97% of companies deployed AI agents. Most can't get value.

Here’s what we’re reading and thinking about this week:

Last week we watched someone demo agentic workflows on stage and call them experimental.

We have been running that same kind of work inside our own company for months. Real tasks, real responsibility, nobody double-checking every step.

The gap between "experimental" and "we depend on this" was not the technology. We both had the same tools.

Here is the strange part of this moment. AI adoption just hit a record. Okta ran a study showing 97% of leaders say they rolled out AI agents this year, and most are already in active use. And yet 79% of companies say they are struggling to get value out of it, and more than half of executives say the rollout is straining their company.

Yes this is a tech-oriented security company, so the numbers may be a bit skewed. But still….

Record adoption. Low payoff. Most people read that and assume they need a smarter model or a more independent agent.

That is the wrong read.

Our take: you do not have an AI problem. You have a sequencing problem. You are trying to climb the mountain in one leap.

There is an order to this. You climb it one step at a time, and almost everyone tries to jump straight to the top.

Level 1: Get every person using AI for repeatable use cases

Of course, we use Playbooking - taking a task you already do and writing it down step by step so AI can run it the same way every time.

But the key is to help people identify repeatable use cases, in plain words, that they can use on a daily basis.

It sounds basic. That is the point. This is where the value actually starts, and it is the step most people skip on the way to something shinier.

Level 2: Team and tools

Once individuals have their own use cases, this is where you can start connecting into shared workflows and team tools. If you skip to the tools first, best case you’ll hit a ceiling of feeling like AI is a fancy google. Worst case, you’ll have agents running amuck.

One operator we were hanging out with last week told us a story of how he connected his AI to their company’s task management tool, and without him realizing, it had gone and commented and caused confusion on a bunch of tasks another team was doing.

Totally controllable, but moral of the story: this opens a fresh set of problems, and it only works if level 1 is solid. Everyone on a team needs to be AI comfortable in order to get the full ROI of any bigger solution.

Level 3: Systems

This is the shiny object.

But if you start with your AI usage and build it up brick by brick … think the playbooks joining together into one system … that’s how you can get an AI that mostly runs itself.

“But can’t it learn instead of me needing to figure out the parts?” No, not reliably.

"But can the agent improve itself?" Yes, eventually.

But asking that before you’ve really identified and created repeat use cases for yourself (eg. playbooks at level 1) and then navigated integrating those into real tools and real team workflows, often stops you before you start.

At least that’s what we’ve seen with the hundreds of companies we’ve now worked with, and thousands of people we’ve trained.

The demo on stage was stuck because they reached for the top of the mountain.

The work we depend on is boring by comparison. We just climbed the steps in order.

Do this this week

Come start learning with us. Tomorrow we are running a free class on exactly this, and we will take a real task and turn it into a playbook AI can run, start to finish, so you can copy the moves.

The Playbook Method Masterclass. Wednesday, July 15 at 5pm ET. It is free.

What level are you on right now? Hit reply and tell us!

LINKS

For your reading list 📚

• Three frontier models shipped the same day this month. GPT-5.6, Grok 4.5, and Meta's Muse Spark all at once. The model race is sprinting…. but our whole point is that’s still not the bottleneck.

• Cisco is about to hand all 90,000 of its employees a personal AI agent. We'll be watching this for sure!

• Terence Tao revived his 1999 math applets with coding agents, and the agent found two bugs in his 27-year-old code along the way.

• Apple sued OpenAI over hiring more than 400 of its people, and Musk and Altman spent the weekend throwing elbows on X about it. We laughed when reading some of this 🙂 

That's all!

We'll see you again soon. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected]

Cheers,

🪄 The AMP Team (formerly: the AI Exchange Team)