Does your business need an AI agent? A 5-question test

Every software vendor now sells AI. Your accountant has heard of it. Your competitor says they use it. Somebody has told you that your business needs an AI agent.

Here is the short answer. Most businesses do not need a clever agent. They need one slow, repetitive task taken off someone's desk, and they need it done safely. Sometimes AI is the right tool for that task. Often a plain automation does the job for less money.

The question to ask is not "which AI should we use?" The question is "which task should AI do for us?" This guide gives you five questions to answer that, before you spend anything.

Agent or automation: what is the difference?

The two words get mixed up, so here they are in plain terms.

An automation follows fixed steps. The same input always gives the same result. For example, seven days after an invoice goes out, the system sends a reminder on WhatsApp. No judgment is needed. Ordinary code does this well, and it does not make mistakes once it works.

An AI agent makes a judgment call. It reads something messy, decides what it means and picks the next step. For example, a customer sends "need 3 boxes picked up from Sector 62 tmrw, same as last time." An agent reads that, finds the customer's last booking and creates a new one for a person to approve.

Most good AI projects mix the two. The fixed steps stay in ordinary code. Only the part that needs judgment goes to the AI model. That keeps the system cheaper, faster and easier to check.

The 5-question test

Answer these about one task in your business. Be honest. A "no" is useful, because it saves you money.

1. Can you name the task and the person who does it today?

"Use AI in sales" is not a task. "Kavita copies order details from WhatsApp messages into the order sheet every morning" is a task. You need that level of detail.

If you cannot name the task, stop here. You do not need an AI agent yet. You need an hour with the people who do the work.

2. Does it happen many times a week?

A task that happens twice a month is not worth automating. A task that happens 50 times a day probably is.

Count it. Look at last week and write down how many times the task happened and how long each one took. That number tells you what the project is worth.

3. If the AI gets it wrong, is that easy to spot and cheap to fix?

AI models make mistakes. Plan for that from the start.

Sorting incoming messages into "booking", "complaint" and "question" is a good first task. A wrong label is easy to see and costs a few seconds to fix. Approving refunds or sending legal notices is a bad first task. One wrong answer costs real money or trust.

If a mistake is expensive, a person must check every result before it goes out. That is still useful, but it changes the plan.

4. Does the task need judgment, or do fixed rules cover it?

This question decides between an agent and a plain automation.

If you can write the task as "if this, then that" rules, you do not need AI. Use an automation. It costs less to run and it gives the same answer every time.

If the input is messy, such as free text, voice notes, photos or documents in different formats, then AI earns its place. Reading a scanned invoice and pulling out the amount and the GST number is a job for AI. Adding 18% to a price is not.

5. Do you have past examples to test it on?

Before anyone relies on an AI agent, test it on your real past cases. Fifty old WhatsApp orders, a month of support emails, last quarter's invoices. Then count how many it gets right.

If you have no examples, you have no way to know if the agent works. Collect some first.

How to read your answers

Your answers	What to do
Yes to all five	A good candidate for an AI agent. Start small.
No to question 4 (rules cover it)	Use a plain automation. Skip the AI.
No to question 3 (mistakes are costly)	Use AI to prepare a draft. A person approves each one.
No to question 1, 2 or 5	Not yet. Name the task, count it and collect examples first.

What a safe first AI project looks like

If your task passes the test, keep the first version small. This is the order we follow on our AI workflow projects.

1. Pick one task. Sit with the people who do the work. Choose a task that is slow, repetitive and easy to check.
2. Design the flow. Decide which steps stay as ordinary code and which steps need the AI model. Then decide what the agent can touch and when a person approves.
3. Test on real examples. Run the agent on your past cases before anyone relies on it. Count how often it gets them right.
4. Start small, then widen. Give it a few users or a share of the work first. Watch the quality and the cost. Then give it more.

Three guardrails belong in every first project:

• A spending limit. AI models charge for each request. A cap stops a bug from turning into a large bill.
• A log of every action. You should be able to see what the agent did, when and why.
• A person in control. Somebody must be able to stop the agent, correct it and do the task by hand.

When the answer is "not yet"

Sometimes the right advice is to wait. Here are the common signs.

• You want to "add AI" without a task in mind. Start from a job to be done, not from an AI model.
• You have no past examples. Without them, nobody can test the agent honestly.
• A mistake would be costly and nobody will check the results. Fix the review step first.
• The work is already fast. If a task takes two minutes a day, an AI project will not earn back its cost.

A good partner will tell you when a plain automation is enough. Be careful with anyone who says every business needs an agent.

Common questions

Which AI model should we use? It depends on the task. We work with models from OpenAI, Anthropic and Google. We pick the one that does your task well at a cost that makes sense. The model matters less than the task you choose and how you test it.

Is our data safe with an AI agent? Decide this before you build. Send the model only the data the task needs. Keep the rest in your own systems, and keep a log of what goes out. If some data must never leave your business, say so at the start, and the design must respect that.

Can the agent work in WhatsApp, email or our CRM? Yes. An agent is most useful where your team already works. It can read from and write to WhatsApp, email, spreadsheets, your CRM or your own database.

Start with one task

You do not need an AI strategy. You need one task that is slow, repetitive and easy to check. Answer the five questions above about that task.

If you have a task in mind and want a second opinion, tell us about it. We will tell you plainly if it needs AI, a plain automation or nothing at all.