# How to Get Your Team to Actually Use AI

Buying your team access to Claude doesn’t give them the time to learn it. They still have deadlines, meetings, and a way of doing their job that already works. Experimenting with AI is another task on their list, and it’s easy to put off until next week.

Sam Parr [asked how people are getting their teams to adopt Claude](https://x.com/thesamparr/status/2032477081695584670). He spends a lot of time on Twitter keeping up with AI, but he doesn’t expect his employees to do the same. How do you get people to use these tools effectively without making following AI news part of their job?

I recently joined Every as head of tech consulting, and this is a problem we’re working on with companies. The biggest obstacle we hear about is that people haven’t had time to try the tools properly. Buying subscriptions is straightforward. Helping someone discover a useful way to change how they work takes more effort.

I [replied with seven things I’ve learned so far](https://x.com/hammer_mt/status/2032591631413567873). Here’s what they look like in practice.

## 1. Buy the model direct, not third-party tools

My default is to start with the model provider’s own tools. If you want to use Claude, start with Claude before evaluating a dozen products that use it under the hood.

Every tool comes with its maker’s opinions about how work should happen. Those opinions can be useful if they match yours. But if they don’t, you spend time working around decisions you didn’t make. Often, I find it quicker to create a Claude skill—a reusable set of instructions for a particular task—that describes how I want the work done.

Take a daily update report. You already have opinions about what belongs in it, how it should be structured, and what your team needs to know. You can put those instructions into a skill and change them as you learn what works. That gives you more control than adapting your process to someone else’s product.

I also think it’s hard to compete with the model companies on their own tools. They know what capabilities are coming and can develop the product and model together. I appreciate the work companies like Cursor do to keep up, but I’m reluctant to add another layer unless it solves a specific problem.

This leaves plenty of room for specialized software. We use Granola for meeting notes, for example. My preference is to buy tools where they add something useful, and build our own instructions where the important ingredient is our judgment about how to do the work.

## 2. Raise the ceiling, not the floor

It’s tempting to announce that everyone needs to use AI now. You’ve bought the subscriptions, so surely people should use them.

But some people are emotionally resistant to AI. They may worry about their job, care deeply about doing the work themselves, or simply haven’t had an experience that made the tools feel useful. A mandate doesn’t necessarily change any of that.

I’ve found it more productive to start with the people who already believe. There’s often someone experimenting with Claude Code in their spare time who could do much more with support from the company. Make it clear that this is encouraged, give them access to the tools, and help them apply what they’re learning to real work.

We’ve had success bringing these people into our courses as teaching assistants. They get recognition for something they care about, and their colleagues get help from someone who understands the business. It also makes the benefits of learning AI visible inside the company.

The incentives matter. If people who use AI well get more responsibility, time with senior management, and opportunities for promotion, their coworkers have a reason to pay attention. Seeing someone you work with accomplish more is more persuasive than another announcement about how important AI is.

You get an immediate benefit from supporting the enthusiasts, too. You don’t have to convince them to start experimenting. They’re already doing it.

## 3. Make workshops at least 50 percent build time

People don’t learn to use AI just by watching someone else use it. They need to try something, see where it fails, and work out how to get a better result.

That’s why I think at least half of an AI workshop should be spent building. Some theory helps people understand what’s possible, but the workshop also needs to give them something they’re missing in their normal workday: permission to spend a couple of hours learning.

Prepare for that time properly. Make sure everyone has access to the tools. If the exercise needs a connector to company information, get that working beforehand. If people can’t use real data, provide synthetic data they can practice with. You don’t want someone spending their entire build session trying to get access to the thing they’re supposed to be learning.

Then give them a project that stretches them beyond what they’d normally do. The useful moment is when someone builds something they didn’t think they could build. They’ve now seen the tool work on a problem they understand, and they have a reason to try it again.

## 4. Assign impossible tasks

If your goal is to write one blog post a week, you can probably keep doing it manually. If your goal is to publish one a day, you’ll need to rethink the process.

That’s what I mean by an impossible task: a goal you couldn’t reasonably achieve with your existing way of working. It forces you to look for a different approach instead of fitting AI into the occasional spare five minutes.

You have to be careful about how you set the goal. Telling someone they need to publish daily starting tomorrow is just giving them more work. Instead, make daily publishing a capability you’re trying to develop together. What would need to change? Could AI help with research? Could it do an editing pass? Where is most of the time going, and what experiments would help you find out?

It might take time to get there. The progress you’re looking for is a better process that makes the target achievable while maintaining the quality you need.

