Steve from accounting just built an LLM-powered deck builder. It generates stakeholder presentations on demand, thank-you slide included. Leadership loves it. Steve didn’t ask anyone, didn’t tell anyone, and can’t tell you what it costs to run.
Steve deserves credit. Self-starters drive most of the early wins in any AI program. But races have lanes for a reason. Gather a crowd in the middle of a field, shout “go,” and anyone can declare themselves finished at any moment, after covering any distance in any direction. You can’t time that race. You can’t improve it. You can’t even say who won.
Most companies run AI adoption exactly that way, then wonder why the results won’t add up.
What unbounded experimentation costs
I’ve been laid off twice in two years, and both related to AI. Both times, I watched the real cause build in the months before. Leadership funded open-ended experimentation with no defined course, no finish line, and a loss of quantifiable impact. When push came to shove, nobody could show what the money bought, what we had to show was more features that read more as pivots than compounding impact.
The symptoms repeat across the industry. Scattered tools. Duplicate licenses. Pilots that never graduate. Teams split between people afraid to touch AI and people quietly running prompts in the shadows.
The fear in those teams is rational, and it deserves a straight answer. Every failure I’ve watched up close traces back to the operating model around the technology. Operating models are fixable.
The real stake is accountability
The task list will keep shifting. Models already draft specs, generate layouts, summarize research, and critique copy, and the list grows every quarter. Chasing it is a losing game.
Judgment and accountability hold steady. Someone decides what ships, owns what it does in the world, and answers for it when it goes wrong. Those layers belong to people.
In many organizations, nobody holds them for AI work. A model writes the release notes, a PM approves them on a skim, and a customer-facing error ships with no one’s name on it. Speed without an owner is a liability that hasn’t been invoiced yet.
So the executive question becomes: who owns each decision AI touches, and do they have the authority and the time to own it well?
Define the course before anyone runs
A useful AI vision reads like a course map: where the race starts, where it ends, and which ground stays off-limits. Before funding another pilot, leadership should answer five questions in writing.
What outcome are we buying? Faster output, less manual overhead, more time for high-judgment work, or better decision quality. Each one implies different tooling, org design, and training.
What stays off-limits, and why? Customer commitments, hiring decisions, regulated content, final calls on brand voice. Edges give the course its shape.
Who owns each decision AI touches? One name per decision, with real authority, including the authority to kill it.
What does a healthy team look like twelve months in? Tools change behavior. Bad tools change it faster. Describe the behavior you want before the tools describe it for you.
How will we measure progress? Pick one or two measures per stage: hours reclaimed from repeat work, tools consolidated, pilots graduated to production, escalations traced to AI output.
If you can’t answer these, you don’t have a strategy yet. You have a field and a starting pistol.
Calibrate the course to the organization
Size shapes the course, and it’s one input among several. Risk exposure, regulatory load, and AI maturity matter just as much. A three-person team at a regulated fintech needs governance on day one. A forty-person team at a seed-stage startup can run looser lanes.
With that caveat, size gives a reasonable starting point.
Team size
Where to start
What breaks without a course
Small (under 10)
Name the repeat work eating everyone’s week. Build a shared prompt library. Give AI use one owner.
Everyone builds a private toolkit, and nothing transfers when someone leaves.
Mid-size (10 to 50)
Write a playbook of what you use AI for and how. Connect AI tools to the design system. Run cross-functional pilots, each with an owner and an end date.
Pilots multiply and none graduate. Three teams pay for the same tool.
Large (50+)
Set ethics and governance guidelines from day one. Consolidate tooling across departments. Open sandboxes that report learnings to a central team. Train people until AI stops feeling like a mystery.
Twelve writing assistants, twelve brand voices, and a compliance problem nobody saw coming.
Vetting your AI direction and implementation
I built an interactive framework around the nine questions where AI adoption most often stalls: the mandate with no workflow, the decision nobody owns, the spec that drifts across three disciplines, and six more. Each card names the failure mode and the moves that fix it.
Pressure-test your AI decision.
Pick the question closest to where you're stuck. Take the framework with you.
Alignment
The usual failure mode is a mandate with no workflow attached. Three disciplines hear “go” and move in three directions.
The framework
01Turn the mandate into one written problem statement the whole team signs.
02Agree on what good and done mean before anyone builds.
03State the single user problem the AI solves, in one shared sentence.
04Name where AI is off the table for now, so the scope has edges.
Roles
When everyone owns it, no one does. The feature stalls, or it ships as three contradictory versions stitched together.
The framework
01Give each AI decision one named owner. Committees don't ship.
