AI agents for business: what they actually do, and where they pay off.

01What an AI agent actually is

Most of what gets sold as an “AI agent” is a chatbot with a better prompt. It talks well, but it cannot do anything. The difference matters, because the value of an agent is in the doing.

An agent runs a loop. It reads a goal, picks a tool, uses it, looks at what came back, and decides the next step. The tools are ordinary software: your CRM, your calendar, your order database, WhatsApp, a payment link, an email inbox. The model is the part that decides which tool to use and when to stop.

Here is the loop on a real task, a new property enquiry arriving at 9:40 on a Sunday night:

1. Look up the phone number in the CRM. Existing contact, enquired twice before about two-bedroom flats.
2. Check site-visit slots for the project they asked about. Three open on Saturday.
3. Send a WhatsApp reply with the brochure and the three slots, using an approved template.
4. When they pick 11 am, book it, update the lead stage and tell the sales executive.
5. If they ask about loan eligibility, hand the conversation to a person. That is outside its job.

Nothing in that list is clever. That is the point. A good business agent does dull, well-defined work quickly, every time, at any hour, and knows exactly where its job ends.

02Five jobs agents do well today

We look for work that is frequent, follows rules a new hire could learn in a week, and already happens on a screen. These five come up again and again.

1. First response to new leads

Speed decides a lot of sales. An agent can reply in seconds, ask two qualifying questions, share the right document and book the next step. On our real estate CRM, automated follow-ups took first contact from 30 to 60 minutes down to about five. Read how we automated WhatsApp follow-ups.

2. Moving information between systems

Copying details from an email, a PDF or a form into the CRM is the job nobody wants. Agents read the source, fill the fields, and flag anything they are unsure about instead of guessing.

3. Support triage

Sort incoming questions, answer the routine ones from your own documents, and route the rest to the right person with a summary attached. The customer waits less, and the human starts with context.

4. Internal operations

Daily reports, reconciliation, checking that every order has an invoice, chasing missing documents. Back-office work with clear rules and a clear finish line is where agents quietly save the most hours.

5. Calls and follow-ups

Voice agents can confirm appointments, remind, and do first qualification on outbound calls, then transfer serious prospects to a person. It works best when your calls are already logged. More in our guide to AI calling agents.

03Where agents fail

The failures are predictable, which means they are avoidable.

• Vague goals. “Improve customer experience” is not a task. “Reply to every new lead within five minutes with the right brochure” is.
• No way to act. If the system it needs has no API and no export, the agent can only talk about the work, not do it.
• No ground truth. If nobody can say what a correct answer looks like, nobody can tell when the agent is wrong.
• Irreversible steps with no check. Refunds, cancellations and anything that moves money need an approval step, at least until the agent has a track record.
• Messy data treated as clean. Duplicate contacts and free-text fields will produce confident mistakes unless the agent is told how to handle them.

None of these are model problems. A better model will not fix a missing API or an unclear goal.

04What it takes to run one in production

A demo needs a prompt. A production agent needs the boring parts too:

• Integrations with the systems it reads and writes, with proper authentication and error handling.
• Guardrails: what it may do alone, what needs approval, and when it hands over to a person.
• An evaluation set: real past cases with known right answers, run before every change.
• A log of every action, so any decision can be traced and explained to a customer or an auditor.
• Cost controls: the right model for each step, with budgets and alerts.
• An owner who watches quality after launch, because prompts drift and upstream systems change.

This is most of the work, and it is the work most pilots skip. We wrote about why in what an AI proof of concept should prove.

05What an AI agent costs

There are two costs: building it and running it.

Building is driven by integrations far more than by the AI. An agent that reads one spreadsheet and sends emails is small. One that works across a CRM, a telephony system and WhatsApp, with approvals, is a real project. At Michi Studios, a fixed-price discovery sprint starts from $1,500 and a production build from $8,000, and you get a fixed quote before any build starts.

Running costs are model usage plus hosting. They scale with volume, so we size them per task in discovery: cost per lead answered, per document processed, per call handled. Often the cheapest reliable setup uses different models for different steps, a small fast model for routing and a stronger one only where judgement is needed. We pick between Claude, GPT, Gemini and open-source models on cost, privacy and quality, not habit.

06How to pick your first agent

List the repetitive work in your business and score each item on four questions:

1. Volume. Does it happen dozens or hundreds of times a week?
2. Rules. Could you write down how a good employee does it on one page?
3. Access. Can software reach the systems involved?
4. Risk. If it gets one wrong, is the damage small and easy to undo?

Start with the item that scores well on all four. For most sales-led businesses that is the first reply to a new lead. For operations-heavy teams it is usually data entry between two systems. Ship that, measure it against the old way for a month, then move to the next.

The first agent's job is not to be impressive. It is to be trusted, so the second one is easy to approve.

07Build or buy?

Off-the-shelf agent builders are a good fit when your workflow is standard and your systems are popular SaaS tools with ready connectors. You will be live faster, and you can switch later.

Custom development makes sense when the workflow is specific to your business, your systems are in-house or regional, you handle sensitive data, or the agent is part of a product you sell. That is the work we do: one team owns the agent from discovery to production, and you own the code.

08Questions people ask