Without that goal, it’s easy to keep doing the work the familiar way. An ambitious target gives experimentation a purpose.

## 5. Make AI meeting notes and an MCP connector mandatory

One of the most useful things we’ve done on the consulting team is give everyone Granola and its MCP connector. The connector lets our AI tools retrieve the meeting notes when we need them.

That sounds like a small change, but it removes a problem I run into constantly. I’ll start working on a proposal or putting together a curriculum and realize that something we discussed in a meeting is essential context. Without accessible notes, I have to reconstruct the conversation or ask someone else what was said. With the connector, I can pull that information into the task.

The same applies to follow-up emails. After a meeting, I can ask AI to draft an update based on what we discussed, then review it before sending. I don’t have to brief it from memory first.

So much of the value I get from AI comes from extracting information from unstructured material and putting it into a useful form. Meeting notes are a particularly good source because they contain decisions, constraints, and preferences that might never make it into a formal document.

This is a narrower requirement than telling everyone to use AI for everything: establish a shared source of context that makes everyone’s tools more useful. Set appropriate boundaries around which meetings are captured and who can access the notes. Within those boundaries, make capturing and connecting the information the default.

## 6. Map workflows and systematically automate them

In our discovery calls, we ask people what they do day to day. What tools do they use? Which tasks repeat? Where do they get stuck?

We turn that into a structured Google Sheet, then work down the list. For each task, we look at what AI could take on and what the person would still need to do. Sometimes we can automate a step. Sometimes the useful result is a skill that produces a first pass for someone to review.

I like to start with the ambition of getting everything off the person’s plate, then work out where human involvement is actually needed. That makes you examine the whole job instead of stopping after you’ve found one easy use case. It doesn’t mean handing over responsibility for the result.

So far, in the work I’ve seen, this hasn’t led to people being let go. It has led to people putting more effort into each task, or teams increasing their output without needing to hire at the same rate.

Our courses are a good example. We used to prepare one project for an entire cohort, or perhaps one project per team. Now we can use Claude Code to create an individual project for each participant.

We’re using the extra capacity to make the course more useful to each person. Preparing that many individual projects wouldn’t have been practical before. If we only measured how much time we saved preparing the original exercise, we’d miss the more interesting opportunity.

## 7. Train people to be managers of agents

When an individual contributor starts delegating work to AI, they take on some of the responsibilities of a manager. They need to explain the task, supply context, define what good looks like, and evaluate work they didn’t produce themselves.

Many people haven’t been trained to do that. They’re used to knowing whether something is good because they worked through every step personally. Now they have to judge a result without having made all the decisions that led to it.

This is why I think management training belongs in AI training. People need to learn how to brief, delegate, review, and give useful feedback. They need to handle switching between tasks and decide which work is worth delegating in the first place.

They also need to turn their judgment into instructions the AI can use again. If you know what makes a good PowerPoint presentation, can you explain it well enough to create a reusable skill? Can you evaluate the output, identify what went wrong, and improve the instructions? Those are skills you can practice, just as you can practice giving feedback to a colleague.

There’s a funny reversal happening on the management side, too. Some managers become more hands-on because they can now do work directly through agents. Once they’ve written a clear brief, they can give it to Claude and start reviewing a result, rather than sending it to a person who may then use Claude themselves.

Managers can take to this surprisingly well once they’ve had their initial breakthrough. They’re already accustomed to specifying an outcome and judging whether it meets their requirements. Individual contributors who love the craft may find delegation harder, even if they’re more comfortable with the technology.

The ability to direct work is becoming useful to more people, regardless of their title.

## Give people a reason to keep using it

Your team shouldn’t need to follow every model release to benefit from AI. They need working tools, access to the relevant context, and time to discover how those tools apply to their job. They also need a goal that makes changing their process worthwhile.

That’s what I’m learning at Every. The people already experimenting need support to go further. The people who haven’t started need a chance to build something useful. And both groups need to learn how to judge the work they delegate.

Once someone has done something they couldn’t have done before—whether that’s building a tool or creating a project for every person in a workshop—they have their own reason to keep learning. They don’t need to borrow your enthusiasm from Twitter.

---

*Editorial note: This draft uses the supplied transcript, tweet text, and your writing sample. I assumed a long-form post with the seven learnings as its structure. Natalia’s consulting post was not present in the supplied files, so I haven’t added background from it or named who runs consulting. External browsing was unavailable; the tweet links are included as supplied, and the draft omits the transcript’s unverified pricing comparison and third-party quotation.*
