LESSON 04
A CHAT CAN BE WRONG.
AN AGENT CAN ACT.
Agent is currently the most oversold word in business software. Underneath the marketing there is a real distinction, and it matters because it changes what happens when the system is wrong.
Three different things with one name
Most confusion here dissolves once you separate three categories that vendors blur together.
A chat
You ask, it answers. Nothing happens in the world. You take the output and decide what to do with it. Every mistake is caught by you, before anything moves.
An automation
A fixed sequence that fires on a trigger. When a form is submitted, create a record and send this email. It does the same thing every time. It does not decide anything.
An agent
A model given a goal, a set of tools, and permission to decide which to use and in what order. It takes actions in your systems without a person approving each one.
The first two are well understood and mostly safe. Automations have been running businesses for twenty years, and their failure mode is boring: they do the wrong fixed thing repeatedly until someone notices.
The third is different in kind, not degree. An agent is a system that is sometimes wrong and is also holding the keys.
What 'taking an action' actually means
Vague language hides the stakes, so here is the concrete version. An agent with tools can do things like this without asking you first:
- Send an email from your address to your customer.
- Create, update or delete a record in your CRM.
- Book, move or cancel something on a calendar.
- Post a message into a channel your team reads and acts on.
- Move a deal to a different stage, which triggers whatever that stage triggers.
- Pull data out of one system and write it into another.
Each of those is reversible in principle and awkward in practice. An email sent to a seller cannot be unsent. A calendar invite cancelled by mistake has already reached the other person's phone. A record overwritten with wrong information is now the record your team trusts.
Everything on this site is an extension of Ben's published resources.
Why permission to act raises the stakes on review
In a chat, you are the review layer and you are in the loop by default. You cannot use the output without reading it, so bad output is caught at roughly a hundred percent.
The moment a system can act, that stops being automatic. Review has to be designed in deliberately, and if nobody designs it, it does not exist. The failure is also quieter. A chat that hallucinates hands you a visibly odd paragraph. An agent that misreads a situation sends a perfectly normal-looking email to the wrong person, and you find out when they reply.
Volume compounds it. A person making a judgment error makes one. A system making the same error makes it on every item in the queue before anyone checks the queue.
None of that is an argument against agents. It is an argument that the question changes. For a chat you ask whether the output is good. For an agent you ask a harder one: what is the worst thing this can do unsupervised, and how would I find out.
The safe way to introduce one
There is a sequence that works, and it is deliberately slow at the start.
- Start with drafts, not sends. Let it prepare the email and leave it in a queue for a person to approve. You get most of the time savings with none of the exposure.
- Give it the narrowest tools that do the job. Read access before write access. One system before three. A system that can only read cannot damage anything.
- Put a person on the irreversible steps. Anything touching money, a contract, or a customer's first impression keeps a human approval, permanently, not as a training-wheels phase.
- Log everything it does. If you cannot review a list of actions taken this week, you do not have oversight. You have hope.
- Decide who owns it. A named person who checks the log, handles exceptions, and has the authority to switch it off.
- Then widen it, one permission at a time, after it has been reliable on the narrow version for long enough to be boring.
What this is worth, honestly
Used well, on well-defined work, this is where the real leverage lives. Not because the model is clever, but because a system that executes a process every time removes the most common failure in a small business, which is that the process depended on someone remembering.
Used badly, it is the most expensive way yet invented to scale a broken process. An undefined process handed to an agent does not become defined. It becomes an undefined process running faster, in more places, with less visibility.
Better processes first. The right technology second.
Ben Lovro
If you cannot write down what the work is, in steps, with a definition of done at each one, you are not ready to hand it to anything. That is true of a new hire and it is more true of a system that never asks a clarifying question.
Frequently asked
Questions people actually ask
Do I need an agent, or is a chat enough?
For most owners starting out, a chat and one or two simple automations cover the majority of the value. Agents earn their place when a well-defined, high-frequency process is already documented and the only remaining problem is that a person has to run it.
What is the difference between an agent and an automation?
An automation follows a fixed path you defined. An agent decides the path, using a model, which makes it more flexible and less predictable. Flexibility is the feature and unpredictability is the cost, and they are the same property.
What is the most common way this goes wrong?
Permissions granted broadly at setup because narrowing them was fiddly, combined with nobody reading the log. The system then does something reasonable-looking and wrong for two weeks before anyone notices.
Should an agent ever email customers directly?
Only for genuinely routine, low-stakes messages, and only after a long stretch of human-approved drafts proving it handles the edge cases. Anything touching price, terms or a difficult situation stays with a person.
Who should own an agent inside a small business?
One named person, not a committee and not whoever set it up. They review the log, handle exceptions, and have unambiguous authority to turn it off without asking permission.
Make your next move
A year from now, what will you be glad you started today?
You don't need another promise that everything will be easy. You need something useful to learn — and a next step you're willing to take.