
Over the summer I had a lot of conversations with executives at mid-market companies, mostly CEOs, CFOs, CMOs, and COOs running privately held businesses. Different industries, different tools and very different levels of AI maturity.. but almost every one of them came back to the same handful of questions.
How much should we budget for AI next year? We bought licenses and nobody uses them, so what went wrong? Where do we actually start? And how do we keep this from going sideways once it's touching customers?
After enough of those conversations I figured I should write the answers down once instead of giving a slightly different version every time. That turned into the 2027 AI Strategy Playbook, 14 chapters and 55 pages written for leadership teams planning AI into next year's budget.
It's free, and you can grab the full playbook here (it lands in your inbox). Below are 4 pages from it that are useful on their own, even if you never download the rest.
1. The most common problem is a tool nobody opens
Most companies I talk to are early in adoption. There's scattered personal use, a few power users who've figured out something great, and nothing that works the same way across the company. That's normal, and it's fixable.
The symptom I see most often is a tool the company already pays for that nobody opens. For a lot of folks using Microsoft that's Copilot, and for others it's a ChatGPT or Claude rollout from earlier this year. People log in a handful of times, ask it a question with no context like it's a search box, get a weak answer, and decide AI isn't for them.
From the playbook:
“"The gap between what a company licenses and what its people open is the most common thing I walk into, and it's rarely a tool problem."”
What usually goes wrong is the plan was built on a guess. Leadership assumes people are further along than they are (or not as far), and the budget follows the guess. So the first move in the playbook is measuring the room before you plan or spend anything. Don't go by how people rate their own AI skill, because almost everyone is either too generous or too hard on themselves. A short anonymous survey works well, and there's one in chapter 04 you can copy.
2. Score yourself before you plan anything
It's a 12-box self-assessment that takes a few minutes and costs nothing, and it's the first page of the playbook I'd hand a leadership team.
You check each statement that's true for your company today, and nobody's grading it but you. The boxes are split into 3 groups:
One person at a time. Things like "Everyone who needs one has a company-approved AI assistant, on a company account" and "Most people with a seat use it every week."
The whole team. Things like "We know what our power users have already built" and "At least one AI workflow runs with a named person approving its output."
Leadership. Things like "We have a written data policy: what can go in a prompt and what can't" and "Our leadership team can say what AI is for in one sentence."
The first 2 groups map to the two ladders I've written about before (I've shared this in past newsletters), i.e., single player AI where one person gets good at it, and multiplayer AI where the whole team shares the same skills/context/workflows.
Then you read your score. 0-4 is getting started, 5-8 is pockets of use, and 9-12 is ready to scale. Each band points you to the chapter to start with, so a 3 and a 10 get very different next steps:
Getting started (0-4). Pick one assistant, connect it to your email/files and train people on their own work.
Pockets of use (5-8). Find what your power users have already built and share the best of it with everyone else.
Ready to scale (9-12). Plan real bets with an owner and a measure attached to each one.
Most teams I talk to land in the first two bands, and that's really where the cheapest wins are.

The way I'd run it is to have each person on the leadership team score it on their own, then compare. The disagreements are the really helpful part. If the CEO checks "our leadership team can say what AI is for in one sentence" and then everyone writes down their sentence and you get 4 different answers, you've learned a ton in just a few minutes.
3. Keep the blast radius small
The question I get right after "where do we start" is some version of "how do we make sure this doesn't blow up on us." It's a fair question, especially once AI is drafting things a customer will see.
“"The governance model that works is embracing AI, but keeping the blast radius small."”
AI output is only useful when people trust it enough to act on it, and that trust comes from how you design the work around it. So you start with AI drafting and a person reviewing, editing and approving. Each workflow earns more autonomy on its own record, and when a team trusts the reviewed version, that workflow can start running on its own.
Treat it like a capable new hire. Someone hired off the street would ask what the work is, where things live and how to tell if it's good, and an agent needs the same answers. You write those into a short document it reads every time (usually called a context file), then you check its work the way you'd check a junior person's, especially early on.
One of the operations examples in the playbook is orders that arrive by email in every format and get re-keyed into the ERP by hand. AI reads the incoming orders and drafts the ERP entry, and the order team checks each draft before it's saved, the way they'd check a new hire's work. What it needs first is the order rules on paper, because today they often live in a few people's heads, and the ERP stays where it is. You know it's working when minutes per order and corrections per draft both go down.
Accuracy comes from the work around the model. That means testing it against answers you already know (including the hardest questions your experts can think of), retesting when your source content changes, and a few firm guardrails that don't bend.
Every new workflow goes through an 8-box trust checklist before it goes live, and these are 3 of the boxes:
Math and business rules run in fixed calculations the AI calls. The model handles language and judgment, and your pricing formula stays a formula.
Access starts scoped and read-only, with no delete. Turning on write access is a decision someone signs.
Before it ships, name who owns AI-written work when it's wrong.
It's the same way we run our own AI teammates at StealthX. They start read-only, a person approves anything before it goes out, and they earn more room one workflow at a time.


Wrapping up
If you scored a 3 out of 12 on the self-assessment, that's not a reason to panic. Most companies I talk to land in the first two bands, so hold onto your score. The rest of the playbook picks up with the question every leadership team asks first: what do we budget? Chapter 03 is about finding the money, chapter 05 is a one-page AI plan for your 2027 budget and chapter 06 is the first 90 days.
You can get the full 2027 AI Strategy Playbook here. It's free and it lands in your inbox as a pdf.
Putting this into action
Have each person on your leadership team score the 12 boxes on their own this week, then compare answers.
Pick the riskiest AI workflow on your list and walk it through the trust checklist with the person who'll own it.
What I'm reading
What Is Jev? The AI Model That Doesn't Generate Text. IBM's explainer on Jev, a decision model from TypeSafe AI that picks from answers you define and tells you how confident it is instead of writing paragraphs. Same idea as the trust checklist above.. let the AI make the judgment call and keep the business rules in fixed code.
Open-Weight vs Closed AI Models for Enterprise. If someone on your team is pitching an open model to save money, read this first. The argument is that the quality gap has mostly closed, so the real decision is how much control you want over cost, data and change.
Notion: your buttons just got a brain. Notion buttons can now message your AI agents, so anyone on the team can kick off a workflow with one click. That's a shared workflow (the "whole team" boxes in the self-assessment) living in a tool a lot of teams already use.
AI Just Crossed the Terrifying Line - Now What?. Kurzgesagt (the animated science channel) walks through the Hugging Face incident for a general audience. Good one to send the person on your team who's still deciding how worried to be.
Onward & upward 🤘
Drew




