# You Don't Get Autonomous Agents on Day One > Source: > Published: 2026-06-11 00:00:00+00:00 · Charlie Holland · [Leadership ](/category/leadership) · 7 min read # You Don't Get Autonomous Agents on Day One The board has seen the demo and wants the autonomous agent live by Q3. That is not how any of this works. Here is the map I actually use to deliver AI: where it goes, and how to reach it without becoming one of the 40% of projects that get cancelled. Somewhere in your organisation, right now, a slide deck promises a fully autonomous AI agent running a core process by the end of the quarter. The board has seen the demo. The demo was magnificent. The demo was also a lie of omission, because nobody on that call mentioned the eighteen months of unglamorous work that sits between a magnificent demo and a thing that actually runs your business. This is most of the reason [Gartner expects 40% of agentic AI projects to be cancelled by 2027](https://www.gartner.com/en/newsroom/press-releases/2025-03-03-gartner-predicts-2025). Not because the technology doesn’t work. The technology mostly works. Projects die because organisations buy the destination and refuse to take the journey. They try to start where they meant to finish. I deliver this stuff for a living, and the failures are almost never technical. They are failures of sequencing. So here is the map I actually use. It answers two questions, in order: where does AI go, and how do you get it there without faceplanting in front of the people who approved the budget. ## Where AI actually goes: Talk, Build, Sense Strip away the hype and every enterprise use of AI lands in one of three buckets. I call them Talk, Build, and Sense. The only reason it matters is that each one needs completely different plumbing, and treating them as one problem is how good money gets set on fire. **Talk to customers.** Support, sales, onboarding. The conversational front door. This is the most constrained and most measurable thing you can point an agent at, because a conversation has defined inputs, defined outputs, and an outcome you can score. Resolution rate, CSAT, cost per interaction, all available from day one. That is exactly why it is the right place to start. Not because it is easy, but because it is the one place you can prove anything at all. **Build things.** Software, content, design. The bucket where individual productivity gains are loudest and organisational ones are hardest to pin down. A developer with a good coding assistant feels twice as fast. Whether the team ships twice as much is a different question, and usually a more disappointing one. Build is real, but measuring it honestly is where most of the work hides. **Make sense of things.** Analytics, reporting, knowledge. The broadest bucket and the slipperiest, because the hard part is not generating an answer. It is knowing whether the answer was any good. An agent that confidently summarises your data is easy. An agent whose summary you would bet a decision on is not. The mistake I see most often is reaching for a single, do-everything agent that spans all three. It encodes none of them well, because each bucket demands its own knowledge, its own guardrails, and its own definition of “good.” Pick the bucket with the clearest payback and the cleanest measurement, which for most businesses is Talk, and earn the right to the others. ## How you get there: Assist, Execute, Operate Once you know where the agent goes, you have to decide how much you let it off the lead. There are three settings, and they are a ladder, not a menu. You climb them in order. **Assist.** The human does the work, the AI helps. It drafts the reply, surfaces the relevant article, pre-fills the form. The human reviews, edits, sends. Nobody puts this stage on a conference slide because it looks unambitious. It is also where the entire foundation gets built, and I will come back to why. **Execute.** The AI does the work, a human reviews it before it reaches the customer. Crucially, you only promote a task type to Execute once your Assist-stage data proves the agent performs at or above your humans on that specific task. Everything else stays at Assist. And if quality on an Execute task drops below your baseline, it drops back to Assist automatically. Not after a steering committee. Not when someone notices. A guardrail does it, the moment the numbers turn. **Operate.** The AI runs the process end to end, humans oversee the exceptions. This is the autonomy everyone asked for on day one, and the only way to arrive here in one piece is to have climbed the first two rungs. You reach Operate with proven numbers, trained escalation paths, and measurement that catches a problem in hours rather than the following quarter. The whole discipline fits in one sentence: earn autonomy, don’t grant it. Two things get organisations killed on this ladder. The first is the dip. [Brynjolfsson’s Productivity J-Curve](https://www.nber.org/papers/w31161) is the well-documented pattern where a general-purpose technology costs you before it pays you, because the gains arrive only after you have done the complementary work. Teams that read the early dip as failure pull the plug roughly three months before the curve would have turned upward. The second is skipping straight to Operate. [Klarna](https://www.bbc.com/news/articles/c4gep52j9n4o) is the cautionary tale everyone now quotes: a headline-grabbing leap to autonomous customer service, followed by a quiet walk-back and rehiring. The lesson is not “AI doesn’t work.” It is “you do not get to skip the staging.” They tried to start at the top of the ladder and discovered the rungs were load-bearing. ## The boring part nobody wants to pay for Underneath all of it sits the least exciting sentence I will write today. The actual work is getting what your best people know out of their heads and writing it down. Not the documented process. The documented process is the easy sixty percent, and a competent agent will execute it on the first afternoon. The work is the other forty. The undocumented part. The enterprise customer who gets a more generous refund than the policy allows because nobody wrote that rule down but every senior agent knows it. The dispute that should be auto-approved because of a billing migration bug. The word “cancellation” that quietly turns a billing query into a retention risk. None of that is written anywhere. All of it is the job. Research keeps putting [a number near 42% on the share of organisational expertise that lives in one person’s head and nowhere else](https://en.wikipedia.org/wiki/Polanyi%27s_paradox). That is your real backlog. Every rung of the ladder is, underneath, a codification exercise. Assist exists precisely so that each time a human corrects the agent, you capture a judgment the model never shipped with. Execute is you trusting that captured judgment on the tasks where the data has earned it. Operate is what you get when enough of it has been captured to run without a hand on the wheel. It is slow, it is political, and the people who hold the knowledge are not always delighted to hand it over. It is also the entire job. The model is the cheap part now. The codification is the work. ## Where this leaves you So if you are delivering a project this quarter, the advice is unglamorous and it is the advice that works. Pick one bucket. Start at Assist. Earn each rung with measurement instead of optimism. Do the boring codification that nobody wants to fund. Do that, and you land in the small minority of AI projects that reach production at all. But I want to leave you with the larger thing hiding inside that boring sentence, because it has been nagging at me. Getting institutional knowledge out of people’s heads and into a system is not a problem AI invented. It is the oldest project in enterprise IT. Every wave did it. ERP, CRM, the document management platform I once wrote a rather long book about, the data warehouse, RPA. Different decade, different brochure, identical work underneath: take what people know, write it down, hand it to the machine. AI is simply the first tool sharp enough to reach the knowledge we always called too skilled to automate. That is excellent news if your job is shipping a project by Q3. It is a more uncomfortable thought if you sit back and ask what happens to the people once the knowledge is finally, completely, out of their heads. That is the next post.