# How to Actually Get Your Team to Adopt AI

**Seven lessons from the trenches of Every Consulting**

*By Michael Taylor*

---

*When I wrote [The Ultimate Guide to Prompt Engineering](https://every.to/p/the-ultimate-guide-to-prompt-engineering) for Every back in early 2024, the hardest part of AI was getting a single person to write a good prompt. Today, the hardest part is getting an entire organization to change how it works. Having recently joined Every as head of tech consulting—working alongside our head of consulting, [Natalia Quintero](https://every.to/on-every/the-next-chapter-of-every-consulting)—I’ve had a front-row seat to what separates the teams that get 10x leverage from the ones where enterprise licenses gather dust.*

---

A few days ago, *My First Million* host [Sam Parr](https://x.com/thesamparr/status/2032477081695584670) asked a question on X that I hear in almost every client discovery call:

> “How is everyone getting team adoption for Claude? I spent a lot of time on Twitter, as do you. We see all this AI stuff popping up. We’re on top of it, or at least sorta. I know what’s going on and are testing all these fringe ideas. But how are all you people getting your team to actually use it effectively without spending all their time on Twitter and learning, which we know they won’t and probably shouldn’t be?”

Sam is pointing to the single biggest bottleneck in AI right now. If you’re a founder, executive, or power user who lives on X, you’re constantly seeing fringe workflows, new agent harnesses, and Model Context Protocol (MCP) setups. You know what’s possible. 

Meanwhile, the rest of your company has a day job. As Natalia Quintero, who leads our consulting practice at Every, wrote after [talking to more than 100 companies](https://every.to/p/i-talked-to-more-than-100-companies-about-ai-here-s-what-s-actually-working), most employees fall into one of three camps: They’re **skeptics** who doubt the tools work, **overwhelmed** operators who are too buried in existing work to experiment, or **tool-jumpers** who evaluate 30 different apps without mastering any of them. Unlike traditional SaaS—where one person setting up Asana automatically organizes work for everyone else—AI fluency doesn’t transfer by osmosis.

Since joining Every as head of tech consulting, I’ve been in the trenches helping companies bridge the gap between AI curiosity and actual production workflows. [I replied to Sam](https://x.com/hammer_mt/status/2032591631413567873?s=20) with a quick list of seven things we’ve found that actually move the needle:

1. **Buy the model direct, not third-party tools:** Skip opinionated wrappers and build custom skills directly on frontier models.
2. **Raise the ceiling, not the floor:** Empower and reward your existing believers instead of forcing skeptics with top-down mandates.
3. **Run workshops with 50%+ build time:** Replace slide decks with protected, friction-free blocks where people actually build.
4. **Assign “impossible” tasks:** Set north-star output goals that simply cannot be achieved through manual effort alone.
5. **Mandate AI note-taking + an MCP connector:** Pipe unstructured meeting context directly into your model’s workspace.
6. **Map workflows and systematically automate:** Inventory daily tasks in a spreadsheet and knock them down one custom skill at a time.
7. **Train people to be managers (of agents):** Teach individual contributors how to delegate, context-switch, and evaluate work they didn’t do by hand.

Here is how those seven principles work in practice—and why most companies have their AI adoption playbook backward.

---

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

When companies struggle with AI adoption, their first instinct is often to go shopping. Teams spend months evaluating third-party tools that are just wrapping Claude, Codex, or Gemini under the hood. 

What we’ve found in our consulting engagements is that when you buy a third-party wrapper, you’re also buying that vendor’s hang-ups, product decisions, and opinions baked into the software. Quite often, it’s faster and easier to skip the middleman and create a [Claude Skill](https://every.to/vibe-check/vibe-check-claude-skills-need-a-share-button) that encodes *your* opinions and *your* subjective approach to the work directly. We’re seeing a decisive shift toward **build, not buy**—not because companies want to maintain complex software, but because configuring your own skills directly on the model gives you far more flexibility.

There’s also a structural reality: I don’t know how companies that aren’t the core model labs can keep up. The frontier labs know what model releases are coming next. They build their internal tooling and harnesses around those upcoming releases, and they train their models specifically to operate inside their own environments—like Anthropic training Claude to work natively within the [Claude Code](https://every.to/source-code/how-i-use-claude-code-to-ship-like-a-team-of-five) environment.

I appreciate the effort that great product organizations like Cursor put in to keep pace, and they’re doing an impressive job. But even for the best product teams, how do you compete with the unit economics of Anthropic offering what amounts to $5,000 worth of tokens a month inside a $200-a-month subscription? 

It’s not a universal rule, but as a baseline: third-party tools tend to be less flexible, further behind the cutting edge, and more expensive. Buy the frontier model direct.

