Your Team Looks More Productive Than Ever. Are They Getting Any Better?

Section 1: The Rep Who Stopped Writing Notes

Let’s start with a scenario:

The call ends at ten past three. Before the rep has even closed the video window, the record has been updated. A neat summary sits in the activity history. The opportunity stage has moved forward, the next step has been filled in, and a follow-up email is waiting in drafts. For the first time in years, the sales manager opens the pipeline on a Monday morning and finds it complete. Every opportunity has notes. Every next step has a date. It looks as though the adoption problem has finally solved itself.

Look a little closer at that summary, though. On the call, the buyer said something like, "If the budget is approved in the next planning cycle, we would probably want to look at this in the second quarter." What landed in Salesforce was a clean line: customer intends to purchase in Q2. The words were all there. The meaning was not. A hesitant, conditional maybe has become a commitment. That commitment is now feeding the forecast, shaping the manager's call with the leadership team, and influencing where the business puts its effort next quarter.

Nobody lied. The rep didn't make anything up, and neither did the AI. It did exactly what summaries are designed to do, which is turn messy human conversations into tidy statements. And the rep, who had three more calls before five o'clock, glanced at it, saw that it looked right, and moved on.

This is the quiet shift happening inside Salesforce orgs right now, and I think most leaders have missed it entirely. Whether it arrives through Agentforce, Claudeforce, a call recording tool or an AI assistant bolted onto the side of the platform, the effect on the user is the same. For as long as Salesforce has existed, the user has been the author of the record. They decided what was worth capturing, what the customer actually meant, and what should happen next. Everything we built on top of that, every report, every dashboard, every forecast, assumed a human had made those judgements first.

AI has changed that job. The user is no longer the author. They are the editor. And nobody told them.

That might sound like a small distinction, but anyone who has ever edited someone else's work knows it isn't. Writing a note forces you to think. You have to decide what mattered in the conversation, which concern was real and which was polite, and whether the buyer was genuinely interested or simply being courteous. Editing a note someone else has written is harder, because the thinking appears to have been done for you, and the result looks finished. To spot what is wrong, you need to know what should be there. You need to remember the tone of the call, notice the missing caveat, and recognise that "intends to purchase" is not what you heard. A good editor needs more understanding than the writer, not less.

Yet look at how most organisations have introduced these capabilities to their users. A short demo. A message in Slack saying the new feature is live. Perhaps a recorded session showing where to click to generate the summary. What almost nobody does is sit down with their teams and explain that their responsibility has changed. That the summary is a draft, not a record. That accepting it means putting your name to it. That the forecast, the handover to the account manager and the renewal conversation in eighteen months' time will all rely on what was approved in the thirty seconds after a call.

This matters because data quality has always started with people. For years the complaint was that users didn't put enough into Salesforce. Fields were empty, notes were missing, and next steps were weeks out of date. Leaders spent enormous amounts of time chasing their teams for updates. AI appears to solve that problem overnight, and that is exactly what makes it dangerous. An empty field is honest. Everyone can see it is empty. A confidently written summary that is slightly wrong looks exactly like good data. It passes every validation rule and survives every dashboard review. It simply isn't true.

I have seen discussions from organisations where leaders proudly showed how complete their records had become since switching AI on. When I question how often their reps changed what the AI produced, there was usually a pause. Nobody knew. When they looked, the honest answer in several teams was almost never. Not because the reps were careless, but because nobody had ever told them that checking was now the job. They had been told the AI would save them time, so they let it.

That is where the human side of this begins. The capability is new, but the pattern is very familiar. We introduce something into Salesforce, we show people how it works, and we assume the behaviour around it will sort itself out. It rarely does. Demo is not training, and with AI the gap between the two is wider than ever, because what users need to learn is no longer where to click. It is what they are now responsible for.

So what would I do if I were leading a Salesforce team that has just switched these capabilities on? I would start by redefining what the user's role actually is, in plain language, and saying it out loud in team meetings rather than burying it in a release note. If the AI drafts, the user owns. I would show the team real examples from their own records where a summary was fluent and wrong, because nothing builds the habit of checking faster than seeing your own forecast distorted by a sentence you approved. I would ask managers to compare a handful of AI-generated summaries against what they know about the deal in every pipeline review, and to recognise the people who corrected the AI, not just the people who used it. And I would make it clear that changing the AI's output is not a sign the tool has failed. It is a sign the user is doing their job.

