Software Seats Meet Salesforce's Agentic Work Units
On August 26, Salesforce put two kinds of software growth into the same earnings release. The company reported $11.3 billion in second-quarter revenue, an amount recognized through contracts and accounting rules. It also reported 7.0 billion cumulative Agentic Work Units, or AWUs, including 3.2 billion during the quarter.
One number describes money already earned across a large software company. The other describes work performed inside a newer agent platform.
Salesforce said quarterly AWUs rose 97% from the previous quarter. Agentforce annual recurring revenue exceeded $1.5 billion, up more than 240% year over year. Starting with this quarter, however, that ARR measure includes Slackbot and Headless 360. The faster growth and the wider scope arrived together.
A procurement team opening its renewal workbook this week therefore has several valid numbers and no automatic way to reconcile them. The user export, consumption report, and service dashboard do not share a denominator. Employees may still carry Salesforce or Slack seats. An Agentforce contract can add consumption charges, a coding agent connected to Slack can carry another license, and a support product can charge for a resolution. Prompts, reasoning, and tool use may all be recorded before the customer’s operation decides whether the result was correct, useful, or final.
The mismatch changes what buyers need to observe about work, not only how they read an invoice.
A software seat identifies a person entitled to use a product. An AWU identifies activity performed by an agent. A credit turns some activity into a bill. A technical completion says a system reached the end of a workflow. An accepted outcome says the business kept the result. Human review, correction, and escalation sit between those states, sometimes by design and sometimes because the agent failed.
All of those units can appear in one contract stack. Substitution is where the trouble starts.
Salesforce has made the distinction urgent by expanding the surfaces on which agents can operate. Headless 360 lets agents act across Salesforce without a person navigating a conventional screen. Slack Code turns a conversation into a shared coding session where people can inspect and redirect agent work. Agentforce Help Agent introduces a pay-per-resolution offer. The same vendor now presents activity, access, consumption, collaboration, and outcome units across one product family.
Buyers need a bridge across them before a large activity chart reaches a budget meeting. Otherwise, an agent can look busier while employees inherit more review. A bill can rise while accepted work stays flat. A vendor-defined resolution can hide a repeat contact. Connected records turn platform activity into operating evidence instead of a loose proxy for value.
Earnings day adds a new unit of work
The familiar accounting measures came first in Salesforce’s fiscal 2027 second-quarter release. Revenue grew 11% year over year. Subscription and support revenue reached $10.8 billion, up 12%. Current remaining performance obligation was $33.5 billion, while total remaining performance obligation reached $66.3 billion. AWUs appeared beside them as a new operating language.
Treat those company-wide measures as financial context. They cover products, contracts, services, and periods far beyond an individual agent task, so AWU growth cannot carry their attribution.
The income statement needs the same discipline. GAAP net income rose from $1.89 billion a year earlier to $3.53 billion, but the quarter included a $2.61 billion gain from strategic investments. Agent activity did not produce that investment gain. Operating margin supplies a cleaner view of the running company: 20.5% on a GAAP basis and 34.1% on a non-GAAP basis.
The agent figures answer a different set of questions. Salesforce reported Agentforce and Data 360 ARR of nearly $3.9 billion, up more than 210%. Agentforce ARR passed $1.5 billion. Cumulative AWUs reached 7.0 billion after 3.2 billion in the quarter. Slackbot users grew 150% from the previous quarter. Data 360 processed 104 trillion records, including 82 trillion through zero-copy access.
Each numerator belongs to its own denominator. Records processed describe data infrastructure. Slackbot users describe adoption of a product surface. ARR describes a recurring revenue run rate under Salesforce’s reporting definition. AWUs describe production activity. None, on its own, identifies accepted customer work or the human effort required to get there.
Scope matters as much as growth. Salesforce states that, effective in the second quarter, Agentforce ARR includes Slackbot and Headless 360. That makes the current figure useful as a measure of the broader agent business Salesforce now wants investors to follow. It makes a strict organic comparison with the earlier, narrower series harder. The release does not provide a bridge showing how much of the 240% increase came from like-for-like product growth and how much came from the expanded definition.