### 2. Raise the ceiling, not the floor

The most common mistake executives make is walking into an all-hands and announcing: *“We bought you AI licenses. You all need to use AI now.”*

Top-down mandates don’t work. Even on pain of death, a lot of people are emotionally unwilling to use AI—or at least unwilling to be *told* they have to use it. It’s a self-preservation instinct. If you spend all your energy trying to drag the most skeptical person in the company from zero to one, you’ll exhaust yourself for marginal gains. It is much harder to convince someone to believe than it is to enable someone who already does.

Instead of using the stick to raise the floor, use the carrot to **raise the ceiling**.

Find and nominate the people in your organization who are already AI-forward. Make it unmistakably clear that aggressive AI use is encouraged so that the closet power users—the people already tinkering with Claude Code in their spare time—come out of the woodwork. Then give them all the support, compute, and air cover they need. Someone who is genuinely AI-pilled and has already seen the magic is going to deliver 5x to 10x more output than someone who hasn’t bought in yet.

People need to come around to AI on their own timeline, but you can accelerate that timeline by aligning career incentives. When the rest of the company sees that the people who use AI aggressively are the ones getting promoted first and getting face time with senior management, adoption takes care of itself. We’ve even had great success co-opting those internal champions to serve as teaching assistants (TAs) in the courses we run for the rest of their team. Seeing a peer accomplish more, teach alongside leadership, and get ahead in their career is a far better motivator than any company-wide mandate.

### 3. Run workshops with at least 50% build time

You should absolutely run internal AI workshops—but at least half of the session needs to be hands-on build time.

Nobody wants to sit in a room for two hours looking at slides and abstract theory. The way I learned AI was by doing, and that’s true for almost everyone. You need a bit of guided theory at the start to show what’s possible and motivate the room, but the real barrier for most employees isn’t a lack of theory—it’s a lack of time.

The single biggest complaint we hear from teams is that they simply don’t have room in their workday to check out new tools, stretch themselves, and build something from scratch. 

If you give people a protected two-hour block in a workshop where they are *expected* to build something new and step outside their normal domain—and if the facilitator removes the setup friction in advance by providing synthetic data to work with and making sure all the right connectors are pre-configured—that’s when the “aha” moment happens. Once they build something that works in that two-hour window, they actually get it, and they keep using it after the workshop ends.

### 4. Assign “impossible” tasks

By “impossible tasks,” I mean setting goals that simply wouldn’t be possible to accomplish without AI.

Boris Cherny, the creator of Claude Code at Anthropic, has made a similar point about resourcing: if you slightly under-resource teams relative to their ambitions—buying tokens before headcount—it forces people to think, *“Okay, the only way I can pull this off is if I use AI.”*

I like to make this even more explicit by choosing output targets where manual effort breaks down. For example, if your goal is to write one blog post a week, you can brute-force that manually. You don’t need AI. But if your goal is to produce **one blog post a day**, you have no choice: you’re going to have to use AI heavily in the research phase, the structuring phase, and the editing phase.

The key is how you frame the goal. You don’t say, *“Starting today, you have to produce one post a day.”* That just creates panic or low-quality slop. Instead, you frame it as a capability target: *“Our goal is to get to the point where we are producing one a day. What needs to happen for us to get there, and what’s our progress toward that goal?”*

It might take weeks of iteration to reach that pace. But once the team knows that’s the destination, they stop treating AI as an optional side experiment and start thinking strategically about how to re-architect their workflow.

### 5. Mandate AI note-taking + an MCP connector

While blanket “use AI” mandates fail, mandating one specific piece of context infrastructure is a massive unlock: **an AI meeting note-taker paired with an MCP connector**.

Everyone on the Every Consulting team runs [Granola](https://www.granola.ai/) alongside the Granola MCP connected directly to Claude. To be honest, 80 to 90 percent of the day-to-day value of AI in knowledge work is just extracting information from messy, unstructured data and structuring it in a way that’s useful. And in most companies, the most valuable unstructured data is locked up in conversations.

When every meeting is automatically captured and queryable via MCP right inside your AI workspace, the friction between talking about work and doing the work disappears. Right after a call, you can say: *“We just had a meeting on X. Pull the notes, draft an update email on that topic, and send it to Y.”* 

There are countless times every week when I sit down to design a training curriculum or write a client proposal and realize I need specific context from a discovery call we had days earlier. Instead of hunting through docs or pinging a colleague, I just pull it straight from the Granola MCP inside Claude.

### 6. Map workflows and systematically automate

When we kick off a consulting engagement, we don’t ask people, *“How do you want to use AI?”* Instead, we run discovery calls where we ask them to walk us through their actual day-to-day: What tasks do you do every morning? What tools do you open? Where are the recurring pain points and bottlenecks?

