“Which AI is better for our managers?” is a question HR and L&D leaders hear a lot right now. It sounds fair, but it sets up the wrong contest. It assumes an AI coach and an AI assistant are two versions of the same product, so one of them has to win. They aren’t the same product. Each one is built to get a different moment right, and each is rewarded for a different result. Once you know what each tool is rewarded for, most of the behavior people worry about starts to make sense.
So this isn’t a feature comparison. General assistants are very good at the work they were built for, and nothing below argues otherwise. This post is part of our series on AI governance for people management. It takes one decision HR keeps facing and asks it a better way: which manager jobs should go to which tool?
What is the difference between an AI coach and an AI assistant?
An AI assistant is a general-purpose tool. ChatGPT, Claude, Microsoft Copilot and Gemini all belong here. They differ in plenty of ways, but for this question they’re one category because they’re built for the same kind of job. You bring a task and the assistant helps you finish it. That might be a first draft of an email or a summary of a long thread. It might be an explanation of a policy you don’t follow, a first pass over a spreadsheet or some background reading before a meeting.
An AI coach is built for a different job. Its purpose is to change what a person does over time. For a manager, that could mean how they run a 1:1 or how they handle a conversation they’ve been avoiding. Our guide to what AI coaching is covers the longer version.
None of this is a quality ranking. An assistant is often the best tool a manager has for getting work done. Many managers already use one for people questions as well, and that habit is one reason this decision lands on HR’s desk at all. We’ve written separately about shadow AI in manager coaching, so we won’t go over that ground here. The useful question is what each one is built to get right.
What is each one built and rewarded for?
Every AI product is shaped by a signal that tells its makers whether it’s doing well. That signal matters more than most people expect, because over time a product gets better at whatever it’s measured on. So the place to start is the measure, not the feature list.
What does a general assistant get rewarded for?
An assistant is judged on the answer it gives you now, in one sitting. Did it help? Was it clear? Did you like it? Part of how these models are trained depends on human preference. People are shown answers and pick the one they prefer, or they press thumbs up or thumbs down. Those choices feed back into training. Over many rounds the model learns to give the kind of answer people pick.
That works well most of the time, because people usually prefer answers that are correct and useful. But it has a known weak spot. A team of researchers led by Mrinank Sharma studied five AI assistants and a large set of human preference data. They found that people sometimes preferred a convincing answer that agreed with them over a correct one. The models trained to predict what people like showed the same lean. The researchers call this sycophancy. Nobody wrote it in on purpose. It’s a side effect of rewarding the answer people like best in the moment.
OpenAI gave a very public example in late April 2025. It released an update to GPT-4o in ChatGPT and then rolled it back within days. OpenAI’s own account of what went wrong said “we focused too much on short-term feedback” and that the model “skewed towards responses that were overly supportive but disingenuous.” A longer follow-up from OpenAI explained the cause in more detail. The update had added a new scoring signal built from thumbs-up and thumbs-down data. Together with other changes in the same update, it weakened the main signal the model had been trained on. OpenAI said the model “aimed to please the user,” and that this went beyond flattery to validating doubts and fueling anger.
This case needs to be read fairly. It was one model from one vendor. OpenAI fixed it quickly and explained the mistake in public, in more detail than most software companies offer when something breaks. So it doesn’t show that ChatGPT flatters people today. What it shows is what happens when a tool leans too hard on immediate approval. The pull is built into the measure, because the thing being scored is how the answer lands in the next minute. Vendors work hard against that pull, and they mostly succeed. But they’re working against it, which tells you which way it points.
What is a coach rewarded for?
A coach is rewarded for something that happens later. The test is whether the manager ran their next 1:1 differently, or finally had the hard conversation they’d been putting off. Whether they enjoyed the session counts for little on its own. That result shows up weeks after the chat ends, in a meeting the tool never sees.
That’s a different target, and it changes what a good session looks like. A session that leaves a manager a bit uncomfortable can still be a good one if it gets them to try something new. A session that leaves them pleased but unchanged has failed, even if it felt helpful at the time. An assistant scored on the next minute would count the second session as the win. A coach has to count it as a miss.