Product groups evolve, and investors need a measure that follows the business as it is sold. Preserving the definition and date lets a buyer use the current figure without pretending it is a like-for-like historical series.
Robin Washington, Salesforce’s president, chief financial and operating officer, now has to explain a company in which financial growth, human access, agent execution, and business outcomes may move at different speeds. The operating file beneath that explanation needs at least six layers.
| Evidence layer | Unit being counted | Useful question | Claim it cannot settle |
|---|---|---|---|
| Access | Named user, seat, plan, or entitlement | Who can use or supervise the product? | Whether the product is used well |
| Agent activity | Prompt, reasoning chain, tool invocation, or AWU | How much production work reached the platform? | Whether the work was accepted |
| Billing | Action, prompt, credit, pack, or contract metric | What created a charge? | Whether the charge created value |
| Technical completion | Finished workflow, response, update, or generated artifact | Did the system reach its defined endpoint? | Whether the endpoint was correct |
| Human work | Review, correction, approval, recovery, or escalation minute | What labor stayed around the agent? | Whether the remaining labor is excessive |
| Business result | Accepted case, order, qualified lead, deployed change, time, revenue, or cost | Did the operation keep the result? | Whether the result will repeat at scale |
The first three layers are usually easiest for a vendor to observe. The last three often live inside the customer’s service desk, code repository, finance system, quality process, and employee workflow. Renewal quality depends on joining both sides without allowing the supplier’s available metric to define the buyer’s objective.
Salesforce’s reporting creates a useful opening for that work. AWUs acknowledge that agent software performs tasks rather than waiting passively behind a login. The next step is to establish what an AWU contains and where it can stop.
An AWU can stop before acceptance
The smallest public example of an AWU is a processed prompt. The most consequential is closer to a resolved case. Salesforce’s Agentic Work Units page also lists a completed reasoning chain or an invoked tool in production; other Salesforce material uses examples such as updating a record and triggering a workflow.
Their consequences are far apart.
A prompt can ask for a summary. A reasoning chain can decide which system to consult. A tool invocation can retrieve a record. A record update can change customer data. A case resolution can close a service workflow. One customer job may require several of these events, while some events may be exploratory, repeated, blocked, or reversed. Salesforce does not say that each example has the same cost, difficulty, risk, or customer value.
The public definition also differs from the latest earnings snapshot. The AWU page currently displays 3.8 billion cumulative units and 1.6 billion in the first quarter, while the August 26 release reports 7.0 billion cumulative and 3.2 billion in the second quarter. That looks like a page updated on a different publishing cycle, not evidence that either figure is false. It does show why the date and source version belong beside any AWU chart.
Tokens do not solve the comparison. Salesforce explains that output can consume far more tokens than input and describes AWUs as a more legible unit for agent work. That is sensible for communicating platform activity. A long answer can use many tokens without changing a record, while a short tool call can alter an account. Yet replacing tokens with AWUs does not turn activity into a uniform job.
Consider a renewal-support agent. It receives a customer’s message, reads the account, retrieves contract terms, calculates an option, drafts an answer, asks for approval, sends the response, and records the interaction. That flow may contain multiple prompts, reasoning steps, and tool calls. If the retrieved contract was stale, the representative may reject the draft. The platform still did work. The customer received no accepted answer.
The opposite case matters too. A reliable agent may perform several invisible steps that spare an employee from opening five applications. Activity can be a valuable leading indicator. It helps a platform team see whether production use is spreading, whether capacity needs are changing, and whether a release caused unexpected repetition. Low activity can reveal a failed rollout before an outcome sample becomes statistically useful.