None of that is complicated. But it does require leaders to accept something uncomfortable. The moment AI started writing the record, the most important skill your Salesforce users have stopped being data entry and became judgement.

Which raises a question that matters far more than any feature. If the user's real job is now to check the AI, what happens to the skills they need to check it, when the AI is doing the very work those skills were built on?

Section 2: Looking Competent Is Not the Same as Being Competent

Now, let’s picture a pipeline review with a rep who joined the business six months ago. On paper, they are doing brilliantly. Their records are complete, their activity is high, their emails are polished, and every opportunity has a clear next step. The manager points at one of the larger deals and asks a simple question. Why do you think this one is at risk?

The rep reads out the explanation the AI has generated. Engagement from the economic buyer has dropped, a competitor has been mentioned, and the close date has moved twice. All accurate. Then the manager asks the follow-up. So what are you going to do about it, and who in that account actually makes the decision? The room goes quiet, because the rep has never had to work that out for themselves. They know how to ask the system. They don't yet know how to read the deal.

I don't think that rep is lazy, and I don't think they were a bad hire. I think they are a product of the environment we have built around them. And I suspect there are a lot of them in Salesforce orgs right now, hidden behind some of the best-looking data those orgs have ever had.

There are two different things happening here, and leaders tend to lump them together.

The first affects your experienced people. An experienced rep, service agent or account manager built their instincts over years of doing the work by hand. They learned to read a pipeline because they had to build the picture themselves. They learned how a case should be resolved because they worked through hundreds of them. When AI starts doing that work, those instincts don't disappear overnight, but they stop being exercised. Slowly, the person who could once look at a pipeline view and feel that something was off starts relying on the summary to tell them. It is gradual and almost invisible, and the people it is happening to are usually the last to notice.

The second affects your newer people, and it is far more serious. A new joiner who arrives in an organisation where AI is already switched on doesn't lose a skill. They never build it in the first place. They learn to ask the agent instead of learning the process. They learn to accept the suggested case resolution instead of understanding why the case happened. They learn what the next-step field should say without ever learning why it matters to the person in finance who depends on it. They become fluent in the tool before they become competent in the job.

Onboarding used to do a great deal of quiet work that nobody ever designed. The tedious tasks we were all glad to see automated were also how people learned. Logging the call made you think about the call. Building the forecast made you understand the deals inside it. Working through a difficult case taught you the product, the customer, and the places where the process broke. When we removed that work, we removed the lessons hidden inside it, and very few organisations have replaced them with anything deliberate.

The problem is that all of this looks like success. Output goes up. Records are fuller. Response times improve. Every metric the business normally checks moves in the right direction. Underneath, capability is moving the other way, and you only find out on the day it matters. The day the AI is confidently wrong about a strategic account. The day a customer escalation doesn't fit the pattern the agent recognises. The day your best rep leaves and you discover that half the team had been running on their judgement and the system's suggestions rather than their own.

There is a behavioural layer to this as well, and it is the part I see most often when I sit with teams. Some users trust the AI completely and never override it, because overriding feels like arguing with the company's expensive new investment. Others don't trust it at all and quietly work around it, redoing the work by hand and saying nothing, which means the business is paying for the capability twice while believing adoption is fine. And a surprising number perform a confidence they don't actually have, because everyone around them seems comfortable and nobody wants to be the one who doesn't get it. None of those behaviours show up in a usage report. All of them affect the quality of what ends up in Salesforce.

This is why I keep coming back to organisational health rather than technology. The question isn't whether the AI works. In most cases it works remarkably well. The question is whether the people using it can tell when it doesn't, and whether they feel safe saying so. If the honest answer is no, you haven't increased your team's capability. You have hidden the lack of it.

So what would I do? I would protect a small number of core skills on purpose, particularly during onboarding. A new rep should build some forecasts and write some call notes by hand before the AI does it for them. Not because the manual way is better, but because it is how they learn what good looks like. I would ask managers to regularly ask the question behind the summary in one-to-ones: not what does the AI say, but what do you think, and why. I would make overriding the AI a normal, visible and even celebrated behaviour, so that "I changed that because it didn't match what the customer told me" becomes something people are proud to say in a team meeting. And I would pay close attention to the quiet workarounds, because a user who avoids the AI is giving you information about trust, clarity or training that no dashboard will ever surface.