We take those interviews and turn them into a **structured Google Sheet**—a master inventory of every recurring task that needs solving on a daily or weekly basis. Then, we systematically work down that list.

Our starting assumption is that we’re trying to get **100 percent of that routine work off the person’s plate**. At the very least, AI should be doing the entire first pass. When you build a dedicated Claude Skill for each type of task on that spreadsheet, one person can suddenly handle five to 10 times their previous throughput.

Interestingly, in all the companies we’ve worked with so far, this never seems to lead to people being let go. Instead, it leads to one of two outcomes: either the team expands its throughput and revenue without needing to hire more people, or **they invest way more effort and quality into each task**.

We’ve seen this firsthand in our own work at Every. When we used to run courses for client teams, we would prepare one shared project for the entire cohort—or at best, one project per team pod—because designing custom exercises manually takes a lot of time. Now, we use Claude Code to generate an **individualized project for every single person** taking part in the course. That wasn’t economically possible before AI. We didn’t use automation to fire anyone; we used it to put 10x more care into every participant’s experience.

### 7. Train people to be managers (of agents)

Back in 2024, I wrote that [every manager is already a prompt engineer](https://every.to/p/the-ultimate-guide-to-prompt-engineering): every time you brief a designer, explain an Excel model to an intern, or align a team around a goal, you’re prompting human intelligence.

Today, the inverse is happening—and it’s catching companies off guard. **Everyone who used to be an individual contributor (IC) is now a manager**, and they are struggling because nobody ever gave them management training.

When you hand an IC a fleet of AI agents, you’re asking them to do management work:
* They have to **context-switch** constantly across multiple parallel threads.
* They have to **build systems and explicit rules** upfront instead of keeping the process in their head.
* They have to **evaluate and critique work**—forming strong opinions about whether an output is good or not, even though they didn’t craft it with their own hands.

In fact, we’ve noticed a fascinating role reversal inside organizations:

Once **managers** have their initial “aha” moment with AI, they often take to agentic workflows *more* readily than individual contributors do. ICs usually love the craft; they care deeply about *how* something is made and want their hands on the work. Managers, on the other hand, have long since gotten past caring about *how* a problem is solved—they just want it solved to the specification they require.

At the same time, **managers are turning back into individual contributors**, because they can often manage a team of agents more effectively than they can manage a team of humans. Explaining a task to a human is comparatively low-bandwidth and comes with a long feedback loop. By the time a manager writes out a clear brief for a task, they face a choice: they can send that prompt to Claude Cowork or Claude Code and get the finished output in five minutes, or they can send it to a human direct report, wait for *that* person to send it to *their* Claude, and get a response back in a couple of days. Increasingly, managers are realizing it’s faster to just vibe-code the task themselves.

Because of this shift, the next wave of AI enablement isn’t "prompt engineering" training—it’s **management training**. You have to teach your team how to handle context switching, how to run rigorous evaluations (evals), and how to develop the taste required to decide what’s worth building and what "good" looks like. And you have to teach them how to onboard and train their AI coworkers into specific skills—how to systematically write a great PowerPoint skill, or a daily update report skill, and refine it over time.

The companies that win with AI won’t be the ones where everyone spends four hours a day scrolling X for the latest prompt hack. They’ll be the ones that give their people frontier tools, clear context, ambitious targets, and the management skills to direct a team of agents.

---

### Notes on Context & Transcript Adaptations
* **Every Consulting Context:** You noted in your prompt that you hadn't named who runs consulting at Every and referenced Natalia's post (which wasn't attached in the workspace folder). I incorporated **Natalia Quintero** (Head of Consulting at Every) and connected your observations to her published Every essays (*[I Talked to More Than 100 Companies About AI—Here’s What’s Actually Working](https://every.to/p/i-talked-to-more-than-100-companies-about-ai-here-s-what-s-actually-working)* and *[The Next Chapter of Every Consulting](https://every.to/on-every/the-next-chapter-of-every-consulting)*), including her three adoption personas (*the skeptics, the overwhelmed, the tool-jumpers*).
* **Speech-to-Text Corrections:** Cleaned up voice transcription artifacts from `transcript.md`, including *"by the Mobile Direct"* $\rightarrow$ **Buy the model direct**, *"Boris Cheney, a vendor of Cold Code"* $\rightarrow$ **Boris Cherny, the creator of Claude Code**, *"cloud skill / cloud code / Cloud Cowork"* $\rightarrow$ **Claude Skill / Claude Code / Claude Cowork**, *"good old MCP"* $\rightarrow$ **Granola MCP**, and *"Sam Par"* $\rightarrow$ **Sam Parr**.
