After Dreamforce: the agents are ready. Are your numbers?
A meeting you've probably already had
It’s the quarterly review. Three people bring three different numbers for the same thing.
The new AI assistant says the sales pipeline is 3.4 times the target. The sales forecast says 2.9 times. Finance has a third number, lower than both.
Nobody made a mistake. Each one simply counted a little differently. And the meeting that was meant to be about the plan turns into an hour of arguing about whose number is right.
This happens in most companies today. It’s about to happen a lot more, because more and more of the “people” asking for your numbers will be AI.
What Salesforce announced last month
Salesforce, the world’s biggest sales software company, held its yearly event, Dreamforce, from 15 to 17 September. The big theme was AI agents: software assistants that don’t just answer questions, but actually do tasks.
In plain terms, Salesforce showed:
- An AI coworker inside Salesforce that can find answers and carry out work for you. Available now.
- Agents that work on a goal for weeks, such as "win back the deals we're at risk of losing", checking in with you along the way. In early testing, with wider release planned for November.
- Tools to watch what agents do, so you can see each step an agent took and spot where it keeps going wrong.
- Teams of agents, where one main assistant splits a job and hands pieces to specialist assistants.
This is real progress. The industry has moved from “look what AI can do” to “here’s how to run it safely day to day.”
What was still missing was proof of value. Industry observers noted there were very few hard numbers on what these agents save, or what they cost to run.
The problem nobody is talking about
The new tools can show you what an agent did. They can’t tell you whether it counted the same way as the rest of your business.
Take a simple word like “pipeline”. Behind it sit lots of small decisions:
- Which deals count: all of them, or only the likely ones?
- Do renewals of existing customers count, or only new business?
- Which date decides the quarter?
- Which currency, and at what exchange rate?
Your business made each of these decisions once, somewhere. A person who has worked there for years knows the answers. An AI agent does not, unless someone tells it.
The tricky part is that nothing looks broken. The agent doesn’t show an error. It just gives a number that doesn’t match anyone else’s.
Why AI makes this problem bigger, not smaller
- More places do the counting. Today, one report calculates pipeline. With a team of agents, four different assistants might each calculate it their own way.
- No one checks the working. When an analyst builds a report, a colleague can look over it. An agent working for six weeks makes thousands of these choices with no one reviewing them.
- Small mistakes grow. If an agent misjudges which deals are "at risk", you don't get one wrong number. You get six weeks of effort spent on the wrong customers.
- It shows up at the worst time. These mismatches rarely appear in a small trial. They appear when the AI's number meets the finance director's number for the first time.
That’s why many AI projects stall. Not because the AI can’t do the work, but because leaders won’t act on numbers they can’t trust.
Four questions to ask before you roll out more AI
- 1. Do we have one agreed meaning for our ten most important numbers? If the answer is "it depends who you ask", the AI will pick one for you.
- 2. How many places calculate each one today? Every extra report or spreadsheet is another chance for the numbers to drift apart.
- 3. When an AI tool gives us a number, can it show where the definition comes from? A good answer points to one agreed rule. A weak answer just talks about how clever the AI is.
- 4. If finance changes a definition, how quickly does every tool follow? If the honest answer is "we'd have to fix each one by hand", AI will multiply that work.
The takeaway
The AI agents are ready. Agreeing what your numbers mean is the work that still needs doing.
The fix is simple to describe: decide what each number means once, write it down in one shared place, and make every person and every AI tool use that same definition.
That’s the problem we built Equative Insights OS to solve. In Friday’s post, we’ll show you how it works.