fallen from 974 million to 956 million. Stock compensation is an economic cost; its accounting expense is not a count of shares issued during the same period.
Assess value creation by comparing repurchase prices with a defensible estimate of intrinsic value, then examine total debt, interest coverage, cash generation and net dilution. EPS accretion alone is insufficient. A comparison of the company's pretax coupon with an earnings yield also does not settle the decision: taxes, risk, growth and the durability of earnings matter.
Cash flow is substantial, but it is not costless
FY2026 free cash flow was $14.4 billion. That supports the argument that Salesforce has an established cash-producing business. It does not eliminate the need to reconcile GAAP and non-GAAP earnings, particularly stock compensation, acquired-intangible amortization and restructuring. These measures answer different questions and should be shown together. The $3.81 non-GAAP EPS figure belongs to Q4; FY2026 non-GAAP EPS was $12.52. Financial results and reconciliation.
Operating cash flow adds back noncash stock compensation, so cash generation does not make dilution disappear. Conversely, subtracting the same compensation cost and then separately charging all dilution-offsetting repurchases can double-count it in a valuation. The model needs one consistent treatment of the cost and resulting share count.
The strongest cautious thesis is therefore conditional: weaker growth would make a larger debt-funded repurchase harder to justify, especially if AI spending fails to create profitable demand. The countercase is that the installed customer base, real cash generation and effective agent products sustain value while the share count falls. Neither outcome can be established from executive share sales or activity metrics alone.
This is independent analysis of historical public disclosures, not a recommendation to buy or sell securities.