02PM owns the problem and the bar. Design owns the experience and the controls. Engineering owns model behavior and guardrails.
03Give the owner real authority, including the authority to kill it.
04Write down who decides before the first disagreement shows up.
Handoffs
An AI feature changes meaning every time it crosses a hand-off. The PM's intent, the designer's flow, and the engineer's build drift apart quietly until the thing that ships isn't what anyone agreed to.
The framework
01Carry the user problem through every hand-off, from spec to build to review.
02Prototype the AI behavior before the hand-off so engineering builds against something real.
03Review the built feature against the original problem. Tickets drift.
04When the model surprises you, loop design and PM back in before shipping.
Adopt?
AI belongs where it makes the product measurably better for the person using it. Trend pressure is not a use case.
The framework
01Name the specific user problem this solves.
02Say how AI beats the alternative you already have.
03Can't do either? You're shipping novelty. Stop here.
Control
People want tools that feel predictable and in their hands. They also want the boring work handled. Those pull in opposite directions, so decide it deliberately.
The framework
01Default to user control for anything people want to own.
02Let the AI act on its own only in low-risk, reversible cases.
03Always show what the AI is doing.
04Make every AI action easy to undo.
Grounding
A generative feature is only as reliable as what you feed it. Grounding is what keeps a confident model from inventing answers.
The framework
01Feed it current, domain-specific context (RAG).
02Never let it answer from nothing.
03Show the source behind each output.
04Test the failure cases as hard as the happy path.
Presenting
AI output needs its receipts attached. Without them, people either over-trust it or dismiss it, and both are expensive.
The framework
01Where it came from: the data source.
02How confident it is.
03What it's based on: the rationale.
04No recommendation ships without all three.
Oversight
Human oversight is the thing you protect, even when it's the slow part. Pull it out to save time and you'll spend the savings, with interest, on the cleanup.
The framework
01Keep a human at every irreversible decision.
02Route borderline cases to a person before the AI acts.
03Treat AI output like a sharp colleague's first draft.
04Verify surprising insights against the source data.
Testing
Find out whether people want the help before you spend a quarter on the model. Then prototype fast and throw most of it away.
The framework
01Fake it first: a human behind the curtain (Wizard of Oz).
02Watch how much autonomy people actually tolerate.
03Then generate three versions with AI tooling, keep one.
04Treat generated prototypes as starting points, then finish them yourself.
Pick the question closest to where you're stuck. Take the framework with you.
Alignment
The usual failure mode is a mandate with no workflow attached. Three disciplines hear “go” and move in three directions.
The framework
01Turn the mandate into one written problem statement the whole team signs.
02Agree on what good and done mean before anyone builds.
03State the single user problem the AI solves, in one shared sentence.
04Name where AI is off the table for now, so the scope has edges.
Roles
When everyone owns it, no one does. The feature stalls, or it ships as three contradictory versions stitched together.
The framework
01Give each AI decision one named owner. Committees don't ship.
02PM owns the problem and the bar. Design owns the experience and the controls. Engineering owns model behavior and guardrails.
03Give the owner real authority, including the authority to kill it.
04Write down who decides before the first disagreement shows up.
Handoffs
An AI feature changes meaning every time it crosses a hand-off. The PM's intent, the designer's flow, and the engineer's build drift apart quietly until the thing that ships isn't what anyone agreed to.
The framework
01Carry the user problem through every hand-off, from spec to build to review.
02Prototype the AI behavior before the hand-off so engineering builds against something real.
03Review the built feature against the original problem. Tickets drift.
04When the model surprises you, loop design and PM back in before shipping.
Adopt?
AI belongs where it makes the product measurably better for the person using it. Trend pressure is not a use case.
The framework
01Name the specific user problem this solves.
02Say how AI beats the alternative you already have.
03Can't do either? You're shipping novelty. Stop here.
Control
People want tools that feel predictable and in their hands. They also want the boring work handled. Those pull in opposite directions, so decide it deliberately.
The framework
01Default to user control for anything people want to own.
02Let the AI act on its own only in low-risk, reversible cases.
03Always show what the AI is doing.
04Make every AI action easy to undo.
Grounding
A generative feature is only as reliable as what you feed it. Grounding is what keeps a confident model from inventing answers.
The framework
01Feed it current, domain-specific context (RAG).
02Never let it answer from nothing.
03Show the source behind each output.
04Test the failure cases as hard as the happy path.
Presenting
AI output needs its receipts attached. Without them, people either over-trust it or dismiss it, and both are expensive.
The framework
01Where it came from: the data source.
02How confident it is.
03What it's based on: the rationale.