Because here is what makes all of this harder. Just as the skills underneath are thinning out, many organisations are about to start measuring their people on how much they use the very tool that is causing it.

Section 3: We're Measuring the Wrong Thing

Let’s consider another scenario: The slide goes up in the quarterly leadership meeting and everybody is pleased. Usage of the new AI capabilities in Salesforce has risen sharply since launch. Thousands of summaries generated, thousands of emails drafted, a healthy share of users interacting with the agent every day. Someone says adoption is going really well, and the meeting moves on to the next item.

I understand why that slide feels like good news. For years, Salesforce leaders have struggled to get people to use the platform at all, so a graph that climbs steadily looks like a problem finally solved. But think about what that slide is actually measuring. It tells you how often people pressed the button. It tells you nothing about whether what came out was right, whether anyone checked it, or whether the people pressing it are getting better at their jobs or slowly getting worse.

Usage is not adoption. It never was, and with AI the gap between the two is wider than it has ever been. A user who generates a summary after every call and approves it without reading it is a power user on that dashboard. A user who generates it, finds two errors, rewrites the next step and corrects the close date looks less efficient. One of them is protecting your forecast. The other is quietly undermining it. Your report rewards the wrong one.

The same pressure is now flowing into how people are judged. AI raises expectations almost automatically. If a rep can produce more emails, more follow-ups and more updated records in a day, then more becomes the new normal. Some organisations are beginning to bring AI usage into performance conversations, and I can see the logic. Leaders want a return on the investment and they want people to engage with it. But if what you count is volume and frequency, you will get volume and frequency. You will train your people, very efficiently, to stop doing the one thing that matters most, which is thinking about what the AI has just given them.

This is where training comes in, because most of the AI training I see is aimed at the wrong target. It is largely about prompting. Here is how to ask the agent a question, here is how to generate a summary, here are five tips for better output. Useful, but it is still a demo with extra steps. It teaches people how to operate the capability, not how to work with it.

To be fair, a lot of Salesforce training has always had this weakness. For years we taught navigation and process: where the field is, which button to press, what order to do things in. That was survivable when the user did all of the work, because they picked up the judgement on the job. AI takes away much of the clicking, and with it much of that on-the-job learning. What is left, and what now matters most, is judgement. Knowing what a good opportunity record actually looks like in your business. Recognising when a summary has flattened a condition into a commitment. Understanding a case well enough to know when a suggested resolution is wrong. Feeling confident enough to override the agent, and to explain why.

That changes what training needs to be. You cannot teach judgement with a click-through demo or a list of prompt tips. You teach it with real situations, ideally drawn from your own org, where the right answer isn't obvious and people have to decide.

So here is where I would start if I were leading a Salesforce team with AI already switched on, or about to be.

Bring enablement in before you switch anything on, not after. Deciding what the user is now responsible for, what they must check, and where they should never simply accept the AI's output are training decisions as much as technical ones. If the first time your users hear about those responsibilities is after go-live, you have already set the habit you will spend the next year trying to undo.

Rebuild your Salesforce training around judgement rather than navigation. Use real examples from your own records: a summary that was fluent and wrong, a next step that sounded right but wasn't, a case the agent would have closed too early. Ask people to spot the problem and fix it. Practise the moment of overriding, so that it feels normal rather than rebellious.

Train your managers, not just your users. Managers are where behaviour gets reinforced or quietly ignored. If they never inspect AI-assisted work in pipeline reviews, case reviews and one-to-ones, the message to the team is that nobody is checking, so nobody needs to. Give them simple questions to ask every time. What did you change? What did the AI miss? What do you think is really happening in this deal?

Change what you measure. Keep tracking usage if you want to, but put it alongside measures of quality. How often are AI outputs corrected, and is that rate healthy or suspiciously close to zero? How accurate are your records when you spot-check them against what actually happened? How long does it take a new joiner to make sound calls on their own deals without the system telling them what to think? Those numbers are harder to collect, but they are the ones that tell you whether your investment is making your people better.

And be honest with your teams about expectations. If AI means you expect more output, say so. But say just as clearly that quality and accuracy are still theirs to own, and that slowing down to correct something is part of the job, not a failure to keep up.

None of this slows AI down. It is what makes it worth the money. The organisations that get real value from AI in Salesforce won't be the ones with the steepest usage graph. They will be the ones whose people are still sharp enough to know when the graph is wrong.