Salesforce finally said the quiet part out loud: per-seat pricing makes no sense when the “user” is an AI agent.
So they’re pushing Agentic Work Units (AWU) as the new unit of value. Not tokens. Not seats. Not vibes. Work. One discrete task completed by an agent. (salesforce.com)
If this sticks, it does not just tweak Salesforce pricing. It rewrites how every CRM vendor sells “AI.”
TL;DR
- Agentic Work Units AWU = Salesforce’s attempt to price AI on “work done,” not “seats used.” (salesforce.com)
- Salesforce says 2.4B AWUs delivered to date, plus 19T tokens processed. They want AWUs to be the board-friendly metric. (s205.q4cdn.com)
- AWUs are still activity, not outcomes. Even CIO.com called out the gap: throughput is not business value. (cio.com)
- For an AI SDR, the only scoreboard that matters is: conversations started, qualified replies, meetings booked, show rate, pipeline created.
- Vendors will game AWUs. Expect “work inflation,” attribution tricks, and dedupe theater unless you demand controls.
- Buyer checklist at the end. Use it. Be annoying.
Salesforce’s Agentic Work Units AWU in plain English
Salesforce defines an AWU as one discrete task accomplished by an AI agent. Think “tool invoked,” “workflow triggered,” “record updated,” “reasoning chain completed.” (salesforce.com)
They frame it as the bridge between:
- Tokens = raw compute consumption (cost signal)
- AWUs = “real work” delivered (output signal) (salesforce.com)
Salesforce also tied AWUs directly to company momentum on its FY26 Q4 earnings release:
- 2.4 billion AWUs delivered
- ~20 trillion tokens consumed
- AWUs growing 57% quarter over quarter
- FY26 revenue $41.5B and RPO $72.4B (s205.q4cdn.com)
This is not subtle. They’re telling the market: “Stop asking what the model cost. Ask what the agent got done.”
Axios put it even more bluntly: Salesforce wants to kill “tokenmaxxing” and measure AI on output and impact instead. (axios.com)
Why this is really about killing per-seat pricing
Per-seat pricing worked when:
- Software was used by humans.
- More usage roughly meant more people.
- “Adoption” meant “logins.”
Agentic software breaks that.
- One agent can do the work of many reps.
- “Usage” goes up while “seats” stay flat or drop.
- Charging per seat punishes the customer for automation.
Salesforce already started moving toward flexible, consumption-style pricing for agents with Agentforce “Flex Credits,” explicitly tied to “digital labor” and swapping between seats and credits. (salesforce.com)
AWUs are the next step: a metric they can point to when finance asks, “What did we get for this spend?”
The uncomfortable truth: AWUs measure work, not value
CIO.com nailed the core problem: AWUs can capture execution, not accuracy, and not business value. A workflow triggering counts even if it was the wrong workflow. (cio.com)
So here’s the stance:
AWUs are useful as an internal efficiency meter.
They are not a purchase decision metric unless you tie them to outcomes.
If vendors sell you AWUs without outcome mapping, you bought a prettier token counter.
Agentic Work Units AWU: what an AI SDR should measure instead
If the use case is outbound and pipeline, do not accept “agent activity” as success.
Measure the funnel like an operator.
1) Conversations started (not emails sent)
Define “conversation started” as:
- A human reply that is not OOO.
- Or a positive signal in another channel (LinkedIn reply, call connect, form fill).
- Or a verified warm intent event (depends on your stack).
Why: volume is cheap. Attention is not.
Baseline reality check: many cold email benchmarks peg average reply rates in the low single digits, often under 5%. (assets.mailshake.com)
2) Qualified replies
Split replies into:
- Qualified: matches ICP, real pain, willing to engage.
- Neutral: “not now,” “send info,” “try later.”
- Negative: “no,” “remove me,” “stop spamming.”
Qualified reply rate is where personalization, targeting, and timing show up.
3) Meetings booked
Booked meeting rate ties outreach to calendar reality.
Track:
- Meetings booked per 1,000 prospects contacted.
- Meetings booked per qualified reply (conversion quality).
4) Show rate
The silent killer. Calendars lie.
Track:
- Show rate by channel (email vs LinkedIn vs phone).
- Show rate by persona.
- Show rate by meeting source (agent booked vs human booked).
If show rate drops, your agent is “booking” low-intent meetings. Congrats on the dashboard.
5) Pipeline created (and attributed)
Pipeline created is the metric that survives CFO scrutiny.
