Your board is going to ask about AI agents this year, if it has not already. Most organizations are not ready to build them. The good news: readiness is testable, teachable, and closer than you think. The bad news: you cannot buy it. Let me explain with a chicken dinner.
Open an AI tool and type this:
"Give me a chicken Milanese recipe."
You will get one. Competent, correct, complete. Bread the cutlet, fry the cutlet, squeeze the lemon. It is also the exact same recipe ten million strangers would get, and reading it feels like eating at an airport.
Now try what I typed instead.
"I came home last week from two weeks in Northern Italy and I cannot stop thinking about the chicken Milanese. Act as an Italian grandmother whose recipe has been handed down for generations. Teach it to me the way you would teach your granddaughter, the details nobody writes down. What to serve with it, which wine to open, what the kitchen should smell like when it is right."
Same tool. Same model. Same monthly subscription.
What came back was not a recipe. It was an evening. The cutlet pounded thin enough to see your knuckles through, because nonna says thickness is where amateurs fail. The breadcrumbs cut with grated Parmigiano. The butter-and-oil argument, settled the way her mother settled it. A bitter greens salad to cut the richness, a chilled Gavi, and a warning to salt the meat, never the crumbs.
The tool did not get smarter between those two prompts. I brought more to the ask.
Hold that gap in your mind, because your agent strategy lives inside it.
Somewhere in your industry right now, a vendor is showing a leadership team a demo of an AI agent doing a full day of work in four minutes. Researching, drafting, updating systems, sending the follow-up. The room goes quiet, someone asks about pricing, and a pilot gets approved before anyone asks the only question that predicts whether it will work.
Can anyone in this building explain that job the way the grandmother explains the Milanese?
Not describe it. Explain it. Every step, every judgment call, every detail nobody writes down, every point where someone should taste before the plate leaves the kitchen.
The numbers say almost nobody asks. Cisco found 85 percent of enterprises piloting AI agents this year. Five percent have them in production. The 80 points between those numbers are not a technology gap. They are the distance between the first prompt and the second.
An agent is someone else cooking dinner in your kitchen.
Strip away the terminology and that is what you are buying. An AI agent is delegation: a tireless new cook who will make the dish ten thousand times, exactly as instructed, including every instruction you got wrong.
Every executive already knows how this fails, because it is how delegation between humans fails. Hand a capable person a vague assignment and you get back something fast, confident, and not what you meant. The failure was never the person. It was the handoff.
Except a human cook has instincts. She smells the burning butter. She tastes the sauce and reaches for the salt. She stops and asks when something seems off. An agent does none of that unless someone taught it to, which means every instinct you rely on has to become an instruction before the kitchen is handed over.
This is why agent readiness has almost nothing to do with technology. It is a question of how well your organization knows its own recipes. And there is a ladder that gets you there. Four rungs. Most organizations are standing on the first one, pointing at the fourth.
Rung one: your people know how to ask.
The distance between those two Milanese prompts is a skill, and the industry keeps declaring it obsolete. The declaration is wrong in a way that costs organizations dearly.
Asking well means knowing what you want before you type, saying it plainly, recognizing when the answer is wrong or invented, and knowing what must never be typed into a public tool at all. Ordinary professional judgment, extended to a new medium.
Here is why this rung carries everything above it: a prompt is thinking made visible. The person who wrote the grandmother prompt understood the dinner they wanted. The person who typed "give me a recipe" had not decided yet, and the generic answer they got back was a mirror, not a failure. When your team's AI output is mediocre, the tool is usually reflecting the ask.
The test for this rung costs nothing. Watch your people work. Are they getting reliable results or getting lucky? Do they catch the confident nonsense? Do they know your rules about what data stays out? If there are no rules, that is also your answer about which rung you are on.
Rung two: your people can bring the trip to Italy.
Look again at what made the second prompt work. It was not politeness or a magic phrase. It was context. Where I had been, what I tasted, who should be teaching me, what mattered about the result. I gave the tool a world to cook inside.
Out of the box, the most capable model on earth produces competent generic, because generic is all it knows about you. It has never met your customers, your standards, your regulatory reality, your voice. Feed it nothing and you get the airport meal. Feed it your world and the output starts belonging to you. This rung is where the visible quality gains live, and it is a learnable skill, not a technical one.
