An agent can do the work. It cannot decide what the work should be.
The demo always goes well. A software agent reads an inbound email, checks the customer’s account, drafts a reply, updates the CRM, schedules the follow-up, and flags the invoice that is thirty days late. Nobody touches a keyboard. The room goes quiet in the way rooms do when people are quietly calculating how many salaries they just watched disappear.
Then the agent arrives at a real company. The inbound email goes to a shared inbox that three people monitor and nobody owns. The customer record exists in two systems that disagree about the billing address. The follow-up rule lives in the head of an office manager who has been there eleven years and does it differently for the customers she likes. The agent does exactly what it was configured to do, and within a month someone is cleaning up after it.
Nothing about that story is a technology failure. The agent worked. The business did not, and it had not been working for some time. The agent simply did the old, unclear work faster and in more places at once.
That is the conversation most companies need to have about agentic AI before they buy anything. It is also the conversation I have built my Lean AI® practice around.
What agentic AI actually is
Strip away the marketing and agentic AI is software that is given a goal instead of a single instruction. It decides the steps, uses other tools to carry them out, checks the result, and keeps going until the goal is met or it gets stuck.
The easiest way to see the difference is to compare it with the AI most people already use.
| Chatbot | AI assistant | AI agent | |
|---|---|---|---|
| You give it | A question | A task | A goal |
| It gives back | An answer | A draft for you to use | A finished outcome |
| Who decides the next step | You | You | The software |
| Touches your other systems | No | Rarely | Yes: email, CRM, calendar, billing |
| Example | “What are our payment terms?” | “Write a reminder to this client.” | “Collect on every invoice over 30 days.” |
That last column is the whole story. A chatbot that gives a wrong answer wastes a minute. An agent that takes a wrong action sends the email, changes the record, or issues the credit before anyone looks. Its value and its risk come from the same place: it acts.
It is also worth knowing that a lot of what is sold as “agentic” is not. Gartner has warned about what it calls agent washing, where older chatbots and automation tools are relabeled, and estimated that only about 130 of the thousands of vendors making agentic claims offer the real thing. The label on the box tells you very little. What the software is allowed to do on its own tells you everything.
Why so many agent projects stall
The enthusiasm is real. In a February 2026 survey of 500 senior executives at large companies, every single respondent said they planned to expand their use of agentic AI this year. The same market is also producing a lot of quiet retreats. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and the reasons it gives are rising costs, unclear business value, and weak risk controls.
Notice what is missing from that list. Nobody is cancelling these projects because the models cannot write an email or update a record. They are cancelling them because the business could not say what the agent was supposed to accomplish, what it was worth, or who was responsible when it got something wrong. Those are operations questions, and in my experience they almost always predate the AI.
When I look underneath a stalled agent project, I tend to find one of three things.
The agent automated waste. If a process has six approval steps and four of them exist because of a problem that was solved years ago, an agent will faithfully perform all six. You now pay for software to do work nobody needed. I wrote about this pattern in more detail in Why Most Businesses Automate Broken Processes, and agents make it more expensive because they run constantly and at volume.
The handoffs were never defined. Most operational failures happen between people, not inside one person’s job. Sales thinks onboarding starts at the signed contract; operations thinks it starts at the first payment. A human team papers over that gap with phone calls and goodwill. An agent has neither. It follows whichever version of the rule it was given and exposes the disagreement at scale.
Nobody owns the outcome. When a person makes a mistake, the business knows whose desk to walk to. When an agent makes one, the vendor points at the configuration, IT points at the business requirements, and the department head points at IT. An agent without a named human owner is not autonomous. It is unsupervised.
The Lean AI® approach to agents
I created the Lean AI® framework because I kept watching artificial intelligence amplify operational weakness instead of operational strength. Agents raise the stakes on that problem, because an agent does not wait for a person to notice the weakness before acting on it.
The principle is simple to say and harder to practice: fix the work first, then make it faster. Applied to agentic AI, that becomes a sequence, and the order matters more than any individual step.
- Diagnose the operation. Find where work actually slows down, where it gets redone, and where accountability is unclear. The goal is the single constraint that limits the business, not a list of everything that could be automated.
- Clarify the process. Write down how the work flows today, who owns each step, and what “done” means. If two people describe the same process differently, that disagreement has to be settled by a human before an agent inherits it.
