Salesforce just gave every sales leader a new problem and a new opportunity: a metric war.
In its fiscal Q4 2026 earnings release (period ended January 31, 2026, reported February 25, 2026), Salesforce introduced Agentic Work Units (AWUs) as a platform-level way to quantify “real work” completed by AI agents, explicitly positioning AWUs as the bridge between raw AI compute (tokens) and business output. Salesforce says it has delivered 2.4 billion AWUs to date and processed 19 trillion tokens all-time. That is not a product update. That is a signal that AI ROI measurement is about to get standardized, whether your team wants it or not.
Source: Salesforce earnings release
TL;DR
- Tokens and “AI usage” dashboards fail because they measure activity, not impact.
- AWUs (agentic work units) are emerging as the AI ROI metric because they count completed units of work that can be audited, quality-gated, and tied to pipeline outcomes.
- To implement AWUs in a lean B2B team, you need:
- an agentic unit-of-work taxonomy,
- quality gates per unit,
- anti-gaming rules, and
- outcome linkage (meetings, conversion, time-to-first-touch).
- Chronic Digital’s positioning: track outcomes, not busywork, using an AI Sales Agent plus pipeline predictions that connect agent work to revenue reality.
What Salesforce earnings tells us: AWUs are the new language of AI value
Salesforce’s Q4 FY2026 narrative was not “we shipped more AI features.” It was “we can now measure AI work.”
Key takeaways that matter for B2B sales teams:
- Salesforce introduced Agentic Work Units (AWUs) as a metric for “tasks accomplished by an AI agent.”
Source: MarketWatch recap - Salesforce framed the shift as moving beyond tokens and seats to quantify “real value created.”
Source: Salesforce newsroom AWU explainer - Salesforce reported scale: 2.4B AWUs delivered, 19T tokens processed, and Agentforce ARR $800M (up 169% YoY) in the same earnings package.
Source: Salesforce earnings release
Why this is bigger than Salesforce
If you sell B2B in 2026, you are already feeling the new buyer posture: skepticism toward AI claims, stricter procurement, and more demand for proof. When the largest CRM vendor starts naming a standardized unit of AI “work,” it changes:
- how boards ask for ROI,
- how finance teams think about AI costs,
- how RevOps teams design dashboards,
- how vendors pitch and price “digital labor.”
This is the same pattern we saw when teams stopped measuring “calls made” and started measuring “pipeline created.” Metrics shape behavior. Behavior shapes revenue.
Why tokens and generic AI usage metrics fail (and will get you budget-cut)
Most “AI ROI” reporting today looks like one of these:
- Tokens consumed
- AI messages sent
- Seats activated
- Time in app
- Number of generated emails
- Number of summaries produced
They are all fragile because they measure motion, not progress.
1) Tokens measure AI talking, not AI working
Salesforce says this directly: tokens tell you how much AI “talked,” not what it accomplished.
Source: Salesforce newsroom AWU explainer
A token-based dashboard can go up while pipeline goes down, for example:
- reps prompt the AI repeatedly to “make it better,”
- agents produce verbose outputs nobody uses,
- teams route easy tasks through AI because it is measurable.
2) “AI adoption” metrics can look great while outcomes stall
Research increasingly shows the right way to think about AI at work is intensity and impact, not just whether it is “on.”
For example, the St. Louis Fed summarized survey evidence that gen AI use varies widely in intensity, with a meaningful share of users spending 15 to 59 minutes per day, and some an hour or more. The point: “used AI” is not the same as “moved business outcomes.”
Source: St. Louis Fed: Impact of Generative AI on Work Productivity
3) Activity metrics invite gaming
If you reward:
- “AI emails generated,” you will get more emails, not more qualified meetings.
- “AI tasks completed,” you will get micro-tasks and duplicates.
- “tokens per rep,” you will get prompt bloat.
A good ROI metric must be:
- auditable (you can inspect examples),
- quality-gated (it only counts if it meets a bar),
- outcome-linked (it correlates with meetings, conversion, or revenue).
That is the logic behind agentic work units.
What are agentic work units (AWUs)? A practical definition for B2B sales
Salesforce’s definition: one discrete task accomplished by an AI agent, where “raw intelligence is converted into real work.”
