Thinking in Public Finding the Collisions Before the Rollout

Finding the Collisions Before the Rollout.

When Everyone Has an AI Strategy, in sixteen minutes, for people who implement AI. AI Practitioners Summit 2026.

The room

The AI Practitioners Summit is a five-day virtual summit for AI consultants, agency owners, and founders – the people who build and sell AI implementation. I recorded this session for it: sixteen minutes, slides on screen, me in the corner. It's the base cut of a talk I gave two days later, at forty minutes, to a room of IT and security leaders at Cyber Cheyenne. Same six people, same five questions. That version asks what the collisions look like from the security seat. This one asks what they look like when you're the one who signed the statement of work. Dan Fazio and the team at Legacy Labs cut a thirty-second highlights reel from it.

Six people, six strategies

It opens with six people at a fintech company, and one question to hold for each: smart use of AI, risky, or both? Sam designs a new logo in Nano Banana. Scott builds a prospecting machine on Clay and n8n. Roger, the CFO, drops the ERP forecast into his personal Google Drive and asks Gemini for the trends. Heather gives a small team of OpenClaw agents write access to Asana and Notion. Lisa's team wires Claude Code into the deployment pipeline. And Mary, the CEO, builds an internal tool in Lovable one Saturday, then tells everyone to experiment, and to use good judgment.

For every one of them, the honest answer is both. And if you implement AI for a living, you've met Mary. She's probably the one who signs your statement of work.

The part nobody scoped

Notice what didn't break in that story: the technology. The logo looked great, the agents wrote good reports, the code shipped. What fails is how the tools meet each other, and the people around them. For anyone selling the implementation, that's what stalls the rollout, and it shows up later as rework nobody wants to pay for.

“If you sell AI implementation, Mary's company is your client. And those six collisions are the part of the engagement that nobody scoped.”

Five questions for the first meeting

AI governance is how an organization decides how AI is allowed to be used. A handful of decisions, each with an owner. Five questions tell you whether a company has made them, and you can ask all five in your first meeting: what AI tools are in use, what data they can reach, where human review is required, who approves new tools, and who owns the outcomes.

For each one, the talk gives what a good answer sounds like, because that's what you're listening for. The first is a list, with owners and an honest gap. The second is boring in the best way: agents on their own service accounts, able to read the CRM but not export from it, and no personal account touching company data. Roger's company can't say that sentence. The third names the specific points where a person checks the work, because “everywhere” is nowhere in particular. The fourth is one intake, someone from IT or security on every review, and an answer within a week. The fifth puts a person's name next to every tool from question one.

Then the question that surprises people: who pays when it goes wrong? Most cyber insurance was written for a stranger breaking in, and Heather's agents aren't strangers. Somebody gave them the keys. That can push the loss toward errors and omissions, and some E&O renewals are coming back with AI exclusions attached. Don't be surprised if nobody in the room knows which policy would respond.

Without losing the engagement

The last part is for the consultant: how to raise all this without losing the work. Ask early, because these are discovery questions. In week one, “walk me through what's already in use and what it can touch” sounds like diligence. In month three, after the agent wrote the wrong thing into the CRM, it sounds like blame. Tie every collision to the project you're delivering: “your governance is a mess” puts a client on the defensive, and “before we connect this to your CRM, I need to know who owns it and what it writes” puts you on their side. Treat governance as a guardrail on a mountain road, the thing that lets you take the curve at speed.

And know when it isn't yours to fix. When every leader gives you a different answer to question five and none of them seem surprised, the leadership team doesn't agree on what AI is for. Treat that as an implementation gap and it will cost you the project.

I've put the five questions on one page, with what each answer tells you. The full transcript is below.

If you're implementing AI for a company right now, bring your company or a client’s to a thirty-minute conversation. I'll walk the five questions against it with you and tell you where it's likely to break.

Book a conversation
Transcript 16 min · 16 sections

From the talk’s caption track. Remarks from the room appear in this lighter tone. Timestamps jump to that moment in the video.

  1. Six people at a fintech company 0:07
  2. Mary, and “use good judgment” 3:00
  3. Six AI strategies in one company 3:42
  4. Herding Wolves, and AI as another wolf 5:04
  5. The part nobody scoped 5:39
  6. AI governance in one sentence 6:06
  7. 1. What AI tools are in use? 6:54
  8. 2. What data can they reach? 8:05
  9. 3. Where is human review required? 9:31
  10. 4. Who approves new tools? 10:45
  11. 5. Who owns the outcomes? 12:02
  12. Who pays when it goes wrong 12:35
  13. Saying it without losing the engagement 13:26
  14. Know when it isn’t yours to fix 14:31
  15. The five questions on one page 15:00
  16. Mary’s company, one last time 15:26

0:07 Six people at a fintech company

I want to introduce you to six people at a fintech company. As you meet each one, keep one question in mind. Smart use of AI? Risky? Or both? Hold on to your answers. This is Sam. Sam runs brand strategy and design. A few weeks ago, Sam opened Google's Nano Banana and spent an afternoon on a new logo concept. The leadership team loved it. It's already on the pitch deck.