Habits are slow, which is part of why the measure sits so far out. Phillippa Lally and her colleagues at University College London followed 96 people as they tried to build a new daily habit. On average it took 66 days before the new behavior felt automatic. The fastest person got there in 18 days, and for the slowest the researchers projected 254. These were simple health habits, like eating a piece of fruit with lunch or going for a run before dinner. The study didn’t look at management skills at all. It did find that more complex habits such as exercise took longer to form. So habits like listening before you answer are likely harder, not easier. Even so, the study shows the basic shape: new behavior takes weeks to months, not one good conversation.
There’s also evidence that workplace coaching can move these later outcomes. Rebecca Jones, Stephen Woods and Yves Guillaume pooled 17 studies in a 2016 meta-analysis of workplace coaching. They found positive effects on skills, on attitudes and on individual results at work. The effect on results was the largest of the three. Two details keep this honest. The review left out managers coaching their own reports, so it’s about coaching by someone outside the reporting line. And the size of the effect didn’t depend on how many sessions people had or how long the program ran. So the claim here is narrow. Coaching outcomes are measured later, and the behavior it aims at forms slowly.
Put the two measures side by side and the difference is clear. An assistant is scored on the answer. A coach is scored on what the person does after the answer, sometimes long after.
None of this is about which model is smarter. A coach can be built on the same kind of AI language model that powers an assistant. The engine can be the same. What differs is the job the tool is given and the result it’s held to. If you give the same model two different jobs and two different scorecards, it will learn to behave in two different ways. That one choice, made early in design, shapes most of what users notice later.
What does that one design choice produce?
Three behaviors follow from it. Each one is reasonable for an assistant doing its job. Each one matters more when the job is a manager’s development.
Why do assistants tend to agree with you?
Myra Cheng and colleagues at Stanford and Carnegie Mellon tested 11 AI models on requests for personal advice. The models affirmed the user’s own actions about 50% more often than humans did. The team then ran two experiments, with plans filed in advance, on 1,604 people. Those who got the agreeable replies were less willing to take steps to repair the conflict they’d described. Yet they rated those replies as higher quality, and they trusted the model more.
That matters for a manager in a specific way. Say a manager asks an assistant for help with a tense situation on their team. The manager types their own version of events, because that’s the only version they have. So the assistant hears one side, and its pull is toward agreeing with the person in front of it. The advice can end up confirming what the manager already thought. A good coach would ask what the other person might say about the same week.
The study has limits, and they’re worth stating plainly. It looked at advice on personal conflicts and used people recruited online. It measured short exchanges, not months of use. Vendors have also adjusted their models against sycophancy since. But the study explains why the pull doesn’t simply go away. People rated the agreeable replies higher, and ratings like that are the kind of signal assistants learn from.
Who does the thinking?
Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using AI at work. The study was presented at CHI 2025, a major human-computer interaction conference. Workers who had more confidence in the AI reported doing less critical thinking. Workers who had more confidence in themselves reported doing more. These are self-reports and correlations, so the study can’t say that AI use causes less thinking. It only shows the two are associated.
The study also found something fairer than a simple decline. Much of the effort moved rather than vanished. People spent less effort producing the work and more on checking the AI’s output and fitting it into what they were doing. That’s a real skill, and for most tasks it’s a sensible trade.
The trouble comes when the thinking is the point. When you ask an assistant for a finished answer, handing you one is the job done well. If a manager asks how to open a hard conversation and gets a ready-made script, the assistant has done exactly what it was asked. But the manager has skipped the part that builds the skill. They haven’t worked out what they want to say or why. They also haven’t thought about what could go wrong. A coach is set up the other way round. Its job is to make the manager produce that thinking, because the thinking is what changes the next conversation.
Who starts the conversation?
By default, the user starts the conversation with an assistant. That follows from how assistants are judged. An assistant is measured on how well it answers the question it was asked, so its natural position is to wait for a question.
That’s fine for most work, because you know when you need a summary. Development is different. It depends on the manager noticing that they need help, and the habits that most need work can be the ones a manager doesn’t think to ask about. A manager who talks over people in meetings may never type “how do I stop interrupting”, because they may not know they do it. Nothing in that gap is a failure by the assistant, which is simply waiting for a question that never comes.
Is the gap between assistants and AI coaches closing?
The gap has narrowed in some ways, and it’s only fair to say so. Assistants now remember. Claude can search earlier chats and use memory to build on past conversations. ChatGPT can also draw on past chats and saved memories. The old line that an assistant starts fresh every time is out of date.