Source: Salesforce newsroom AWU explainer
For a lean B2B sales team, here is the operational definition that makes AWUs implementable:
An agentic work unit (AWU) is a completed, reviewable unit of revenue work produced by an AI agent that (1) triggers or updates a real system of record and (2) meets a defined quality gate.
This definition matters because it excludes:
- drafts that never ship,
- research that never updates a lead record,
- “notes” that do not change next steps,
- automation noise.
Agentic work units taxonomy (sales): the 6 AWUs most teams should start with
You do not need 40 categories. You need a small set that maps to your funnel and your CRM objects.
Below is a starter taxonomy aligned to your brief, designed for implementation in a B2B SaaS, agency, or consulting sales motion.
1) AWU: Account research brief
What it is: An AI agent produces a 5 to 10 bullet research brief about an account, including relevant triggers and a tailored angle.
Counts when:
- the brief is attached to the Account record,
- at least 2 sources are cited (website pages, job posts, product pages, press release),
- it includes 1 recommended first step (email, call, LinkedIn, partner intro).
Quality gate (simple):
- Must mention: ICP fit signal, business model, and one plausible pain point.
- Must not include hallucinated funding, headcount, or tech stack without a source.
2) AWU: Lead enrichment completed
What it is: The agent fills missing CRM fields (company size band, industry, region, key persona, tech stack hints) and flags uncertainty.
Counts when:
- enrichment updates at least X required fields (you choose X, usually 5 to 8),
- confidence is recorded per field or a “needs review” tag is set,
- duplicates are checked before creation.
Quality gate (simple):
- Email domain matches company domain.
- Title and seniority are normalized (no free-text chaos).
3) AWU: Personalized email draft approved-ready
What it is: The agent drafts a cold email that is personalized to an account signal and ready for human approval.
Counts when:
- the draft is saved to the Prospect record,
- it includes a single CTA,
- it references a verified account signal (not generic “I saw you’re growing”).
Quality gate (simple):
- Under 120 words (or your deliverability best practice).
- No unsupported claims, no fake case studies.
- Reads like a human, not a press release.
(If you want to operationalize deliverability alongside AWUs, pair this with your internal deliverability SOP content. Link: Deliverability Ops SOP for Agencies)
4) AWU: Sequence launch (with guardrails)
What it is: The agent builds and launches a multi-step sequence (or prepares it for launch), segmented to the right persona and trigger.
Counts when:
- audience is deduped,
- suppression lists applied,
- steps have defined stop conditions (reply, meeting booked, bounce threshold).
Quality gate (simple):
- Segment definition stored (ICP + trigger + exclusions).
- A/B test rules or learning objective is specified.
5) AWU: CRM update from unstructured input
What it is: The agent reads an email thread, call notes, or Slack message and converts it into structured CRM updates: stage, next step, risks, stakeholders.
Counts when:
- at least 3 structured fields are updated (stage, next step date, deal risk, competitor, etc.),
- provenance is stored (link to the source email/call note),
- changes are logged.
Quality gate (simple):
- No stage changes without evidence in the source text.
- If uncertain, it creates a “review required” task instead of updating.
(Related internal architecture link: Conversation-to-CRM)
6) AWU: Meeting booked handoff package
What it is: After a meeting is booked, the agent prepares a handoff pack: agenda, hypothesis, stakeholders, relevant context, and a discovery plan.
Counts when:
- pack is attached to the meeting or opportunity,
- it includes 3 discovery questions tied to hypothesized pains,
- it includes one “mutual next step” recommendation.
Quality gate (simple):
- Must include ICP fit statement and a disqualifier check.
- Must list what the rep should verify first 5 minutes.
AWU quality gates: make “work” count only when it meets a bar
AWUs fail if they become “AI did something.” Your AWU framework needs explicit gates.
Recommended gate types (use 2 to 3 per AWU)
- Completeness gates: required fields present, required sections included.
- Evidence gates: must include sources, must reference a CRM field, must link to an event or signal.
- Compliance gates: opt-out language present, brand voice, no restricted claims.
- Impact readiness gates: ready-to-send, ready-to-launch, ready-to-handoff.
A lean team-friendly approach: 3 gate levels
- Auto-pass (deterministic checks): format, word count, field presence, dedupe.
- Sample audit (human QA): 5 to 10% weekly review.
- Outcome audit (RevOps): monthly correlation to meetings and conversion.
