AI assistant or AI agent? And how to let it into your system without losing control.

First: an assistant and an agent are not two names for one thing

Three different things are all sold as "AI inside the system", and they are built differently:

1. A chat window stuck on the system — it answers from general knowledge, not from your data. It looks clever in a demo and invents figures when it does not know.
2. An AI assistant — you ask, and it answers from your own system: it reads invoices, stock and payroll with tools, shows the figure and the table, and does what you ask with your permissions.
3. An AI agent — it works on its own on a task you gave it: every morning, when a document is submitted, at month end. It reads, prepares, and brings you whatever needs your decision.

What each one does on a real working day

The assistant is for the question you have now; the agent is for the work that comes back every day:

1. Sales — you ask the assistant "this month against last?" and get the figure. You give the agent a report that reaches you every morning.
2. Collections — the assistant tells you who is late and since when. The agent reviews the late payers every Sunday, drafts the reminders, and waits for your OK.
3. HR and payroll — the assistant answers "what is this month's payroll?" The agent reviews leave requests and shift swaps every day and prepares the payroll at month end for you to approve.
4. Stock — the assistant lists the items running low. The agent tells you the moment a stock movement breaks a reorder level.

Seven questions to ask any vendor

Print these and take them with you. Each one reveals something the demo does not:

1. Whose permissions does it work with? If it can see what the person asking cannot, you have handed a new hire the keys to every department.
2. Where does the number come from? Ask to see the steps: which report it read, which filter it used. A figure you cannot trace is not fit for a decision.
3. What happens before a step that cannot be undone? Submitting an invoice, paying, deleting, sending to a customer — does it do it alone or wait for explicit approval?
4. Can you see every step afterwards? Who asked, what it read, what it changed, and when. Without a log, nobody is accountable.
5. How is it priced? Per message, per user, or unlimited? Ask for a monthly cap where it stops, not a bill that surprises you.
6. Where does your data go? Is the model trained on your conversations, and can you say no? The answer should be written, not spoken.
7. How do you stop it? One switch that stops an agent, or all of them, at once — ask to see it.

Levels of autonomy: start at the bottom

Do not give an agent everything on day one. The four sensible levels:

1. Report — it reads and tells you what it found, and changes nothing.
2. Draft — it prepares documents, comments and assignments as drafts for a person to review; it never submits or sends.
3. With approval — every change it wants becomes a request that waits for someone to approve it.
4. Act — it makes small changes on its own within a limit; everything else waits for approval.

What breaks in the second month

In week one everything looks wonderful. The problems appear once people get used to it:

A clerk asks it about colleagues' salaries and gets an answer. The AI bill arrives at three times the estimate. The agent sends a customer the same reminder twice because the task ran twice. It gives a figure that differs from the report and nobody knows why. It stops working and tells no one.

Red flags in the demo

1. It answers with figures and will not show where they came from.
2. It answers questions beyond the permissions of the user trying it.
3. "It can do anything" — with no word on what waits for approval.
4. No cost cap, or "pricing depends on usage" with no number.
5. It is shown on prepared demo data and you are not allowed to ask your own question.
6. No written answer about training the model on your data.

Checklist before you switch it on

1. You saw its answer match the system's report.
2. You tried it with a limited user, and it stayed within those limits.
3. You know what waits for approval, and who approves it.
4. You know the monthly cost cap and what happens when it is reached.
5. You saw the log of steps yourself.
6. You have the answer in writing about your data and training on it.
7. You know where the stop switch is.

And where we stand

We built Managely's AI Team on the answers above. The assistant and the agent work inside the system itself, with the permissions of the user who asks and nothing above them, and they read your data with tools, so no figure is without a source. Submitting, sending, paying and deleting wait for a confirmation, and every step is logged under the name of whoever asked. An agent works on tasks a manager sets — on a schedule or when a document changes — at one of the four levels, and no task is switched on before a test run has passed. Each agent has a monthly and a daily cap from a separate wallet, and a switch that stops it at once. And we learn from your conversations only if you explicitly allow it.

We will not tell you it replaces your team. It answers, prepares and follows up; the decisions stay with the people who own them. And this guide works with any other vendor too.