Agentic work units: the token count Salesforce dropped

AI providers bill by the token, the unit a language model reads and writes, about four characters of English text. In December 2025 Salesforce began reporting how many tokens its AI agents had consumed, and its chief executive, Marc Benioff, told analysts that the company was “going to start talking about that concept.” The count appeared in three releases in a row. In August 2026 it was gone, and a newer metric, the agentic work unit, stood alone. The token count was the metric that could be turned into energy and carbon, and Salesforce’s own earnings releases hold enough to rebuild it.

Three releases with a token count

Salesforce’s fiscal year ends on 31 January, so its fiscal 2027 began in February 2026. The token counts and work unit totals below come from its quarterly results, which it furnishes to the US Securities and Exchange Commission as an exhibit to Form 8-K.

The token counts are running totals. The December 2025 total of 3.2 trillion tokens covers only Agentforce’s traffic through Salesforce’s LLM gateway, while the totals from February 2026 cover all of Salesforce, so most of the jump to 19 trillion tokens is that change of scope. Beside each total Salesforce printed its growth in the quarter itself, and that growth rate is what lets each quarter’s tokens be worked out.

Salesforce’s executives promised more token counts on the earnings calls. In May 2026 Benioff called the totals “record token counts” and told analysts, “you’re going to see these token numbers continue to expand and grow.” The August release, its call, its investor presentation and the investor day deck of 16 September 2026 contain no token count. Salesforce’s own page on the new metric, last modified on 30 June 2026, still says the company tracks tokens processed as one of two key figures.

What an agentic work unit counts

Salesforce introduced the agentic work unit on 25 February 2026, beside the token count, as an additional metric. Its definition, printed in its results, is a measure of discrete tasks executed by AI agents in production across the Salesforce platform, including Agentforce and Slack, such as resolving customer cases, updating records or triggering automated workflows. Its launch post says the unit is the moment raw intelligence is converted into real work, and that tokens only measure how much an AI talks.

Salesforce made the case for the new metric on that February call. “The value isn’t in the token,” Benioff said. Patrick Stokes, Salesforce’s president and chief marketing officer, who devised the unit, described watching the ratio of tokens consumed to work produced, customer by customer. Benioff added that Stokes “was very unhappy that I keep bringing out this token number because I’m very impressed that we have 19 trillion tokens.” Salesforce’s launch post says the ratio between tokens and work units is “not a fixed ratio; it’s elastic,” and that it expects them to diverge as the platform does more work with fewer tokens.

The work unit measures activity, and it is a sensible thing for a platform to count. A work unit cannot be turned into energy without the tokens behind it, because the unit covers a record update and a long reasoning chain alike.

Rebuilding the missing quarter

Salesforce printed tokens and work units side by side for two quarters, which gives its own ratio of tokens to work units. From February to April 2026 it processed 9.6 trillion tokens (28.6 trillion tokens to date less 19 trillion tokens) and delivered 1.6 billion work units, so 6,000 tokens per work unit. For November 2025 to January 2026, working back from the 152% growth, it was about 3.81 trillion tokens over 771 million work units, about 4,900 tokens per unit. The totals to date give 7,526 tokens per unit.

Applied to the 3.2 billion work units of May to July 2026, the ratio of the latest quarter gives 19.2 trillion tokens. The ratio of the totals to date gives 24.1 trillion tokens. We lead with 19.2 trillion tokens because it compares one quarter with one quarter. Either estimate would have been a record, two to two and a half times the 9.6 trillion tokens of the quarter before.

The rebuild assumes the ratio held. Salesforce expects its ratio of tokens to work units to move: simple, repeatable tasks should use fewer tokens per unit, and complex reasoning more. The mix is shifting too, with Slack’s AI growing fastest. In the only two quarters where Salesforce printed tokens and work units together, though, tokens per unit rose, from about 4,900 to 6,000 tokens per unit, the opposite of the divergence Salesforce predicted.

The carbon of the missing quarter

The rebuilt quarter goes through Crbon Labs methodology v1.2, whose factors are all on the methodology page. It counts the electricity the chips draw, adds the host machines and idle capacity around them, and then the manufacture of the hardware and the training of the models. Agentforce’s reasoning runs by default on GPT-4.1 for new agents and GPT-4o for older ones, with Claude Haiku 4.5 and Gemini 3.5 Flash as options, and Benioff has said the lowest-cost model is the best one for Salesforce. Those are mid-sized models, so the medium tier is our default here, with the large and reasoning tiers as the upper cases. The count carries no split of input and output, so we use the methodology’s default for an aggregate, 75% input and 25% output, with nothing cached.

