AI agents for business: what they are, what they are worth

The term AI agent has been used to mean five different things in the last twelve months. Most demos look magical. Most production systems are quieter. This is the version we actually ship for clients, and the version that earns its keep.

What an AI agent actually is

Strip away the marketing and an AI agent is software that can do three things in a loop:

1. Read a request, in your business's language.
2. Use a small toolkit to look things up, take actions, and check its work.
3. Decide what to do next, then either reply or hand off to a human.

That is it. The "intelligence" part is the language model that does the reading and deciding. The "agent" part is the toolkit and the loop. The custom build is the toolkit and the loop fitted to your business, not a generic chatbot.

Where it pays back

The agents that earn their keep are usually unglamorous. Some examples from the work we do:

• Lead qualifier. Reads a new enquiry, checks past records, asks the two questions a human salesperson would ask, and books a call slot. Replaces the back-and-forth that used to take 18 hours.
• Inbox triage. Reads incoming support emails, tags them, drafts a reply that matches your tone, and drops it in your queue ready to send.
• Document reader. Watches a folder. When a new contract or invoice lands, it pulls out the structured fields, checks for issues, and posts a summary to the right channel.
• Internal Q&A. A WhatsApp or web bot trained on your SOPs, your product catalog, your FAQs. Stops the team asking the same five questions every week.

Where it does not

The places we tell clients not to use an agent yet:

• Anything where a wrong reply costs more than a few minutes to fix.
• Anything that needs a human's judgement on tone (negotiating a refund, breaking bad news to a vendor).
• Anything where the underlying data is messy. The agent will just learn the mess.

The right test is not "can the agent do this?" The right test is "if the agent gets it wrong twice a week, does anything bad happen?" If yes, keep a human in the loop.

What it costs to run

A typical small-business agent runs on around USD 30 to 80 a month in API costs, depending on volume. The build is a one-time engineering cost, after which you own it. Compare that to the cost of one full-time employee doing the same work, and the math is usually short.

How we approach it

We do not start with the agent. We start with the workflow. If the same result can be hit with a simple workflow automation, we build that, because it is cheaper, faster, and easier to debug. The agent only comes in when the work genuinely needs language understanding or judgement.