What I see on the ground
The vast majority of companies that call us to implement AI are not ready to implement AI. I have no figure to back that up and I am not going to invent one. It is what happens engagement after engagement.
And that is fine.
Artificial intelligence changes daily and not every company moves at that speed. The same thing happened with the dot-com wave. It happened with social media. It happened with digital marketing. There are still companies today that have none of it sorted out, and they run well, they bill well, and they are references in their region. Missing the first wave was never the same as being left out.
What is a problem is buying the tool before knowing what it is for.
The first half of the work is measuring the before
Back to the original question. To know whether your team is 30% more efficient, you first have to know what it produces today. And that is not one number, it is several: how long a person takes to answer a case, how many proposals they put together in a week, how much time passes between an enquiry arriving and an answer going out, how many contracts they close a month, how much of the day goes to work nobody would call work.
That exercise is half the project and almost nobody does it, because it has nothing attractive about it and it cannot be shown in a demo. But without that picture of the before, any number you quote afterwards is a feeling.
Once the baseline exists, the rest falls into place. You know what to look at, you know what should move, and you can argue about whether it moved.
Before the platform, map the process
On most projects we recommend mapping the processes before touching any platform. Not for tidiness. It is so you do not drag hidden vices into the new technology.
A process that today carries three approvals nobody remembers the reason for, once automated, becomes a process with three useless approvals that now run faster. Automating a mess does not clean it up, it speeds it up. And it adds a layer on top that makes the original problem harder to see.
The hard question: what does your team do with the time you freed up
Say everything goes well. A task that took an hour now takes forty-five minutes. What does that person do with the fifteen minutes?
This is where it gets genuinely difficult, and there is data that helps explain why. Workday surveyed 3,200 full-time employees at companies with revenue above US$ 100 million, all active AI users, across North America, Asia-Pacific and Europe. 85% say they save between one and seven hours a week.
Now the two numbers that matter. Close to 40% of that saved time is lost to rework: fixing errors, rewriting output, verifying results. And only 14% of employees get consistently net positive results.
Among those who do get positive results, 57% use the freed time for higher-value work. On the other side, 32% of organisations simply pile more work on their people. That is not a productivity strategy, it is a fairly quick way to burn out a team.
Workday is a vendor and commissioned its own study, exactly like Salesforce did. I mention it to hold both to the same standard.
The problem with asking your team whether they feel faster
There is one experiment I keep coming back to. In July 2025 METR published a trial with 16 experienced developers across 246 real tasks in their own repositories, randomising task by task whether they could use AI.
With AI they took 19% longer. Before starting they expected to be 24% faster. After finishing, already knowing how it had gone, they still believed AI had sped them up by 20%.
It is a small study, on a very specific profile of people, with early-2025 tools. The authors are the first to point out it does not support concluding that AI speeds up nobody, or that it fails outside software development. But the finding that does travel is a different one: perceived productivity and actual productivity can move in opposite directions without anyone noticing.
If your only measurement is asking the team whether they feel faster, you already know what they are going to say.
Buying the licence is the easy part
Paying for a per-employee licence of Claude, GPT or Agentforce costs nothing next to the rest. The hard part is getting people to work with the tool and turning that time into something. You do not buy that, you build it, and it looks a lot more like a change of culture than a technology project.
There is one piece of that change almost nobody talks about: teaching the team not to sit and wait for the agent to finish. If a person hands off a task and then watches it run, the time saved is zero. The gain shows up when they start the next thing while the first one resolves itself, and that has to be taught and measured like anything else.
Where the 30% comes from
The question always arrives the same way. We implemented the platform, the team is using it, and someone in the meeting asks how much more efficient we are now. The answer everyone expects to hear is 30%.
That figure has been bothering me for months, because I hear it everywhere and I had never seen anyone explain where it comes from. So I went looking.
It comes from somewhere specific. In May 2025 Salesforce published research conducted with NewtonX: a double-blind survey of 200 chief human resources officers worldwide. The finding people quote is that, once agentic AI is fully implemented, CHROs expect an average employee productivity gain of 30%.
Expect.
It is not a measurement. It is what two hundred executives believe will happen. The same study projects a 19% reduction in labour costs, around US$ 11,064 per employee based on OECD salary averages, and that 23% of staff will be redeployed to other roles. All of that is expectation, and the original study presents it as exactly that. The trouble starts later, when the number breaks loose from its source and walks into a meeting room dressed as a promise.
We are a Salesforce partner and we work with Agentforce every day. This is not a criticism of the platform. It is an observation about how a number travels.
What happened when someone went to measure
In August 2025 MIT's NANDA project published a report combining 150 executive interviews, 350 employee surveys and an analysis of 300 AI projects. Of those 300, 95% showed no measurable impact on the profit and loss statement.
That 95% gets quoted constantly and almost always wrong. It does not say the projects failed technically. It says the impact could not be measured. And the reason the report itself gives is the one that matters to me: the companies had no documented baseline from before they implemented.
To be fair to the data: the 300 projects were not randomly selected, they came from the research team's own access. It works as a signal, not as an industry statistic. But the signal points at something I see every week.
All of this sounds abstract until it happens to someone. We tell one in full, anonymised and with the self-criticism first, in Failing with flying colours: an agent that works, that converts, and that part of the team considers a failure.
My conclusion
Not every company is ready to implement AI and agents today, and that is not a defect. This wave is here to stay and it is not going to be over by tomorrow night. There is time to sort out the processes, measure how the work happens today, think, and only then decide.
When someone asks us whether they are ready, the honest answer is almost always the same: not yet, and it does not matter. What matters is what you do in the next three months to get there.