If you need a governance model, do not improvise it. Use a clear approval matrix. Link: AI Governance for RevOps in 2026
The missing piece: tie agentic work units to pipeline outcomes
Counting AWUs alone is still an activity metric. The point is to connect AWUs to outcomes that finance and leadership already trust.
Here are the most defensible outcome links for sales:
Outcome 1: Meetings booked per 100 AWUs (by AWU type)
Why it works: it forces the team to value “work that creates conversations,” not “work that looks impressive.”
Track:
- Meetings booked
- Meetings held
- Qualified meetings (SQL, SAO, however you define it)
Outcome 2: Stage conversion lift where AWUs were used
Examples:
- MQL to SQL conversion on enriched vs non-enriched leads
- Stage 1 to Stage 2 conversion when a handoff pack exists
- Proposal sent rate when CRM updates are up-to-date
Outcome 3: Time-to-first-touch and speed-to-lead
If AWUs include enrichment + first email draft + sequence launch, you should see first touch times drop.
Speed matters, especially for inbound. Link: Speed-to-Lead in 60 Seconds
Outcome 4: Rep time saved (but only if redeployed)
Time saved is not ROI unless it is reinvested into:
- more accounts touched,
- more follow-up,
- higher-quality discovery,
- more pipeline created.
If you need external evidence that gen AI can raise productivity, the Stanford/MIT field evidence (customer service context) is widely cited and found meaningful productivity lift with AI assistance.
Source: QJE: Generative AI at Work
The AWU Scorecard Template (copy-paste)
Use this as a weekly operating doc for RevOps and sales leadership. The goal is to make AWUs legible, comparable, and tied to revenue movement.
AWU Scorecard (Weekly)
1) AWU Volume (by type)
- Research briefs completed (AWU-RES): ___
- Enrichments completed (AWU-ENR): ___
- Email drafts approved-ready (AWU-EML): ___
- Sequences launched (AWU-SEQ): ___
- CRM updates from unstructured (AWU-CRM): ___
- Meeting handoff packs (AWU-HOF): ___
- Total AWUs: ___
2) Quality (pass rate + audit)
- Auto-pass rate: ___%
- Human audit sample size: ___
- Audit pass rate: ___%
- Top 3 failure reasons:
-
3) Outcomes (lagging, but essential)
- Meetings booked: ___
- Meetings held: ___
- Qualified meetings (SQL/SAO): ___
- Pipeline created ($): ___
- Stage conversions (pick 2 key transitions):
- Stage A -> B: ___%
- Stage B -> C: ___%
4) Efficiency
- Median time-to-first-touch (inbound): ___ minutes
- Median time from lead created -> first sequence step: ___ hours
- Rep admin time (self-reported or measured): ___ hours
5) AWU ROI snapshot (simple)
- Meetings booked per 100 AWUs: ___
- Qualified meetings per 100 AWUs: ___
- Pipeline created per 100 AWUs ($): ___
Anti-gaming rules: stop “AWU inflation” before it starts
If AWUs become a target, they will be gamed. Put constraints in writing.
Rule 1: Minimum unit size
Do not count micro-actions (for example “generated subject line”).
- An AWU must be a shippable artifact or a system update.
Rule 2: Deduplicate by object and time window
Example:
- Only 1 enrichment AWU per lead per 30 days unless a material change occurred (new role, new company, new funding, new tech event).
Rule 3: Cap AWUs per opportunity stage per week
If a deal is in “Negotiation,” 50 email draft AWUs are not progress.
Put a cap, then require justification to exceed it.
Rule 4: Penalize reversals and error correction
If the AI updates CRM fields incorrectly and humans have to undo it, you need a negative adjustment:
- AWU clawback, or
- a “rework unit” recorded separately.
Rule 5: Separate “attempted” vs “accepted”
Track two numbers:
- Attempted AWUs (agent produced)
- Accepted AWUs (passed gates, shipped, or approved)
You reward Accepted.
Rule 6: Outcome-weighted AWUs (optional, but powerful)
Not all AWUs should be worth the same.
A simple weighting model:
- Research brief: 1.0
- Enrichment: 1.0
- Email draft: 1.0
- Sequence launch: 2.0
- CRM update: 1.5
- Meeting handoff: 2.5
Then monitor: weighted AWUs per qualified meeting.
How to implement agentic work units in 30 days (lean team plan)
Week 1: Define AWUs and gates
- Pick 6 AWU types (start with the taxonomy above).
