AI Adoption
The Agent With No Job Description
By John J. BakerJuly 21, 20266 min read
The pitch landing in your inbox this year is different from the one you ignored last year. Last year the offer was a tool. You would open a chat window, type something, and get help. Whether you used it was up to you, and mostly you did not, because there was always something more urgent than learning a text box.
This year the offer is a colleague. The software you already pay for now ships with agents that have job titles. Intuit sells a roster inside QuickBooks: an Accounting Agent, a Payments Agent, a Customer Agent, a Marketing Agent, a Payroll Agent. ServiceTitan describes itself as an agentic operating system for a contracting business. Procore has agents for RFIs, submittals, and daily logs. Nobody is asking you to learn anything. They are asking you to delegate.
It is a good pitch, and the products behind it are often real. It also quietly removes the one question that decides whether any of this works.
Nearly everyone has adopted. Almost nobody has integrated.
Almost everyone has adopted. Almost nobody has integrated.
A tool is open somewhere in the business. Someone is getting value from it.
It is part of how the work actually gets done, not a thing a few people do on the side.
Goldman Sachs 10,000 Small Businesses Voices, published March 2026. 1,256 respondents, surveyed by Babson College and David Binder Research. The sample is drawn from Goldman’s own program alumni, who are more growth minded and better resourced than small businesses generally, so read the adoption number as a ceiling.
Read those two bars again, because they are measuring the same businesses. Adoption is close to universal. Integration is close to absent. Whatever is happening in that 62 point gap is the actual story of AI in small business right now, and it is not a story about capability.
The same survey found that 73 percent of these owners say more training and support would help them implement, and that 87 percent describe AI as augmenting their people rather than replacing them. So the owners are not confused about what the technology is for, and they are not resistant to it. They are stuck somewhere between having it and using it.
I would treat the adoption figure as a ceiling. The sample comes from Goldman's own small business program, which skews toward owners already inclined to try things. The gap is probably wider in the general population, not narrower.
An agent with a title and no job cannot be evaluated
Here is what the delegation framing costs you. When you hire a person, the job description comes first. You know what the role is responsible for before anyone starts, which means at the end of the first month you can say whether it is working. That sentence is doing a lot of work: you can say whether it is working.
An agent sold by job title arrives with the title and without the description. It is responsible for accounting, or for marketing, or for the customer. Ask what specifically it is supposed to remove from your week and there is no answer in the product, because the product was deliberately scoped wide enough to appeal to every business that might buy it.
Could someone tell you on Friday whether it worked?
- A category
“We are adopting AI.”
Nothing here can be true or false on Friday. There is no work described.
- A headcount
“We are putting an AI agent in the office.”
Names a worker, not a job. You cannot review the performance of someone with no assigned duties.
- A department
“We are automating estimating.”
Closer, and still too wide to check. Estimating is fifteen tasks, and this commits to none of them.
- A named problem
“Nobody retypes an approved estimate into the accounting system.”
Answerable by Friday, by asking one person. If it is wrong you can see how, and you can turn it off.
Look at the difference between the top three rungs and the bottom one. The top three are all things an owner can honestly say they are doing. None of them can be wrong. There is no Friday on which someone walks in and reports that adopting AI did not happen this week, because the sentence does not describe an event.
The bottom rung describes an event. One person used to retype approved estimates into the accounting system, and now either they do or they do not. That is checkable by asking them. It is correctable, because when it fails you can see the specific place it failed. And it is reversible, because you know exactly what you would be turning off.
This is the same discipline behind building a box around the AI: a system you can reason about when it surprises you is worth more than a system with a broader mandate. Scope is not a limitation you accept reluctantly. It is the thing that makes the result knowable.
The firm that sells the roadmap is telling you to slow down
If you have been feeling behind, this part matters.
In May 2026, Gartner issued a warning about what it called agent washing: vendors relabeling conventional automation as agentic, which it said "obscures meaningful distinctions and increases risks of misaligned investments and extended vendor lock-in." Its analysts went further, saying that vendors claiming end to end autonomous operation before 2027 are overstating what the technology can currently do, and that most capability marketed as agentic today amounts to interpreting questions, making recommendations, and holding a conversation.
Their guidance was this, from senior director analyst Jan Snoeckx: "The priority today is not full autonomy, but building the operational discipline, architectural flexibility, and decision frameworks that allow agentic AI to scale as the technology matures."
That is a research firm whose business is selling companies their technology roadmap, telling those companies that the discipline comes before the autonomy. Gartner had already predicted, in June 2025, that more than 40 percent of agentic AI projects would be canceled by the end of 2027, citing costs, unclear value, and weak risk controls.
You have probably also seen the figure that 95 percent of AI pilots deliver no measurable financial impact. It comes from an MIT project in 2025 and it is worth knowing that the number is contested. The sample was small and not random, and "no measurable impact" is not the same finding as "failed." I mention it because it gets quoted at owners constantly, usually by someone selling the cure. The honest version of that number is narrower and more useful: pilots that sit beside the work rather than inside it tend not to show up in the results.
None of this says the technology is hollow. It says the gap between the pitch and the delivery is currently wide, and that the people best positioned to know are advising patience about scope rather than enthusiasm about scale.
What the strongest trade data actually shows
In construction, the numbers are moving in the right direction. Commercial contractors reporting measurable business impact from AI went from 17 percent to 38 percent in a year, with the gains concentrated in cost estimating, budgeting, and bid management. That is a real doubling and it is worth noticing.
It also means 62 percent report no measurable impact. Same survey, same year, and a more accurate description of the average experience. The report comes from ServiceTitan, a vendor with an obvious interest in the finding, though it was fielded by an independent research firm across more than a thousand commercial construction leaders.
What separates the 38 from the 62 is not budget and it is not which model they picked. It is whether anyone named the thing that was supposed to change.
Name the problem, then decide if AI is even the answer
The useful conversation does not start with AI at all. It starts with two questions that have nothing to do with technology.
Where does your week actually go, and which of those hours would you buy back first if you could?
Most owners already know the answer and have never said it out loud, because it sounds too small to bother anyone with. Someone rekeys the same numbers into a second system. Quotes go out and nobody follows up until the client follows up first. The person who knows how a job gets priced is the only person who knows, and every question routes through them.
Those are the right size. They are small enough to be named, which makes them small enough to be finished. And when one of them is genuinely gone, you have something more valuable than a working automation: you have a team that watched the first one work, which is the only reason the second one gets adopted.
Sometimes the answer is not AI. A rekeying problem is often an integration problem, and the honest recommendation is a connector and no model at all. That is a better outcome than an agent, and you only get to find it out by naming the problem before choosing the tool.
Where ClearOak comes in
We train teams and build automations, and the work always starts in the same place: finding the specific problems worth removing, in order, and being clear about which ones are not worth touching.
In practice that means a conversation about where the time goes and what it is costing you, in either revenue you are not capturing or margin you are giving away. Not all three at once. Pick the two that are actually true for your business right now, and the sequencing tends to become obvious.
You are not behind. The businesses in that 14 percent did not get there by adopting harder. They got there by removing one thing, watching it hold, and then removing the next one.
Schedule a call and bring one problem, described the way you would describe it to a new employee on their first day. We will work out together whether it is worth removing, and whether AI is the right way to remove it.
Prefer email? Reach me directly at john@clearoakconsulting.com and tell me where your week goes. That is usually enough to find the first one.