Agents explained
Why the scripted assistant failed
and what an Agent does differently
The old kind matched your words against a list somebody wrote. An Agent reads what was said, works out what is being asked, and has code underneath that checks the result.
The thing you tried before was matching your words against a list somebody wrote, so anything off the list fell through. An Agent gives the question to a model that understands what was said, and has code underneath it that makes sure the result is correct.
The second difference matters more than the first. The old kind could only reply. An Agent can do the thing: write the record, book the slot, file the document, send the follow-up. That is exactly why it asks a person before it changes anything, and writes down what it did afterwards. You can read the log.
It reads the question
A model works out what somebody meant, including spelling mistakes, Hinglish, half a sentence and a photograph of a bill. There is no list of accepted phrasings to fall off.
The model works it out, code checks it
The model works out what is being asked. The arithmetic, the lookup and the writing into your records are done by code, so the answer is correct rather than plausible.
It asks before it changes anything
Anything that writes a record, moves a file or sends a message is confirmed by a person first. That confirmation is part of the build, rather than a setting somebody can forget.
It writes down what it did
Every action goes into a log with the time, who asked and what changed. When somebody says the system did something odd, you can look instead of guessing.
It lives in the app you already use
WhatsApp, a team chat, email, a sheet. There is no new screen to log into, which is the other reason the last one failed. The WhatsApp version.
It can say it does not know
When the records do not hold the answer, it says so and names who to ask. A scripted assistant had to pick its nearest menu item instead.
What the one you tried was actually doing
The assistants most businesses tried were decision trees. Somebody wrote out the questions customers ask, listed the words those questions contain, and wrote a reply for each. When it ran, the thing matched your sentence against that list and played back the closest reply. If your sentence was not on the list, it either guessed the nearest one or asked you to rephrase.
That is why it felt stupid. It was never understanding anything. It was looking you up. Every unusual question, every spelling mistake and every sentence in two languages was a miss, and the only fix was somebody adding another branch by hand.
What is different now
The part that changed is the understanding. A model reads what was actually written, in the words the person used, and works out what they want. Nobody has to list the phrasings in advance. A message asking for tomorrow's gate pass for the trucks coming from Bhiwandi is understood the first time it is sent, and so is the same thing said badly.
Understanding on its own would be no better, because a model can be fluent and wrong. So the understanding is used only to work out what is being asked. The actual work, looking up the record, doing the arithmetic, writing the entry, is done by ordinary code that either works or fails visibly.
The two of them side by side
This is the comparison I draw on a whiteboard in a first meeting.
| What happens | The scripted kind | An Agent |
|---|---|---|
| Somebody asks something unexpected | Falls through to a rephrase request, or the wrong branch | Reads it, works out the intent, answers or says it cannot |
| A message in Hindi, Hinglish or half a sentence | Misses, unless somebody listed that phrasing | Understood, because nothing was listed in advance |
| It has to do something, rather than reply | It cannot, beyond handing you a form | It writes the record, files the document or books the slot |
| Making sure the answer is correct | The reply was written by a person in advance | Code does the lookup and the arithmetic, so it is checked |
| Changing one of your records | Outside what it can do | It asks a person first, every time |
| Finding out what it did yesterday | Nothing to look at | A log with the time, who asked and what changed |
| Where your team uses it | A widget on a website somebody has to visit | Inside WhatsApp, the team chat, email or a sheet |
| Adding a new kind of request | Somebody writes another branch by hand | Usually nothing. It already understood the sentence |
The model decides, the code guarantees
This is the rule I build to, and it is worth understanding, because it is what separates a demonstration from something you can run a company on. The model is allowed to decide what is being asked and which tool to use. It is never allowed to be the thing that produces a number, a date or a record.
So when somebody messages asking how much leave a person has left, the model works out that this is a leave question about that person. Code then reads the leave table and does the subtraction. The answer is right because arithmetic did it. A system that let the model do the subtraction would be right most of the time, which is useless for payroll.
It asks before it changes anything
Reading is the safe half. Writing is where the care goes. So anything that changes a record, moves a file, or sends a message to somebody outside the team is written as a proposal first. The Agent says what it is about to do, in plain words, and a person says yes. Only then does it happen.
That one rule is what makes an Agent usable in a business where mistakes cost money. It also keeps the thing honest about itself. Because every action is proposed and then logged, you can read back a week of its work and see exactly what it touched.
- Reading your records, answering questions, drafting a message: no confirmation needed
- Writing into a record, booking a slot, filing a document: confirmed by a person
- Messaging a client or a vendor: confirmed by a person, with the text shown first
- Every one of those, confirmed or not, written into a log with a time against it
Where an Agent is still the wrong answer
I turn this work down regularly. An Agent is the wrong tool when the job is already working, when there is no record to read, or when what people actually need is a form and a reminder.
Part of what the audit produces is the list of things to leave exactly as they are. It is usually longer than the build list. What the audit looks at.
- Work that runs well today and nobody complains about
- A decision that needs a person's judgement and should keep needing it
- A process so inconsistent that nobody can say what the correct outcome is
- Anything with no record of what happened, because there is nothing to read or write
- A job that happens rarely, where doing it by hand costs less than maintaining anything
What to ask before you buy another one
If the last one disappointed you, the difference shows up in what you ask before buying another. You can also just use the ones already running: several work in a browser, with no form to fill in. Try them yourself.
These are the questions that separate the two kinds. Ask them of me as well.
- What happens when somebody writes something nobody anticipated
- Who does the arithmetic, the model or the code
- What it can change in our records without asking a person
- Show me the log of what it did yesterday
- Which app does my team open to use it
- Who owns it, and what happens to it if we stop working with you
Built, not described
Questions people ask
What is the difference between an AI agent and a scripted assistant?
The scripted kind matched your words against a list somebody wrote in advance, so anything off the list failed. An Agent gives the question to a model that works out what was meant, then uses code to do the actual work. It can also act, which the scripted kind could not.
What is an AI agent, in plain words?
Software that reads a request in ordinary language, works out what is being asked, does the job inside your own systems, asks a person before it changes anything, and writes down what it did. It lives in an app your team already uses.
Why did our last one fail?
Usually two reasons. It could only match phrasings somebody had listed, so real messages fell through. And it lived on a website nobody visited, instead of in the app the team was already using all day.
Can an Agent make a mistake?
Yes, in what it understands, which is why anything that changes a record or sends a message is confirmed by a person first. The numbers and the lookups are done by code rather than by the model, so those are either right or visibly broken.
Does it work in Hindi or Hinglish?
Yes. The model reads what was actually written, so a mixture of languages, a shortened sentence or a spelling mistake is understood without anybody listing the phrasing in advance. Most of the messages my own systems handle are written that way.
Will my team have to learn a new app?
No. The Agent sits in WhatsApp, the team chat, email or a sheet, whichever your team already uses. At my own studio eighteen people use the systems every day and none of them sat through a training session.
Can an Agent act on its own without asking us?
It can read and answer on its own. Anything that writes a record, moves a file or messages somebody outside the team is proposed to a person first, and only happens once they agree. Every action, confirmed or not, goes into a log.
Next step