Short answer. Salesforce says queries to its agent grew 21x, and it charges $50 a seat. Business Wheel in Dubai counted what its own 37 agents made, threw out more than half as empty, and kept about 152 hours for $100 of AI costs. A query is not work. Before you pay for seats, count what an agent delivers on one process.
Two agent stories came out in the same week, and they measure different things. On 4 November Salesforce said usage of its Agentforce Coworker grew 21 times in three months and set a price of $50 per user per month. On 8 November Business Wheel, a Dubai firm, published what 37 agents built by its own staff had actually produced.
One number tells you people use the agent. The other tries to tell you what the agent did. If a vendor is offering you seats this quarter, you need the second kind of number, and you can build it yourself.
What actually grew 21 times?
Queries. Salesforce's own post puts it at nearly six million Coworker queries in September, worldwide, with more than 185,000 users.
That is a real adoption figure, and I am not arguing with it. It answers a different question. A query is something a person typed. It does not tell you whether anyone used the answer, whether someone had to redo it, or whether the task would have got done anyway.
Two days later TahawulTech ran the announcement under the headline that UAE enterprises are handing real work to AI, with a quote from Salesforce's head of the Middle East. The 21x and the 185,000 are global figures. Nobody published a UAE count. By Salesforce's own description, Coworker returns results and recommended actions that a person reviews and approves, so a query is at best where a piece of work starts.
How did Business Wheel count its agents?
It counted everything, threw out the empty part, and called the rest capacity rather than money.
The detail sits in the notes to the press release. After a company AI bootcamp on 28 September, staff in six departments built 37 agents on ClickUp, the work platform they already used. Since they went live, mostly in September, the agents have produced more than 2,300 tasks, comments and messages.
Then the firm did three things most vendor case studies skip:
- It priced a record. A record is anything an agent posted: a task, a comment, a message. Working from ClickUp's own benchmark, the firm put each one at about nine minutes of human work.
- It threw out the empty part. It removed routine "nothing new" checks, repeated verifications and setup conversations. They made up more than half of the raw total.
- It named what was left. About 152 hours, worth roughly $7,600 at ClickUp's rate of $50 an hour, against about $100 in AI running costs. The release calls these values "capacity freed, not cash saved".
| Salesforce Coworker | Business Wheel | |
|---|---|---|
| What was measured | Queries and users | Records the agents produced, minus empty ones |
| Unit | Query | Hour of human work |
| Whose figures | Salesforce, worldwide | The firm, about itself |
| What is not counted | Whether a query became finished work | Time staff spent building and reviewing agents |
The rate is the weak point. Nine minutes a post comes from the platform vendor's benchmark, and the firm applied it to every agent. It could not see private conversations between staff and their own agents, so it calls its numbers a floor. The release states both caveats itself, which is more than most case studies give you.
What does that do to the arithmetic?
You get smaller numbers, and you can check them.
Take the release's own figures. Multiply 2,300 records by nine minutes and you get about 345 hours. Drop the empty half and you are left with the 152, about 44% of the raw count. Spread that over 37 agents and you get roughly four hours per agent since most of them went live in September. A few busy agents on schedules carry most of it, and the firm says it will retire agents nobody uses.
Now put a seat price next to it. If you buy twenty seats at $50, you pay $1,000 a month. At the same $50 an hour, you need those seats to free about 20 hours a month, one hour per person. That bar looks low, and it is. The hard word is "free". An hour you count at a vendor's rate clears the bar only if somebody spends it on work you would otherwise pay for.
We compared subscription and custom builds in money in Subscription or custom build. You do this count before that comparison. Until you know what the agent delivered, neither price tells you much.
How do you run this count yourself?
Pick one process, time it before the agent, decide what a result is, and throw out everything else.
- One process, one owner. Choose something your team repeats every week, such as drafting replies to supplier requests. Whoever does it today owns the count.
- A baseline before the agent. For two weeks, log how long the work takes and how often your team does it. Business Wheel had owners record before and after times for four agents. Those are the strongest numbers in its release.
- Define a result in advance. A reply you send without rework is a result. A draft somebody rewrote from scratch is not. Neither is a check that found nothing new. Write this down before the agent goes live.
- A month of counting. Count results only, multiply by the baseline time, and subtract the minutes your people spent reviewing. Call the total hours freed. Turn it into money only when you can say where those hours went.
Step three is the hard one. An agent produces a lot of activity, and a model cannot tell you which of its own outputs mattered. You have to decide that, and deciding is most of the work.
Training will not decide it for you either. We wrote about that in training is not adoption. If your staff build agents on their own, list those agents first. That is the subject of who in your company has already connected AI to your data.
When we build an AI employee for a company, we start with this count: what the process costs today and what will count as done. The agent comes after.
Danil Ivanov
Founder, KAIVIX
Builds AI systems for companies in the UAE and beyond.