Track:
- $ pipeline created in 30/60/90 days from first touch.
- Stage progression speed.
- Win rate and ACV by source.
If your vendor cannot do defensible attribution, AWUs will turn into a vanity scoreboard.
How to translate AWUs into outcomes per dollar
Use AWUs like electricity usage. Useful. Not the bill’s purpose.
The board question stays the same:
“How many meetings and how much pipeline did we buy?”
So map:
- Cost per AWU (vendor pricing)
- AWUs per outcome (your conversion data)
- Cost per meeting and cost per $ pipeline
Salesforce itself says tokens-to-AWUs is “elastic” and will diverge as systems get more efficient. (salesforce.com)
That’s good. It also means your cost model must live at the outcome layer, not the AWU layer.
A simple ROI model for Agentic Work Units AWU (AI SDR edition)
Keep it stupid simple. Then get stricter.
Step 1: Define your unit economics
You need these inputs:
Volume and conversion
- Prospects contacted per month
- Conversation rate (reply rate or multi-channel response rate)
- Qualified rate (qualified replies / total replies)
- Meeting book rate (meetings booked / prospects contacted)
- Show rate (shows / meetings booked)
- SQL rate (SQLs / shows)
- Close rate (wins / SQLs)
Money
- Average deal size (ACV or first-year ARR)
- Gross margin (if you want contribution margin ROI)
- Sales cycle length (for payback period)
Costs
- Vendor cost (monthly)
- Data costs (if separate)
- Email infrastructure costs
- Human oversight time (hours, fully-loaded cost)
Step 2: Convert outcomes to dollars
Example (use your numbers, not mine):
- 10,000 prospects contacted/month
- 3% replies = 300 conversations started (assets.mailshake.com)
- 35% qualified = 105 qualified replies
- 40% convert to meetings = 42 meetings booked
- 70% show rate = 29 shows
- 30% to SQL = 9 SQLs
- 25% close rate = 2.25 deals
- $12,000 ACV = $27,000 new ARR/month
Step 3: Compute ROI
If total monthly cost is $4,000, then:
- Net-new ARR: $27,000
- ROI multiple: 6.75x (before retention, churn, expansion)
Also compute:
- Cost per meeting booked = $4,000 / 42 = $95
- Cost per show = $4,000 / 29 = $138
- Cost per SQL = $4,000 / 9 = $444
This is how you evaluate an AI SDR. Not by “agent tasks completed.”
Step 4: Now bring AWUs back in
If vendor charges per AWU, you add:
- AWUs per conversation started
- AWUs per meeting booked
- AWUs per $1,000 pipeline created
That’s your guardrail against runaway consumption.
How vendors will game Agentic Work Units AWU (because of course they will)
If you pay per unit, vendors will manufacture units. Not always malicious. Often just incentive gravity.
1) Work inflation
Make one “task” into five:
- “Research account” (1)
- “Find contact” (2)
- “Draft email” (3)
- “Check policy” (4)
- “Log activity” (5)
Congrats. Your AWUs went up. Your pipeline did not.
Fix: require a task taxonomy and enforce “compound tasks” reporting.
2) Tool-call spam
Salesforce emphasizes “tool invoked” as the most important AWU moment. (salesforce.com)
So vendors will invoke tools constantly.
Fix: track tool calls per outcome. Penalize runaway ratios.
3) Attribution laundering
Credit the agent for pipeline that came from:
- An existing inbound lead
- A partner intro
- A renewal expansion
- Someone already in an active sequence
Fix: hard attribution rules. First-touch, last-touch, and multi-touch. Pick one as the commercial standard.
4) Dedupe theater
Count the same person multiple times because:
- Variations in email
- Multiple domains
- Multiple CRMs
- Bad enrichment
Fix: dedupe at contact + account level. Enforce identity resolution rules.
5) “Activity” success criteria
They’ll show:
- emails sent
- tasks completed
- sequences launched
- records updated
All true. All useless without conversion.
Fix: require outcome dashboards tied to revenue stages, not activity streams.
What Chronic would measure (and why AWUs are still useful)
AWUs are a decent internal counter for “how much the agent did.”
But outbound lives and dies on:
- Fit
- Intent
- Timing
- Deliverability
- Follow-up discipline
- Attribution integrity
That is why an autonomous SDR must run end-to-end, till the meeting is booked. Not “generate some tasks.”