AWUs become misleading only when the next layer disappears. Three ratios preserve their operational value:
technical completion rate = completed agent workflows / initiated workflows
acceptance rate = results accepted without correction / technically completed workflows
repeat-work rate = corrected, reopened, reversed, or repeated workflows / accepted results
The exact definitions should change with the job. A customer-service acceptance may require the issue to stay closed for a defined period. A record update may require validation against a source system. A generated code change may require tests, review, merge, deployment, and no later rollback. A qualification may require a sales team to accept the lead under a documented rule.
These ratios also expose where human labor went. A high completion rate with low acceptance means the system is reaching the end of its own workflow while people discard the output. A high initial acceptance rate with many reopenings suggests the acceptance window is too short. Rising AWUs with stable outcomes may reflect deeper reasoning, broader use, retries, or wasted motion. The dashboard cannot choose among those explanations without customer-side evidence.
Planned review also differs from repair. Approval of a sensitive discount can be a control the company wants to keep. Correcting a hallucinated contract term is rework; escalating a threat or legal question is proper routing; repeating a tool call after an integration timeout is recovery. One “human in the loop” total conceals all four.
AWUs therefore fill a real measurement gap. Seat counts become less informative when an agent works across systems without a person opening each one. But the unit has to remain where Salesforce places it: platform-level production activity. Accepted work requires another record.
Headless 360 moves execution behind the screen
An IT operator no longer has to open the user-administration screen before an agent can attempt to freeze an account. That is the practical break Headless 360 makes with seat-era observation.
Salesforce’s August 19 product announcement describes more than 60 Model Context Protocol tools and more than 30 skills that can connect external agents with Salesforce work. The examples cover data, analytics, customer service, commerce, revenue operations, development, and administration.
The technical design is narrower and more useful than the large catalog number. Salesforce’s developer reference says the hosted MCP server presents four stable tools: Discover, Describe, Dispatch, and a read-only form of Dispatch. An agent can discover a growing set of operations, inspect the requirements for one, and invoke it through the stable interface.
That architecture avoids publishing a separate permanent MCP tool for every Salesforce operation. It also lets the available operation set expand without requiring every connected agent to learn a new top-level interface. The agent asks what exists, reads the schema, then acts.
Current examples include querying, creating, and updating records; creating or deactivating users; freezing access; assigning permission sets; deploying Apex; managing integrations; and working with commerce orders. These are consequential tasks. A wrong summary is inconvenient. A wrong permission, user-state change, order update, or code deployment can alter security, revenue, or production behavior.
Salesforce says every transaction runs as an authenticated user. Existing object permissions, field-level security, sharing, and other authorization controls still apply, and audit records attribute the action to that user. Those controls are essential. They establish who or what had authority to attempt the action and leave a record for investigation.
Authorization does not establish correctness. A human administrator can make an authorized mistake, and an agent can do the same at greater speed. A valid API response confirms that Salesforce accepted the request. It does not confirm that the request matched the customer’s intent, that the source data was current, or that downstream staff accepted the change.
Picture an offboarding request arriving at 5:17 p.m. The agent finds the user, freezes access, removes a permission set, and records four successful operations. If the HR source carried the wrong employee identifier, the technical record can be immaculate while an active worker loses access before a deadline. The business needs the source event, identity match, approvals, executed changes, reversal time, and worker report in one trace. Tool success covers only the middle.
Headless 360 is also in beta. Buyers can test the architecture and its controls, but they should not treat a beta surface as proof of universal production reliability. Availability, product maturity, customer deployment, and accepted results are separate evidence states.
Work shifts away from the conventional screen. A sales operations analyst may ask an external agent to update a field without visiting the record page. A developer may inspect an object schema from a coding environment. The named Salesforce user remains relevant for authorization and accountability, even when that person spends less time inside Salesforce’s interface.
That creates a difficult renewal conversation. If an agent performs more of the navigation and routine manipulation, buyers may ask why the old seat count should grow with employee headcount. Vendors still bear infrastructure, security, support, and product-development costs when fewer people click through screens. A pure outcome fee may be impractical for work whose value appears later or depends on customer decisions. Seats, usage, and outcomes are likely to coexist for some time.