The quarter whose tokens Salesforce chose not to report comes to about 4,200 tonnes of CO2e on the medium tier, and to between about 3,000 and 9,300 tonnes of CO2e across the three tiers and the uncertainty band. The token count is what made an estimate like this possible for three quarters, and the work unit cannot replace it, because a work unit counts a task and not the computation behind it. Without the token count, nobody outside Salesforce can estimate the carbon of its AI.

On the ratio of the totals to date, the quarter comes to 5,316 tonnes of CO2e on the medium tier. Per work unit, 6,000 tokens on the medium tier is about 1.3 grams of CO2e, 2.2 grams on the large tier and 2.9 grams on the reasoning tier. We found no other published estimate of the carbon of an agentic work unit.

Where a year of Salesforce’s tokens sits in its inventory

A year at the pace of May to July 2026 is about 77 trillion tokens, and about 17,000 tonnes of CO2e on the medium tier. That is close to what the same tier gives for a single month of Meta’s employees’ tokens: 16,268 tonnes of CO2e for 73.7 trillion tokens, from reporting on a leaked memo.

Salesforce owns no data centres. Its AI runs on OpenAI’s and Microsoft’s, Amazon’s and Google’s infrastructure, so the tokens sit in its Scope 3, Category 1, purchased goods and services, which its fiscal 2026 inventory puts at 833,000 tonnes of CO2e on the location-based method and describes as including cloud computing services from third parties. A year of these tokens at the medium tier would be about 2% of that line. The inventory does not break AI out of it.

Who else reports tokens to the SEC

Salesforce was neither the first nor the only company to put a token count in a document it gave the SEC. Alphabet has printed a Gemini token rate in every earnings release since October 2025, five weeks before Salesforce’s first, and still does: 22 billion API tokens a minute in July 2026. Microsoft’s proxy statement of October 2025 says Azure AI Foundry processed over 500 trillion tokens in its fiscal 2025, and Sprinklr’s annual report of March 2026 says it processes over 9 trillion tokens a month through its own models. A few companies, among them the Chinese industrial supplier ZKH, have reported their own consumption.

Salesforce stood out for how prominently it used its token count. It put the token count in the headline or subhead of two releases, its chief executive promised more of it, and its website still says the company tracks tokens. Then the fourth release came out without a token count, and nothing in the release or on the call said why.

What a voluntary count is worth

Nothing required Salesforce to report tokens, and nothing required it to keep reporting them. On Salesforce’s own ratio of tokens to work units, the quarter it left out would have been a record of about 19 trillion tokens, so slowing growth does not explain the gap, and Salesforce has given no reason. A voluntary AI usage metric that its publisher can drop from one quarter to the next is useful while it lasts, but nobody else can build their reporting on it. Of everything Salesforce discloses about its AI, the token count was the only measure that could be turned into energy and carbon, and it is the one Salesforce stopped printing.

Salesforce has argued for this kind of disclosure. It endorsed a US bill in February 2024 that would have set up voluntary reporting of AI’s environmental impacts, helped launch the AI Energy Score in February 2025, and in June 2026 added energy and emissions figures to the model cards of the models it builds itself. The models behind most of Agentforce’s reasoning come from OpenAI, Anthropic and Google, and have no Salesforce model card.

What the missing quarter means for a company that uses Agentforce

Rebuilt from Salesforce’s own ratio of tokens to work units, the quarter it skipped comes to about 19 trillion tokens from May to July 2026, about 4,200 tonnes of CO2e on our medium tier, and about 1.3 grams of CO2e for each agentic work unit.

These emissions do not belong to Salesforce alone. For a company using Agentforce, the agents’ tokens are part of a purchased Salesforce service and sit in that company’s own Scope 3, Category 1, because Scope 3 counts the same emissions in more than one company’s inventory by design, as our guide to Scope 3 emissions from AI explains. Each buyer needs its own share, and a vendor’s disclosure that can vanish between one quarter and the next cannot be that share. Ask vendors for tokens by model and by month, and where you call models directly, keep the token counts that come back with every request, whatever a vendor decides to publish.

CrbonFree meters those token counts for the providers it connects and turns them into tonnes of CO2e, with the boundary and factor version stated, and the OpenRouter case study shows the same method on a public count of tokens that anyone can check.