If you want pipeline on autopilot, the backbone looks like:
- ICP definition that doesn’t rot: ICP Builder
- Data that doesn’t bounce: Lead Enrichment
- Copy that doesn’t read like AI: AI Email Writer
- Prioritization that reflects reality: AI Lead Scoring
- A pipeline that stays clean: Sales Pipeline
And if you want the unglamorous math behind why inboxes punish sloppy outbound, read:
Buyer checklist: what AWUs must include so this doesn’t become another vanity dashboard
Print this list. Bring it to the demo. Interrupt them.
1) AWU definition and taxonomy (in writing)
Demand:
- A clear definition of what counts as an AWU.
- A list of AWU categories (research, enrichment, messaging, logging, workflow, escalation).
- Whether AWUs include retries, failures, and backoffs.
If they cannot explain it in 60 seconds, it’s not a metric. It’s a spell.
2) Audit logs you can export
Must-have:
- Timestamped event logs
- Actor (agent vs human)
- Input source (CRM field, enrichment, web, manual note)
- Action taken (tool call, email sent, record updated)
- Result (success, fail, rollback)
No audit log means no trust. Full stop.
If you care about agent governance, start here:
3) Dedupe rules (contact and account)
Must-have:
- Exact match and fuzzy match rules
- Merge strategy
- Prevent duplicate outreach across domains and inbox pools
If they “dedupe later,” you pay twice and burn reputation.
4) Attribution rules that match finance
Must-have:
- Clear model (first touch, last touch, weighted multi-touch)
- How they handle:
- existing open opportunities
- influenced pipeline vs sourced pipeline
- territory conflicts
- recycle rules (old leads re-contacted)
If attribution is vague, AWUs will “create” pipeline by redefining pipeline.
5) Outcome mapping: AWUs to funnel metrics
Must-have dashboards:
- AWUs per conversation started
- AWUs per qualified reply
- AWUs per meeting booked
- AWUs per show
- AWUs per $ pipeline created
This stops work inflation.
6) Quality controls, not just throughput
Must-have:
- Scoring rubrics for “qualified reply”
- Disposition definitions
- Human review sampling
- Automatic suppression rules (competitors, customers, do-not-contact)
Otherwise, the agent optimizes for the cheapest win: noise.
7) Cost predictability guardrails
Must-have:
- Monthly caps
- Per-workflow caps
- Alerts on anomalous AWU spikes
- Budget-based throttling (not “surprise invoice” mode)
Salesforce itself positions AWUs as a way to make AI spend legible. Do not accept illegible billing. (axios.com)
FAQ
What does Agentic Work Units AWU mean?
Agentic Work Units (AWU) are Salesforce’s metric for “one discrete task accomplished by an AI agent,” meant to represent work performed across its agent ecosystem. (salesforce.com)
Are AWUs better than tokens for pricing AI?
For buyers, yes and no. Tokens measure compute consumption. AWUs measure “work done.” But AWUs still don’t guarantee business outcomes like meetings booked or pipeline created. Salesforce explicitly positions AWUs as a replacement for tokens-as-success. (salesforce.com)
What should I measure for an AI SDR instead of AWUs?
Measure funnel outcomes:
- conversations started
- qualified replies
- meetings booked
- show rate
- pipeline created and revenue
Use AWUs as a secondary efficiency metric, not the KPI.
How do I prevent vendors from gaming AWU metrics?
Demand:
- audit logs
- a written AWU taxonomy
- dedupe rules
- attribution rules
- outcome mapping (AWUs per meeting, AWUs per $ pipeline)
If they won’t commit, they’re planning to win the dashboard, not the deal.
Did Salesforce provide real traction numbers for AWUs?
Yes. Salesforce reported 2.4 billion AWUs delivered and ~19-20 trillion tokens processed, and introduced AWUs as a core “agentic enterprise” metric in its FY26 Q4 earnings materials (dated February 25, 2026). (s205.q4cdn.com)
Does AWU pricing mean per-seat pricing is dead?
Not automatically. Salesforce has discussed a mix of models, including consumption options and packaging changes. The shift is real, but many enterprises still demand predictability, so expect hybrid pricing for a while. (salesforce.com)
Run the AWU test in your next vendor call
Ask one question and refuse to move on:
“How many AWUs does it take to book one qualified meeting, and what controls stop that number from drifting up over time?”
If they answer with a slide about “agent activity,” they’re selling you motion.
If they answer with:
- conversion rates,
- attribution rules,
- audit logs,
- and a hard cost-per-meeting model,
then you’re finally talking about outcomes per dollar. The only metric that matters.