- Define:
- entry criteria (what triggers the agent),
- completion criteria (what must be produced),
- quality gates (2 to 3 checks),
- where the artifact lives (CRM object, attachment, activity).
Deliverable: a 1-page AWU spec.
Week 2: Instrumentation and logging
You need event logs that answer:
- what the agent did,
- when it did it,
- what record it affected,
- whether it passed gates,
- who approved it.
If your CRM cannot answer “what changed and why,” you cannot run AWUs.
This is also the week to align on security and governance requirements. Link: The 2026 AI Sales Tool Buying Checklist
Week 3: Roll out with sampling audits
- Start with 1 team (or 3 reps).
- Audit 10% of Accepted AWUs.
- Fix the top 3 failure modes (usually: bad data, weak personalization, missing sources).
Week 4: Tie to outcomes and set thresholds
- Build the scorecard.
- Agree on 2 leading indicators and 2 lagging indicators.
- Set initial baselines, not targets.
Targets come after you have baselines, otherwise you incentivize gaming.
Why Chronic Digital’s take is different: outcomes over busywork
Sales teams do not need another “AI activity” dashboard. They need a way to prove:
- pipeline impact,
- cycle-time impact,
- and forecast impact.
That is where Chronic Digital’s model fits AWUs better than legacy CRMs:
- AI Sales Agent produces work that maps cleanly to AWU types (research, enrichment, drafts, launches, updates, handoffs).
- Sales Pipeline with AI deal predictions lets you quantify whether AWUs are improving win probability, stage velocity, and forecast quality.
- AI lead scoring + enrichment ties upstream AWUs to downstream conversion, not vanity usage.
- Campaign automation makes AWUs “real” by shipping sequences with controls, not dumping drafts into a folder.
If you want a framework for separating “real agents” from marketing automation, use this: What Is Agent-Washing? 12 Tests
FAQ
FAQ
What are agentic work units (AWUs)?
Agentic work units are a way to count discrete, completed tasks performed by AI agents. Salesforce introduced AWUs as a metric to move beyond token consumption and better reflect “real work” delivered by agentic AI systems. For implementation, a sales team should define AWUs as completed, reviewable outputs that pass quality gates and connect to CRM records and outcomes.
Why are tokens a bad ROI metric for sales teams?
Tokens measure compute and language volume, not business progress. A team can spend more tokens generating longer emails, re-prompting, or producing unused content without booking more meetings or creating more pipeline. AWUs are more useful because they can be tied to completed tasks and then linked to pipeline outcomes.
What AWU categories should a lean B2B sales team start with?
Start with 6 categories that map to revenue motion:
- account research, 2) lead enrichment, 3) email draft, 4) sequence launch, 5) CRM update from unstructured input, and 6) meeting booked handoff. Each should have a clear definition, completion criteria, and quality gates.
How do we prevent AWU gaming and inflation?
Use anti-gaming rules: define minimum unit size, dedupe by record and time window, cap AWUs per opportunity stage, separate attempted vs accepted AWUs, and add clawbacks for rework. Most importantly, monitor AWUs alongside outcomes like qualified meetings per 100 AWUs and stage conversion lift.
How do we tie AWUs to revenue without waiting a full quarter?
Use leading indicators that move faster than revenue: time-to-first-touch, meetings booked, meetings held, and early stage conversion (lead -> meeting, meeting -> qualified). Then add lagging indicators like pipeline created and win rate once you have enough volume.
What does an “AWU scorecard” look like in practice?
A useful AWU scorecard tracks: AWU volume by type, quality pass rates and audit outcomes, and business outcomes like meetings booked, qualified meetings, pipeline created, and stage conversion. It also includes efficiency metrics like time-to-first-touch and a simple ratio like “qualified meetings per 100 AWUs.”
Build your AWU system, then enforce it like a revenue process
If Salesforce is right, 2026 is the year AI ROI stops being vibes and starts being accounting. Your advantage is not adopting the new metric first. It is adopting it correctly:
- define agentic work units that reflect real revenue work,
- gate them for quality,
- connect them to pipeline outcomes,
- and put anti-gaming rules in place before incentives warp behavior.
When you do that, AWUs become the metric that finally makes AI investments legible to leadership, and finally makes “agentic” mean “it shipped work that moved pipeline.”