04No recommendation ships without all three.
Oversight
Human oversight is the thing you protect, even when it's the slow part. Pull it out to save time and you'll spend the savings, with interest, on the cleanup.
The framework
01Keep a human at every irreversible decision.
02Route borderline cases to a person before the AI acts.
03Treat AI output like a sharp colleague's first draft.
04Verify surprising insights against the source data.
Testing
Find out whether people want the help before you spend a quarter on the model. Then prototype fast and throw most of it away.
The framework
01Fake it first: a human behind the curtain (Wizard of Oz).
02Watch how much autonomy people actually tolerate.
03Then generate three versions with AI tooling, keep one.
04Treat generated prototypes as starting points, then finish them yourself.
Pick the question closest to where you're stuck. Take the framework with you.
Alignment
The usual failure mode is a mandate with no workflow attached. Three disciplines hear “go” and move in three directions.
The framework
01Turn the mandate into one written problem statement the whole team signs.
02Agree on what good and done mean before anyone builds.
03State the single user problem the AI solves, in one shared sentence.
04Name where AI is off the table for now, so the scope has edges.
Roles
When everyone owns it, no one does. The feature stalls, or it ships as three contradictory versions stitched together.
The framework
01Give each AI decision one named owner. Committees don't ship.
02PM owns the problem and the bar. Design owns the experience and the controls. Engineering owns model behavior and guardrails.
03Give the owner real authority, including the authority to kill it.
04Write down who decides before the first disagreement shows up.
Handoffs
An AI feature changes meaning every time it crosses a hand-off. The PM's intent, the designer's flow, and the engineer's build drift apart quietly until the thing that ships isn't what anyone agreed to.
The framework
01Carry the user problem through every hand-off, from spec to build to review.
02Prototype the AI behavior before the hand-off so engineering builds against something real.
03Review the built feature against the original problem. Tickets drift.
04When the model surprises you, loop design and PM back in before shipping.
Adopt?
AI belongs where it makes the product measurably better for the person using it. Trend pressure is not a use case.
The framework
01Name the specific user problem this solves.
02Say how AI beats the alternative you already have.
03Can't do either? You're shipping novelty. Stop here.
Control
People want tools that feel predictable and in their hands. They also want the boring work handled. Those pull in opposite directions, so decide it deliberately.
The framework
01Default to user control for anything people want to own.
02Let the AI act on its own only in low-risk, reversible cases.
03Always show what the AI is doing.
04Make every AI action easy to undo.
Grounding
A generative feature is only as reliable as what you feed it. Grounding is what keeps a confident model from inventing answers.
The framework
01Feed it current, domain-specific context (RAG).
02Never let it answer from nothing.
03Show the source behind each output.
04Test the failure cases as hard as the happy path.
Presenting
AI output needs its receipts attached. Without them, people either over-trust it or dismiss it, and both are expensive.
The framework
01Where it came from: the data source.
02How confident it is.
03What it's based on: the rationale.
04No recommendation ships without all three.
Oversight
Human oversight is the thing you protect, even when it's the slow part. Pull it out to save time and you'll spend the savings, with interest, on the cleanup.
The framework
01Keep a human at every irreversible decision.
02Route borderline cases to a person before the AI acts.
03Treat AI output like a sharp colleague's first draft.
04Verify surprising insights against the source data.
Testing
Find out whether people want the help before you spend a quarter on the model. Then prototype fast and throw most of it away.
The framework
01Fake it first: a human behind the curtain (Wizard of Oz).
02Watch how much autonomy people actually tolerate.
03Then generate three versions with AI tooling, keep one.
04Treat generated prototypes as starting points, then finish them yourself.
Every executive asks the same follow-up: how do we consolidate tools when the market changes monthly? Standardize in Q1 and you lock in whatever led in Q1. Let every team choose and you pay for twelve writing assistants.
No universal answer exists, and anyone selling one is selling a vendor. The right balance depends on your risk tolerance, data sensitivity, procurement cycle, and how fast your teams actually learn. Three questions start the conversation:
Which capabilities need a single standard (data handling, security review, brand voice), and which can stay open?
How often will you revisit tool decisions, and who holds the authority to switch?
What does a pilot need to prove before it graduates, and what retires it?
This is the work I do with leadership teams. We map the course for your business, name the owners, and set the measures, so the next budget review has answers.
The finish line
AI lets everyone run. That’s the opportunity and the risk in one sentence. Leaders who draw the course, name the owners, and define the finish line get a team that gets faster every quarter and can prove it. Leaders who fire the starting pistol and walk away get a field full of Steves, each one declaring victory in a different direction.