Gartner calls the pressure agentic arbitrage. Its July forecast estimates that $234 billion, roughly 20% of enterprise application software spending, could be exposed by 2030 as agents weaken the traditional connection between user growth and vendor revenue. That is a forecast about a category, not realized Salesforce displacement in 2026. It identifies the incentive conflict rather than settling the amount.
The work design is immediate. A platform owner has to define which operations an agent may discover, which require read-only access, which need human approval, and which are prohibited. Security teams need identity, least privilege, audit retention, incident response, and kill controls. Business owners need acceptance rules. Frontline employees need a way to see, correct, and contest agent actions that change their queue or customer record.
A broad tool catalog increases the possible AWU surface. A controlled operation, accepted update, and lower loaded cost determine whether the surface deserves to remain open.
Slack Code returns review to the channel
Slack Code reverses the camera angle. Headless 360 lets software act behind a screen; the code channel brings the people who frame, inspect, and approve a task into view.
Slack’s August 20 changelog describes a code channel created from a conversation. The workspace can contain the original discussion, an agent’s plan, code changes, a live HTML preview, and a canvas that preserves decisions. Teammates can add context, redirect the agent, review a diff, and decide what should ship.
Salesforce’s launch release positions product managers and designers beside engineers. It lists partner agents including Claude Code, Devin, GitHub Copilot, ChatGPT, and Vercel. Slack Code is available across Slack plans, while access to a partner agent remains a separate requirement.
The commercial stack can therefore contain a Slack seat, access to a coding agent, model or platform consumption, a source-control product, CI capacity, deployment infrastructure, and employee review time. Counting only one line will understate the cost. Counting every generated diff as delivered software will overstate the result.
Rob Seaman, Slack’s executive vice president and general manager, presents the shared channel as a way to make coding accessible to the broader product team. Malte Ubl, Vercel’s chief technology officer, describes a version of software work that stays close to the conversation where intent began. Mario Rodriguez, GitHub’s chief product officer, and Jeff Wang, Cognition’s president of new enterprise, appear in the launch material from the agent-provider side.
Their product case has merit. A product manager can catch a misunderstood requirement before an engineer reviews hundreds of changed lines. A designer can inspect the rendered interface. An engineer can challenge the implementation and tests. The history can preserve why a decision changed. Review is part of the product, not evidence that the agent failed by definition.
Review still has a cost and a quality profile. The useful coding record separates at least seven events: request accepted, plan approved, diff generated, tests passed, human review completed, change merged, and change deployed without rollback. An agent can produce many plans and diffs while the merge rate stays flat. A faster merge rate can coexist with more escaped defects. A shared preview can reduce design rework, or invite a larger group into low-value iteration.
The employee effect differs by role. A product manager may gain the ability to make a small change without waiting for a sprint, while assuming responsibility for specifying it safely. An engineer may spend less time writing boilerplate and more time reviewing unfamiliar code. A designer may intervene earlier. A security reviewer may receive a larger queue. Junior engineers may get more executable examples but fewer opportunities to build a solution from first principles.
Managers can distort the evidence without touching the model. If teams are praised for agent acceptance, a reviewer may approve a marginal diff rather than become the person who slowed the metric. If rejected work lowers a manager’s reported adoption, employees can learn to repair it quietly after approval. Rejection therefore needs protection as a quality signal. Reviewers should be able to mark a result unsafe, incomplete, or wasteful without the act counting against their performance.
None of those outcomes can be read from agent activity alone. They require repository, incident, review, and employee evidence.
A practical weekly view would put generated diffs beside accepted diffs, median review minutes, reviewer count, test failures, post-merge corrections, reversions, and escaped incidents. It would also sample whether reviewers understood the change. A fast approval from an overwhelmed reviewer should not score as strong oversight merely because a button was clicked.