Assistants have also added modes that hold back the answer. In April 2025 Anthropic added a learning mode to Claude for universities that guides students with questions instead of handing them answers. Google built Guided Learning into Gemini, and Google says plainly that it “won’t do it for you.” OpenAI’s study mode for ChatGPT asks questions and explains ideas step by step instead of only giving a final answer. It also checks your understanding as you go. These modes close much of the “it just hands over answers” gap for studying. Their limits are about fit. The user has to choose them, and they’re built around learning a topic. A manager trying to stop avoiding hard feedback isn’t learning a topic. They’re trying to change a behavior at work.
Assistants are also starting to reach out first. In September 2025 OpenAI started rolling out ChatGPT Pulse, beginning with a preview for Pro users on mobile. OpenAI says it “proactively does research to deliver personalized updates based on your chats, feedback, and connected apps.” The company describes this as a move “from answering questions to a proactive assistant.” ChatGPT can also run scheduled tasks that a user sets up, either once or on a repeat. So the default is shifting. What these features bring you, though, is what you asked for or what serves your work. They aren’t built around a development goal that a manager agreed to and is being held to.
Most assistants also let a user save standing instructions and reference files for one kind of work. So a motivated manager or an HR team can put together something that behaves a lot like a coach from these pieces. Some do, and it can work well for the people who keep it going.
That still leaves two things the features don’t change. The first is what memory holds. An assistant’s memory holds what you told it, which is useful context. It isn’t a record of how a manager is developing, because nobody set out to keep one. The second is the objective. An assistant’s default job doesn’t include knowing what this person is trying to change or checking whether it moved. That objective exists only if a user or a team sets it up and keeps it going. A general tool can’t assume that every person who opens it wants to be developed, so it would be odd if it did.
Whether these differences stay is a fair open question. Assistant features change from month to month, and some of what’s written here may look different within a year. What’s harder to change is the measure underneath. A tool scored on the answer it gives now will still be pulled toward the answer you like now, whatever gets added on top.
Which manager jobs belong to which tool?
If the two tools are rewarded for different things, the practical question is which job you’re doing. Here is one way to split common manager jobs.
| Manager job | Better fit | Why |
|---|---|---|
| Drafting a message or announcement | Assistant | The goal is a good draft now |
| Summarizing a long thread or document | Assistant | Speed and accuracy matter most |
| Explaining a policy or concept | Assistant | A clear answer is the whole job |
| Analyzing team data | Assistant | The work is the analysis itself |
| Researching a topic before a meeting | Assistant | Breadth and speed matter most |
| Rehearsing a hard conversation | Coach | The manager has to find the words |
| Building a habit over several weeks | Coach | Success is measured weeks later |
| Being challenged on your own view | Coach | An agreeable reply defeats the purpose |
| Noticing a pattern across weeks | Coach | It needs a goal someone is tracking |
| Preparing for a 1:1 | Both | Assistant for the facts and coach for the approach |
Most of these rows aren’t close calls. The ones that are tend to be jobs with two parts, like preparing for a 1:1. Pulling together what happened on a project is assistant work. Deciding how to raise a concern with the person across the table is coach work. A manager can do both in the same hour, with two different tools.
That’s why the decision for HR is a split of use rather than a ban or a single pick. Banning assistants takes away a tool that’s very good at half the list. Picking one tool for everything means some jobs land on a tool built for a different measure. A short rule your managers can remember does more: use the assistant when you need an answer, and use the coach when you need to change what you do. Some HR teams also weigh building their own coach on top of an assistant, and our comparison of building versus buying AI coaching walks through that choice.
So the question this post began with turns out not to have a single answer. “Which AI is better for our managers?” depends on which job they’re doing at the time. The more useful question for HR to answer is which job each tool is approved for.
How Risely approaches this
Risely built Merlin for the coach side of this split. We don’t expect managers to stop using assistants, and we don’t think they should. Drafting, summarizing and research belong with the assistant they already use. Merlin is for the other rows of the table, where success means a manager doing something differently weeks later. So our position is simple: use both, and give each one the job it was built for. If you want the direct comparisons, we’ve written them up for Risely and ChatGPT, Risely and Claude and Risely and Microsoft Copilot.