I watched someone climb it in a single afternoon. A senior producer at an insurance client, a woman who had never once used an AI tool, sat down and dictated the quoting process she had run for years. It lived entirely in her head. No manual, no documentation, nothing written anywhere. The AI turned her dictation into a process document her colleagues could follow, and it was nearly perfect on the first pass, because she was not vague about a single step. She was the grandmother. She knew the recipe cold, and the machine amplified her clarity into something the whole kitchen could use.
Nearly perfect matters too. The draft contained one factual error about state insurance rules, and she caught it in seconds, the way nonna catches under-salted meat with one taste. That is the entire model of safe AI adoption in one scene: human clarity in, machine speed out, expert judgment on top.
The readiness question was never "is your team technical enough." It is "does your team know the work the way the grandmother knows the dish." Different question. Better answer. Fixable through teaching.
Rung three: the recipe gets written down.
Now the uncomfortable part of the grandmother story, the part every family knows. The recipe handed down through generations usually is not written anywhere. It lives in her hands. And the family finds out what that means the year she is gone and nobody can make Christmas taste right again.
My insurance client was living that story with a revenue process. Until that afternoon, the quoting workflow existed in exactly one place: one person's memory. If she had left, it left with her. Institutional knowledge in one person's head is not institutional knowledge. It is institutional risk with good attendance.
The third rung is where individual clarity becomes organizational capability: the recurring work written down, judgment calls included, clear enough that a new hire could follow it without a hallway conversation. The recipe card, written for someone who has never cooked the dish.
This is not bureaucratic housekeeping. In an agent-shaped economy it is the most valuable asset you can build, because a process document a new hire could follow is nine tenths of an agent specification. The steps, the inputs, the exceptions, the taste-tests: that is what an agent runs on. Organizations that wrote their recipes down will wire up agents in weeks when they choose to. Organizations that did not will discover, mid-build, that nobody can say how the dish is made. The project stalls, and everyone blames the oven.
Rung four: you hand over the kitchen.
Only here does the board conversation belong, and by now it is the easy conversation, because everything hard is done. The recipes are written. The people directing the agents can tell a good plate from a bad one. The rules about what stays out of the pot are habits, not posters.
One discipline remains, and you never outgrow it: a human tastes before the plate reaches the guests. Money moving, contracts signing, anything leaving your walls under your name. Those checkpoints are not training wheels. They are the reason the speed everywhere else is safe.
The test takes one afternoon.
Pick your highest-volume process. Ask the person who runs it to write the instructions so clearly that a new hire could execute them without asking a single question.
If a clear document comes back, you are closer to agents than most of the market, and your people deserve more credit than the readiness-assessment industry gives them. If what comes back is three bullet points and a shrug, you have learned the true state of things: the recipe lives in someone's hands, unwritten. Nothing wrong with that person. Everything wrong with the handoff you were about to attempt.
One exercise, zero dollars, and it measures the only thing agents run on.
The mistakes I keep watching smart leaders make.
Buying agents like software. Software works when installed. Agents work when specified. A budget approval is the start of the work, not the end.
Skipping rungs because the top one is exciting. Automation wired on top of unwritten, unspecified work does not skip the foundation. It scales a recipe nobody ever wrote, and serves the same confusion faster, to more people, with nobody tasting.
Dismissing the bottom rung as beneath investment. The typing may become optional someday. The clarity never will, and the bottom rung is where clarity gets trained.
Training by exposure: a lunch-and-learn, a license, a hope. Capability that matters gets assessed. A workshop your people sat through proves attendance. What you want is a level they can prove.
What this is about underneath.
I run a workforce AI literacy association alongside my consulting practice, and the reason is in everything above. The agent conversation keeps being framed as a technology race between companies. Underneath, it is a capability question about people.
Every worker who learns to ask with intent, bring the context, and judge the output becomes more valuable in an agent-shaped economy, not less. Every worker handed a license without teaching becomes a risk surface, and eventually a statistic in a company that concluded AI did not work. Same technology. The difference is whether anyone built the ladder.
Leaders get to decide which story their organization tells. That is not a burden. It is one of the clearest opportunities on your desk, because the readiness everyone is trying to buy is standing in your building right now, waiting to be taught.
The agent future is coming either way.
The organizations that thrive in it will be the ones that knew their own recipes.