Slack Code shows why human work should not be treated as residue that disappears once an agent is installed. Collaboration can turn agent output into accepted work. Buyers still have to separate useful review from repetitive repair, then price both.
Pricing follows seats, credits, actions, and outcomes
The invoice can combine units that the product dashboard keeps apart. Salesforce’s usage and billing documentation describes consumption pricing, a hybrid of per-user monthly access plus consumption, and pricing based on a business metric. Agentic usage is tracked in actions, while embedded prompt usage is tracked in prompts. Native units are converted to Flex Credits through rate-card multipliers.
The metering boundary starts before a public rollout. Design and configuration activity may be unmetered, but preview interactions, voice preview, Testing Center runs, sandbox use, grid evaluation, and production activity can consume billable units. A team that treats testing as free experimentation can arrive at launch with a bill it did not model. A team that suppresses testing to save credits can create a larger quality cost in production.
Salesforce also offers a different unit through Agentforce Help Agent. Its second-quarter product summary says the product uses pay-per-resolution pricing and charges when an inquiry is resolved autonomously without human intervention. In one product family, Salesforce can therefore report platform activity in AWUs, meter underlying use through actions and prompts, preserve seat or hybrid access, and sell a named customer result.
Competitors expose the same design problem with different currencies.
Microsoft’s Copilot Studio licensing guidance offers prepaid packs and pay-as-you-go Copilot Credits, currently listed at $0.01 per credit for the pay-as-you-go option. Its billing-rate documentation assigns different consumption to different functions: a classic answer uses one credit, a generative answer two, an agent action five, and tenant graph grounding ten. Some employee use is included with a Microsoft 365 Copilot license.
Intercom’s Fin outcome rules charge $0.99 for a resolution, procedure handoff, or disqualification and $9.99 for a qualification under the current schedule. A conversation can generate at most one billed outcome. A failed procedure and an explicit escalation are not charged. A late reopening can deduct a resolution. The rules also permit an assumed resolution when a customer leaves without confirming the answer.
These products are not interchangeable, and their public prices should not be ranked as if one credit, action, resolution, or qualification carried equal work. Their definitions reveal what each supplier is willing to meter.
| Product or model | Commercial unit | Trigger in public documentation | Buyer-side evidence still needed |
|---|---|---|---|
| Salesforce platform activity | AWU | Production agent work such as a prompt, reasoning chain, or tool invocation | Technical completion, acceptance, rework, business result |
| Salesforce consumption | Action, prompt, and Flex Credit | Native usage converted through a rate-card multiplier | Contract price, test versus production use, cost per accepted result |
| Salesforce hybrid | User access plus consumption | Monthly entitlement and metered agent use | Seat utilization, agent utilization, overlap, human review |
| Salesforce Help Agent | Resolution | Autonomous resolution without human intervention | Repeat contact, quality, customer impact, exclusions |
| Microsoft Copilot Studio | Copilot Credit | Feature-specific rate, with pay-as-you-go or prepaid options | Accepted output, included seat use, unused capacity, correction cost |
| Intercom Fin | Outcome category | Resolution, procedure handoff, disqualification, or qualification under product rules | Durable resolution, sales acceptance, repeat contact, lifetime value |
Outcome pricing sounds closest to value, but the definition can still be gamed or simply differ from the customer’s. An assumed support resolution may be a reasonable billing convention when a user leaves. It is not proof that the issue stayed solved. A qualification can satisfy a configured rule while sales rejects the opportunity. An autonomous resolution may reduce agent labor while harming retention through a technically complete but unsympathetic answer.
Usage pricing has a valid defense. Vendors pay for models, data infrastructure, security, support, and product development before a customer’s revenue or cost outcome is visible. A customer can change its sales process after the agent produces a valid lead. It would be unfair to make a platform absorb every downstream management failure. Credits give both sides a countable resource and make marginal use visible.
