Access is not a workflow
The expensive miss in AI for business is rarely a bad demo. It is a license that never becomes a job. A team signs the contract, names an initiative, and waits for judgment to appear in a chat window. A sharp person sometimes gets a real win. Often the only thing that changed is the line item.
Treat that as a pattern operators recognize, not as a published scorecard. There is no universal failure rate sitting in the purchase order. If you cannot point to a task the software now does, and to a person who owns the miss, you bought access and called it a plan.
You have seen the human version. A crowded room, a coffee station, a hope that the person who can change your quarter is standing there too. Hope can make a pleasant morning. It is a poor substitute for a system, whether the system is a conversation or a piece of software.
An agent is not a chatbot with a longer leash
An AI agent, in the sense that matters on an operating review, is software that can carry a goal across steps. It pulls a record, uses a tool, drafts an action, and stops when a rule says stop. No taste. No quota. It will follow a bad instruction with the same energy it follows a good one.
That is why “turn on agents” is not a plan. A plan names the job, the inputs, the stop condition, the narrowest permission that still lets the job get done, and the human who owns the miss. If you cannot explain what the agent touched, you cannot defend the outcome.
AI automation works when the work already has a shape
The workflows that repay AI automation are usually dull on purpose. Invoice exceptions. Renewal reminders. Support tickets that are really “where is my order?” Lead records missing fields your reps already know how to find. The shared shape is a repeated task, structured inputs, and an output a person can check in under a minute.
If you cannot write the task as a checklist a competent new hire could follow, you are not ready to hand it to an agent. You are ready to write the checklist. Another meeting will not make undefined work safer to automate.
Time the human version before you automate it. Write down one number that matters: minutes per renewal, or the share of renewals a person forgets. Skip that baseline and every later claim about productivity is a mood. Vendors are happy to supply the mood. Your finance partner should not be.
Picture a renewal desk as a sketch, not a case study. Ninety days out, the agent lists accounts with a contract date you already store, checks whether usage fell, and drafts a note a manager can edit. It does not set the discount or email the customer on its own. Retrieval is “what is the date, and did usage drop?” Judgment is “do we concede anything, and does a person need to call?” The win, if you earn one, is fewer forgotten renewals and a shorter path from signal to that human decision. You measure it against the baseline. You do not infer it from the software being switched on.
AI for sales is where fluency gets expensive
Sales is where a well-written email gets mistaken for a qualified opportunity. AI for sales can still help, and the help is unglamorous. An agent can watch inbound interest, compare it with account data you already trust, draft a first note in the buyer’s language, and park anything ambiguous for a person. It can keep the CRM from turning into a junk drawer, and flag that a champion went quiet.
It should not invent a discount, describe a feature you do not ship, or contact a list you do not have permission to use. A model that sounds like your best account executive can still be wrong about the contract, the product, and the law. Accuracy still does not amount to consent. The cost of that miss is a customer, a regulator, or both.
More outbound is not more pipeline. If the agent multiplies messages faster than your team can tell a real buyer from a polite no, you have automated noise. Noise scales. Trust does not.
Do not let a myth write the quarter
A louder conversation runs beside this one, about superintelligence and whether some future system will make this quarter’s process work look quaint. Maybe. That possibility is not a reason to skip the unglamorous work. It is not a reason to hand an unbounded agent the customer inbox and the payments stack because a keynote sounded inevitable.
The durable move is narrower. Pick one revenue workflow and one internal workflow. Write the steps. Mark which steps are retrieval and which are judgment. Give the agent the retrieval, the formatting, and the reminders. Keep the judgment. Review a fixed sample every week, and expand only after the error rate is low and predictable.
The license can wait in the budget line. The checklist is what shows up in next quarter’s numbers, or does not.
Questions worth asking out loud
-
What has to be true before a task goes to an agent?
You can write it as a checklist a competent new hire could follow: repeated work, structured inputs, a clear stop, and an output a person can check in under a minute. If you cannot describe it that way, the task is not ready.
-
What do you measure before and after?
Time the human version first: minutes per case, and how often it is missed or done wrong. After you automate, compare those same numbers on a fixed weekly sample. A switched-on tool is not a result.
-
What must a sales agent never do?
It must not invent a discount, describe a feature you do not ship, or contact people you do not have permission to reach. A fluent note can still be wrong about the contract, the product, and the law.
-
What stays with a human?
Judgment. Pricing, exceptions, anything ambiguous, and the decision to send. The agent can retrieve, format, and remind. A named person still owns the miss.