There's one small detail. The company may have a hard time protecting it. The U.S. Copyright Office ties copyright to human authorship. And nobody wrote down how much of that logo was Sam and how much was the model. This is Scott. Scott runs sales. A few months ago, Scott found Clay and n8n. And over a couple of weekends, he built a prospecting machine. Every morning, it finds new companies, enriches the contacts, pushes the best ones into the CRM, and drafts a first outreach email with AI.

Pipeline is up. His team loves it. Scott isn't entirely sure where all that enrichment data comes from, but it works. This is Roger, the CFO. Every Monday, Roger exports the forecast from the ERP. Revenue projections, hiring plans, cost models. He drops the spreadsheets into Google Drive and asks Gemini to find the trends and flag the risks. It saves him half a day a week. It's his personal Google account because that's the one he already had open.

Roger assumes the data stays private. He's not entirely sure. This is Heather. Heather is a project manager, and she discovered OpenClaw. She built a small team of AI agents connected to Asana and Notion. They read every project update, summarize progress, and write the weekly status report for the leadership team. Nobody has missed a status report since. To keep everything current, Heather gave the agents API keys with write access.

So when something looks blocked, they update the task themselves. This is Lisa, the director of engineering. Her team has gone all in on Claude Code. Developers use it to write large parts of every new feature, and they've wired it into their CI/CD pipeline. Some changes are now written by the model, committed to the repository, and deployed automatically. The pipeline doesn't care who wrote the change, and the review step has gotten pretty thin because everything keeps passing.

Velocity has never been better.

3:00 Mary, and “use good judgment”

And this is Mary. Mary is the CEO. A few weeks ago, Mary sat down with Lovable. In about 30 minutes, she built an internal tool the team had been talking about for months. Mary has never written a line of code in her life until that Saturday. And she thought, if I can do this, imagine what the team could do. So at the next all-hands meeting, she stood up and said, "AI is moving really fast.

If you find tools that help you work faster or smarter, experiment with them. Just use good judgment." Mary wasn't being irresponsible. She was being a normal CEO.

3:42 Six AI strategies in one company

So, smart, risky, or both? For every one of these, the honest answer is both. If you implement AI for a living, you've met Mary. She's probably the one who signs your statement of work. Let's look at her company again. Sam, Scott, Roger, Heather, Lisa, Mary. None of these people talk to each other. And none of them talk to IT. This isn't six people experimenting with AI. This is six different AI strategies inside the same company.

A brand strategy with no answer on ownership. A sales strategy built on data nobody can trace. A finance strategy running on a personal account. An agent strategy with write access. An engineering strategy with a review step nobody designed. And a cultural strategy that amounts to "use good judgment." Here's the part worth paying attention to. Every one of these people was trying to help the company. And every one of them succeeded.

Individually. Nobody in this story is careless. They're the smart ones. The people who moved fast. AI doesn't create chaos in organizations. It just exposes the chaos that was already there.

5:04 Herding Wolves, and AI as another wolf

I'm Korey Mercier. I specialize in Herding Wolves. I help leaders negotiate organizational change. And lately, most of that change shows up when AI starts disrupting how the business actually works. AI is just another wolf. Fast, capable, and usually let into the pack before anyone agreed which way the pack was running. Here's why that matters to you. If you sell AI implementation, Mary's company is your client. And those six collisions are the part of the engagement that nobody scoped.

5:39 The part nobody scoped

Notice what didn't break in that story. The technology. The logo looked great. The agents wrote good reports. The code shipped. But a fair coin doesn't make a fair game. The tools all worked. What fails is how they meet each other and the people around them. That's what stalls your rollout. And that's what shows up later as rework nobody wants to pay for.

6:06 AI governance in one sentence

So here's one sentence you can take with you. AI governance is how an organization decides how AI is allowed to be used. That's it. A handful of decisions that somebody owns. When a company does have an AI policy, it's usually written after the tools are already in place. A policy written on top of that gap reads well and changes nothing. Governance is the set of decisions underneath it.

Without it, every employee writes their own rules for AI. Mary's company has six rule books and nobody has read the other five. There is a fast way to tell whether a company has made those decisions. Five questions. You can ask all of them in your first meeting with them.

6:54 1. What AI tools are in use?

Question one. What AI tools are in use? Ask Mary and she'll give you a number. It'll be wrong and she won't know it. She can't see it all from where she sits. Look in three places instead. Tools people signed up for on their own. AI that got switched on inside software the company already had approved. And AI inside whatever the company sells. Bryan Kaplan, a CISO I presented with recently, described one company that went looking and found more than 40 of these.

So if leadership answers with confidence, treat it as a guess. And here's the catch. Finding the real answer takes more trust than people expect. The worst thing a company can do is tell everyone "no AI." They'll keep using the tools. They'll just stop telling you. A good answer to this sounds like a list with owners and an honest gap. "Here's what we've approved. Here's what we found when we asked.

We haven't looked inside the CRM yet." That company can actually see itself.

8:05 2. What data can they reach?