- Remove the waste. Cut the steps, approvals, and reports that no longer serve a purpose. Every step you remove is one the agent never has to be built, tested, monitored, or paid for.
- Introduce the agent where it relieves the constraint. Put autonomy where it reduces labor, shortens a cycle, or improves a decision at the bottleneck. Anywhere else, it is decoration.
- Enable the team. People need to understand what the agent does, when to trust it, and how to override it. An agent the team works around is the most expensive tool in the building.
- Measure and refine. Track the business result, not the agent’s activity, and adjust as the operation changes.
Most agent projects start at step four. That is the vendor’s natural starting point, since step four is what they sell. It is rarely the business’s natural starting point, and the gap between the two is where the cancelled projects come from.
Two of the Lean AI® principles matter more for agents than for any other kind of AI. The first is that humans stay accountable: the agent can execute, but ownership and judgment remain with a named person. The second is that adoption matters more than capability. An agent that can do forty things and is trusted to do none of them has produced nothing.
Is this work ready for an agent?
Before any agent is configured, I ask leadership to answer six questions about the specific process in front of them. None of them are technical. All of them are predictive.
- Can you describe the process the same way twice? If the answer depends on who is asked, the process is not ready.
- Does every step have a named owner? A department is not an owner. A person is.
- Do you know what the process costs today? Hours, delays, errors, or lost revenue. Without a baseline there is no way to know whether the agent paid for itself.
- Is the data the agent will rely on trustworthy? If your team double-checks a report before they believe it, an agent will believe it without checking.
- What happens when it gets something wrong? Know who finds out, how fast, and how the action is reversed.
- Is this the constraint? Speeding up a step that is not the bottleneck produces a faster pile of work waiting in front of the step that is.
A business that can answer all six with confidence is ready for an agent. A business that cannot has found its real project, and it is usually worth more than the agent would have been.
There is a second judgment to make even when a process passes: how much freedom the agent gets. I use two questions for that. How often does this work happen, and how costly is a mistake? The answers sort most tasks into one of four places.
[embed: node/1574f89d-1d24]
The lower right is where agents earn their keep. The upper right is where most of the expensive mistakes happen, because the volume tempts people to remove the human signature before the agent has earned it.
The guardrails that make autonomy safe
Autonomy is not a switch. It is a dial, and it should be turned up only as an agent earns trust in a specific job. Four guardrails make that possible, and none of them require an enterprise budget.
A named human owner for every agent. Not the vendor and not IT. The owner is the person accountable for the outcome the agent produces, with the authority to pause it.
Checkpoints sized to the risk. An agent can send a meeting confirmation without approval. It should not issue a refund, change a price, or contact a regulator without one. Decide in advance which actions need a human signature, and write that down.
A clear way to stop and reverse. Every agent needs an off switch that a non-technical manager can use, and every action it takes should leave a record of what it did and why. If you cannot reconstruct a decision after the fact, you cannot defend it to a customer.
Measurement tied to the business, not the bot. “The agent handled 4,000 tickets” is activity. “Average resolution time fell from two days to six hours, and customer complaints did not rise” is a result. Measure the second kind, and compare it against the baseline you set before you started.
This is also where frameworks like the NIST AI Risk Management Framework become practical rather than theoretical. A small business does not need to adopt the entire framework. It does need to know what could go wrong, who would notice, and what happens next.
Back to the demo
The agent in that demo was never the impressive part. What made it look effortless was everything the vendor had quietly settled before the meeting: one inbox with one owner, one customer record, one written rule for when a follow-up goes out. The demo ran on a clean operation, and the software got the credit.
That is the honest case for agentic AI. It can take real work off real people, and the companies that use it well will run leaner than the ones that do not. But the advantage does not come from buying the agent. It comes from building the kind of operation an agent can succeed in, which happens to be the kind of operation that runs better with or without one.
So before the next demo, look at the shared inbox. Ask who owns it.
Ryan Gartrell is the creator of the Lean AI® framework and founder of Ryan Gartrell, P.A., an operational excellence advisory firm serving owner-led businesses. If you are weighing an agent project and want to know whether the operation underneath it is ready, start with a conversation or take the free Operational Readiness Assessment.
Sources
- Gartner prediction on agentic AI cancellations and agent washing, as reported by CRN India and OODA Loop
- CrewAI, 2026 State of Agentic AI Survey (Business Wire, Feb. 11, 2026)
- NIST AI Risk Management Framework
