Three of the four listed companies that sell software for running corporate customer relationships have now put a number on their artificial-intelligence business. All three also reported a lower gross margin than a year earlier. That is the whole argument in one line: an agent answering a customer question consumes computing power, that cost sits in cost of revenue, and none of these vendors has yet priced it back out.

The exception is the largest. Salesforce, whose Customer 360 suite plus Slack, Tableau and MuleSoft runs sales pipelines and service desks for 83,000-employee-scale enterprises, earned a gross margin of 76.92% in its April quarter against 76.96% a year before. It did that while disclosing $1.2bn of annualized Agentforce revenue, up 205%, and 28.6 trillion tokens processed.

What the month actually was

The group's roughly 23% equal-weight gain over 30 days is two things stacked. About half arrived in one session on 28 July, when money rotated out of semiconductors into oversold enterprise software — Salesforce up about 7%, ServiceNow 8%, Workday 10%, on no company news. The rest is a grind: across the 19 sessions since, Salesforce rose 11 times and fell 8, with no gain above 5.1%. On market value the group is Salesforce — $170.5bn against $12.3bn, $5.9bn and $3.5bn — so its 27.7% month effectively is the group's. The backdrop is a valuation unwind, not new demand: the Morningstar US Software Application Index fell about 27% from October to mid-July while the broad market rose, and roughly $2 trillion of software value has been removed on the argument that agents do the work seat licences pay for.

The seat, the meter and the outcome

Salesforce is trying to graft consumption onto a subscription base and is running three prices at once for the same product: $2 per conversation at launch, Flex Credits at $500 per 100,000 credits, and a flat per-user digital-labour licence from $125 a month. That is a company that has not settled on a billing unit. The underlying business is nonetheless accelerating — 8.6%, then 12.1%, then 13.3% revenue growth — with operating margin at 21.80% from 19.76% and current remaining performance obligation of $33.6bn, up 14%. Agentforce is still only about 3% of a $41.5bn revenue base, so the flat gross margin proves less than it looks: inference cost has not landed because the volume has not arrived.

HubSpot, which sells an all-in-one marketing, sales and service platform to mid-market companies on seats and contact tiers, is the contradiction. It added 7,000 net customers in the June quarter against 9,000-10,000 expected and guided the second half down to 5,000-6,000 a quarter. Net revenue retention was 102%, down a point, with average subscription revenue per customer of $11,800 — up only 2% in constant currency. Gross margin fell 159 basis points. Its agents are being priced on outcomes — $0.50 per resolved conversation, $1 per qualified lead — and adoption is far ahead of billing. The shares fell 18% on 6 August, the worst session in the company's history, and had recovered the entire loss by 19 August.

Braze, which orchestrates cross-channel messaging for consumer apps and bills on messages and data points rather than seats, is the meter in its purest form. Revenue grew 30.2% in a fourth straight quarter of acceleration, with retention at 110% and customers above $500,000 of annual revenue up 33%. Gross profit grew only 24.7%: the meter fills at 65.72% gross margin against 68.62%. Its own guidance implies about 22% growth next quarter.

NICE, which sells the CXone contact-center platform and Actimize financial-crime software and is not a customer-relationship vendor at all, is the control that decides the question. It has the most AI attach of any of them — $362m of AI and self-service annualized revenue, up 52%, now 15% of cloud — and it is the most damaged by it: revenue grew 9.5%, gross profit 5.1%, and GAAP operating income fell 31.5%. It reported a beat on 5 August and the shares fell 6% in two days.

What the prices assume

On price per dollar of trailing gross profit — the lens that normalizes a 77%/82%/66%/64% margin spread — Salesforce is at 5.13x trailing and 4.76x forward, above 4.62x three months ago but below 5.68x in February, on a 8.6% free-cash-flow yield and 14.70x forward earnings against 24.01x trailing. Consensus still models fiscal 2028 revenue growth of 9.4%, slower than the company just delivered; JPMorgan initiated at Overweight on 13 August with a $250 target. HubSpot at 4.28x has expanded 13% since May in the window its guidance came down, yet sits 8.5% below February. Braze is the stretch: 3.94x in February, 5.23x in May, 6.72x today — a 71% re-rating on no new disclosure since 27 May, at a 1.9% free-cash-flow yield. NICE, the cheapest at 2.89x and 9.03x forward earnings, is the weakest performer.

The durable risk sits behind all of it. Klarna publicly abandoned Salesforce for an assembled AI stack, the second such defection after Workday, and rival platforms are now indexing Salesforce and Slack data into their own cross-application context graphs — Atlassian's latest shareholder letter treats Salesforce data as an ingested source rather than a partner. Coordination value migrates to whoever holds the graph, not whoever holds the record.

The setup

Where it stands — Every vendor here that monetizes AI has lost gross margin doing it; the one that hasn't has barely started. Would confirm — Salesforce reporting Q2 constant-currency cRPO growth at or above the 14% Q1 pace with gross margin near 77%. Would invalidate — Salesforce gross margin falling below 76% as Agentforce volume scales, matching the pattern at NICE and Braze. Watch next — Salesforce fiscal Q2 on 26 August 2026; Braze fiscal Q2 in early September. Valuation — Salesforce 5.13x trailing and 4.76x forward gross profit, versus 5.68x in February; Braze 6.72x, versus 3.94x.