Outcome pricing transfers definition risk back to the supplier. That can focus product work on completion, yet it also encourages a narrow event the supplier can observe and defend contractually. Usage pricing leaves more risk with the buyer but can support exploratory jobs whose value has not settled. Neither allocation is inherently honest or dishonest. Contract quality depends on which party controls the failure and which party can verify it.
Seats also remain useful where entitlement, collaboration, support, and accountability attach to named people. A Slack workspace does not cease to have value because a coding agent generated a diff. A Salesforce user identity can remain essential when a headless action inherits that person’s permissions and audit trail.
A workable contract may use a portfolio rather than one pure unit. Access can stay predictable. Consumption rates can stay visible. Where the product controls enough of the process, both parties can agree on an outcome definition and attach service levels or rebates to technical failures. The buyer still has to reconcile that portfolio against its own work record.
A work-unit bridge for the renewal meeting
Start the renewal file with a verb: resolve, update, provision, produce, qualify. That keeps the customer job above the vendor meter. Each row can then follow the job from authorization through activity, billing, technical completion, human handling, and business acceptance.
| Customer job | Platform activity | Vendor bill | Technical endpoint | Acceptance rule | Human work | Business evidence and decision |
|---|---|---|---|---|---|---|
| Resolve a support inquiry | Retrieval, reasoning, response, record update | Actions, credits, or resolution fee | Conversation closed | Correct answer; no avoidable repeat within the chosen window | Review, correction, escalation, recovery | Customer effort, repeat contact, retention; expand or retrain |
| Update an account record | Read source, validate fields, invoke update | Actions or credits | API accepts the write | Source match; no unauthorized field change | Approval, exception handling, rollback | Data quality, downstream error rate; widen or restrict permission |
| Provision an employee | Discover user, assign role, apply permissions | Platform use plus access products | Account and permissions created | Manager request matched; least privilege verified | Approval, security review, correction | Time to access, access incidents; automate, hold, or narrow |
| Produce a code change | Plan, generate, test, revise | Agent seat, model use, CI and platform cost | Diff and tests complete | Reviewed, merged, deployed, no defined rollback or defect | Product input, design check, code review, incident response | Lead time, rework, reliability; add use cases or cap queue |
| Qualify a sales conversation | Retrieve account, ask questions, score, route | Actions, credits, or qualification fee | Lead routed under configured rule | Sales accepts; required data present; duplicate removed | Rule design, review, follow-up, rejection coding | Conversion, cycle time, seller capacity; renew or redesign |
Three calculations turn the file into a decision instrument.
loaded cost per accepted result = access + partner agent + consumption + data + integration + testing + review + correction + escalation / accepted results
acceptance yield = accepted results / technically completed jobs
human handling per accepted result = review + approval + correction + recovery + escalation minutes / accepted results
The numerator and denominator need parentheses in an implementation workbook, and every input needs a time window. Access charges may be annual while consumption is monthly. Integration costs are high during setup and may fall when reused. Review can be required for a sensitive job even after quality improves. A quarterly average can hide a bad release week.
Ownership matters. Finance reconciles the invoice and contract. The platform team maps AWUs, actions, credits, tests, and environments. The business owner defines acceptance and the counterfactual process. Security defines authorization and prohibited actions. Frontline employees record correction, escalation, and work moved elsewhere. An employee representative or people leader checks whether workload, training, performance measures, or job access changed in ways the cost model misses.
NIST’s recent work on evaluation measurement warns that evaluation methods can collapse different notions of system performance into one score. Its research on deployed-system monitoring identifies human-AI feedback loops and oversight as unresolved monitoring problems. NIST did not assess Salesforce, Microsoft, or Intercom in those reports. The research supports the separation of measurement targets, not a verdict on these products.
The bridge should stay smaller than the work. A company does not need fifty measures for every agent. It needs the few that can change a decision. A low-risk internal summary may require sample accuracy, correction, and time saved. User deactivation needs authorization, exact account match, completion, rollback readiness, and an audit review. A customer-facing resolution needs quality, escalation, repeat contact, and customer impact.