Question two. What data can they reach? This is Roger, Scott, and Heather. Roger handed the company's forecast to an account the company doesn't control. That's reach. Scott's automation pulls in personal data from sources he can't name and writes it straight into the CRM every morning. Heather handed her agents the keys to Asana and Notion. That's reach with hands. The useful question is what the tool can see, change, send, or delete.

A chatbot that answers questions and an agent connected to your email, your CRM, and your file share are different risks, even when they come from the same vendor. If you give an agent your inbox, you've given it your password resets and your two-factor codes. Don't do that. Guardrails define what a tool supposedly won't do. Access controls define what it can't do. You want both. Only one of them holds when someone is actively trying to get around them.

A good answer here is boring in a way. "The agents run under their own service accounts. They can read the CRM, but they can't export from it. No personal account touches company data." Roger's company can't say that sentence.

9:31 3. Where is human review required?

Question three. Where is human review required? That's Lisa and Heather. Code that the model writes, commits, and ships with a review step nobody actually sits at. Agents that can change a task status on their own. If the answer is "everywhere," it's nowhere in particular. Human in the loop is a stopgap that laziness erodes over time. The first 50 approvals get read carefully. The next 500 just get clicked through.

People like to call AI an intern. I think of it more as a teenager. It follows the letter of the law and nothing more. It thinks in black and white. So ask where, specifically, a person checks the work and whether that check was designed or just assumed. A good answer names these points. "Nothing reaches production without a person approving the change. Nothing reaches a customer without a person reading it.

Agents can suggest a status change, and a person makes it." Those three sentences – Mary's company can't produce one of them.

10:45 4. Who approves new tools?

Question four. Who approves new tools? Mary's answer was, "We use good judgment." That is the approval process. And Scott's automation is what that process produces. A new business process running every morning, pulling data from sources nobody checked, that nobody ever approved. And it runs under Scott's own login. The day Scott leaves or IT disables his account, a process the sales team depends on just stops. And nobody knows why.

What you're listening for here is a path to yes. Without one, shadow AI is inevitable. And watch the order of operations. Too often, the business picks a tool, then asks IT to make it safe. That's backwards. IT and security belong in the selection process. Because they ask the questions the business doesn't know to ask. A good answer is short. "Requests go through one intake. Someone from IT or security is on every review.

Most requests get an answer within a week." The speed matters here as much as the rule. If a yes takes three months, people will route around it.

12:02 5. Who owns the outcomes?

Question five. Who owns the outcomes? Let's start with Sam's logo. Who owns it? And who answers if it turns out the company can't protect it? Then Heather's agent marking the wrong tool as blocked. And Lisa's 2 a.m. deploy that breaks checkout for a customer. Whose name is on those? If the answer is "the vendor" or "the AI," the answer is nobody. A good answer puts a person's name next to every tool from question one.

12:35 Who pays when it goes wrong

Most cyber insurance policies were written for a stranger breaking in. Heather's agents aren't strangers. Somebody gave them the keys. When a tool the company authorized causes the loss, there may be no break-in to claim against. That can push the loss towards technology errors and omissions instead. And most E&O renewals are coming back with AI exclusions attached. So ask, if one of these tools causes a loss, which policy responds?

Cybersecurity? E&O? Or neither? Don't be surprised if no one in the room knows. If a company can't answer these five questions, it has AI experimentation. Governance starts when somebody owns the answers.

13:26 Saying it without losing the engagement

So how do you say that to a client without losing the engagement? First, ask early. These are discovery questions. Asked in the first meeting, they make you look careful. Asked after something breaks, they make you look like the person who broke it. In discovery, it sounds like ordinary diligence. "Before we scope this, walk me through what's already in use and what it can touch." Nobody gets defensive about that in week one.

Ask in month three after the agent wrote the wrong thing into the CRM, and it begins to sound like blame. Second, tie every collision to the project you're delivering. "Your governance is a mess" Puts the client on the defensive. "Before we connect this to your CRM, I need to know who owns it and what it writes" puts you on their side. Nobody fires the consultant who protected the launch.

Third, keep the purpose straight. Governance works like a guardrail on a mountain road. It's what lets you take that curve at speed. The goal here is safe speed.

14:31 Know when it isn’t yours to fix

And last, know when it isn't yours to fix. Sometimes the five questions surface something bigger. Mary and her leadership team don't actually agree on what AI is for. Treat that as an implementation gap, and it will cost you the project. You'll recognize this when every leader gives you a different answer to question five, and none of them seem surprised. That's usually the moment to bring someone else in.

15:00 The five questions on one page

I've put these five questions on one page with what each answer tells you. This QR code will let you download it. And if you're implementing AI for a company right now, or you want a second set of eyes, bring one company, yours or a client's, to a 30-minute conversation. I'll walk through these five questions with you and tell you where it's likely to break.

15:26 Mary’s company, one last time

So let's go back to Mary's company one more time. Six smart people. Six good intentions. Six accidental AI strategies. AI adoption is going to keep happening through experimentation. That's not the problem. AI doesn't need less experimentation. It needs better leadership.