That constraint matters most for smaller buyers. A ten-person support team cannot build a measurement department around one agent. It can still export a monthly invoice, sample twenty completed conversations, review every escalation and repeat contact, record correction minutes for one week, and compare the result with the prior queue. Suppliers can reduce the burden by exposing event-level exports, metric-version history, environment labels, and a stable join key. A glossy aggregate without exportable rejects shifts too much reconciliation work to the customer.
Sampling can contain the measurement burden. Review every high-consequence action, a random share of routine actions, all rejected outcomes, and all incidents. Increase the sample after a model, prompt, permission, integration, or product release. Reduce it only when the evidence supports the change. Keep a stable control group or baseline where feasible, because a faster workflow during a quiet month may reflect demand rather than the agent.
A rejection reason is as important as an acceptance count. Use a short controlled list: wrong source, wrong action, incomplete result, policy violation, poor format, duplicate, timeout, human preference, downstream failure, or business rule changed. Free text can preserve detail, but the controlled field shows whether the agent is improving or merely producing more.
The bridge also protects a successful agent from the wrong conclusion. If AWUs rise because the team expands a reliable workflow, acceptance stays high, human correction falls, and loaded cost improves, the activity increase is good evidence. If a sensitive workflow deliberately retains human approval, the approval minutes should not be labeled waste. The record lets the buyer reward productive scale and preserve necessary control.
Renewal then becomes a set of bounded choices. Expand jobs with improving acceptance and economics. Redesign a workflow when one fixable cause dominates rejection. Restrict permissions when consequence exceeds control. Stop jobs whose activity and bill grow without accepted output. Experimental work can keep a separate budget, with learning goals that do not masquerade as mature return.
Monday’s activity chart includes rejected work
Imagine a Monday renewal meeting with four lines on the first slide. Agent activity rose 70% during the quarter. Consumption charges rose 45%. Technically completed workflows rose 30%. Accepted customer outcomes rose 12%.
Those figures are illustrative, not Salesforce results. They show the decision the AWU era creates.
Finance can ask whether higher activity came from more useful jobs, deeper reasoning, a pricing multiplier, testing, retries, or an expanded product scope. The platform owner can identify which environments and operations generated the work. The business owner can show how much reached acceptance. Employees can show where review and correction moved. Procurement can compare the loaded cost with the previous process and the next-best product.
The rejection column keeps the meeting honest. If most rejected work came from one stale data source, the answer may be a repair rather than a license cut. If reviewers repeatedly rewrote acceptable work from habit, the rollout may need training and clearer authority. If an agent performed authorized but unwanted record changes, permissions and approval gates need attention. If customers reopened supposedly resolved cases, the outcome definition is too generous for the buyer’s purpose.
The same column helps the supplier. A vendor can improve a product when customers distinguish timeouts from wrong answers, policy blocks from poor routing, and required approvals from rework. A blunt demand to price everything by revenue would assign the platform responsibility for decisions it does not control. A blunt insistence that activity equals value would assign the customer responsibility for work it never accepted.
Salesforce’s 7.0 billion AWUs are evidence that agent activity has become large enough to deserve a first-class operating measure. The 3.2 billion quarterly count and 97% sequential growth make capacity, observability, and economics material. Headless 360 widens the action surface. Slack Code makes review visible. Help Agent tests an outcome unit. None of those facts requires the software seat to disappear.
They require the seat to meet other units in the same workbook.
At the end of the meeting, the CFO should be able to trace a dollar from entitlement to consumption, and the operations leader should be able to trace an agent action from initiation to acceptance or rejection. The employee who reviewed it should appear in the labor count. The customer who returned with the same problem should appear in the outcome record. A scope change in a vendor metric should appear in the chart’s footnote.
The 7.0 billion figure can remain on the earnings slide. The renewal decision begins one row lower, with a named customer job and a place to record the work the business refused to keep.