When Should You Use AI Agents in Your Business (And When You Really Shouldn't)

Straight answer: use an AI agent when you have a task that repeats at least 50 to 100 times a month, follows rules you can write down in a page, and currently sits between two systems that don't talk to each other. Don't use one to paper over a process that's still a mess in your head, because the agent will just do the mess faster and with more confidence. Most businesses need three or four agents doing narrow jobs, not one grand assistant doing everything.

What an AI agent is (and isn't)

People throw "AI agent" around to mean everything from a chatbot to a full digital employee, so let's be clear first. A chatbot answers a question and stops. An automation follows a fixed if-this-then-that script and breaks the moment reality doesn't match the script. An agent is different: it's given a goal, a set of tools (your calendar, your CRM, your inbox, a database), and it decides the steps itself, adjusting as it goes. It can check a customer's order status, notice the delivery is late, draft an apology email, offer a discount code within a limit you set, and log the whole thing, without you writing out every branch of that conversation in advance.

That flexibility is the whole point and the whole risk. A script can only do what you told it to. An agent can do things you didn't quite mean it to.

The volume test I use with clients

Before I recommend building anything, I ask one blunt question: how many times a month does this task happen? If the answer is under 20, I tell people not to bother. Below that volume, a good template, a checklist, or a part-time VA on 10 hours a week will beat an agent on cost and on peace of mind, because you're not maintaining a system for something you barely do.

The threshold that justifies the build time (and the ongoing token cost) sits around 50 to 100 repetitions a month, and it gets compelling above 200. I worked with a small property management company doing roughly 40 tenant maintenance requests a week (about 170 a month). That's squarely in agent territory: triaging urgency, checking which contractor covers that postcode, booking the slot, and updating the landlord, all following clear rules. We built that in under three weeks. Compare that to a solo consultant handling 8 client onboarding calls a month. Not worth automating. A checklist and a calendar link does the same job for free.

My own experiment, and the mistake I made

I ran this test on myself last year. My inbox had a repeating pattern: people asking about pricing for one-off consulting sessions, roughly 15 to 20 a week, all wanting the same three pieces of information back. I built an agent to read the enquiry, pull the right rate card, and send a tailored reply with next steps.

The first version was too eager. It started quoting availability dates without checking my actual calendar, because I'd given it access but hadn't told it to double-check against holds and tentative bookings, only confirmed ones. Within a week it had promised a Tuesday I didn't have. That's on me, not the tool. I'd skipped the boring part: writing down my actual booking rules before I let the agent touch my diary. Once I fixed that (a two-hour job, not a technical one), it worked well and now handles about 80 percent of pricing enquiries without me touching them. The lesson stuck: the agent is only as good as the process you hand it, and most businesses don't have that process written down anywhere, they just have it in the owner's head. If you want to know whether your business could survive you disappearing for a fortnight, that gap shows up fast, which is exactly what I dig into in the founder dependency audit.

Where agents earn their keep

• Customer support triage, sorting incoming tickets by urgency and topic, answering the repeatable 60 percent, and escalating the rest with full context attached.
• Lead qualification and booking, especially for businesses with high enquiry volume and a clear set of qualifying questions, like estate agents fielding viewing requests. If that's your sector, I've mapped out exactly which tasks pay back fastest in my piece on AI for estate agents.
• Inbound phone handling for businesses drowning in calls outside office hours, where the comparison isn't "agent versus human," it's "agent versus voicemail nobody returns." I've written a full breakdown of that trade-off in AI voice agents versus traditional call centres.
• Data reconciliation between systems, like matching invoices to purchase orders, or checking stock levels across three platforms that were never designed to talk to each other.
• Research and drafting at volume, pulling competitor pricing weekly, drafting first-pass proposals from a brief, summarising meeting notes into action items.

Where they don't, and why nobody likes admitting it

Here's the uncomfortable bit that most posts on this topic skip: an AI agent will confidently execute a bad process at speed. If your onboarding is inconsistent because your team improvises it call by call, an agent won't fix that, it will just improvise faster and with less accountability, because nobody can ask it "wait, why did you do that?" and get a useful answer. Businesses that are already a bit chaotic often reach for agents hoping the tool will impose order. It won't. It will expose the chaos, usually to a customer, usually in writing.

Skip agents (for now) if:

• Your process changes every few weeks. Agents need stability to be worth building; if the rules shift constantly, you'll spend more time rebuilding the agent than it saves.
• The task requires judgment calls with real consequences and no clear rule, like deciding whether to refund an angry client beyond your stated policy. That's a human call, keep it one.
• You don't have anyone who can check its output weekly. An unsupervised agent that quietly starts making small mistakes for two months is worse than doing the task manually.
• Your data is a mess. An agent pulling from three spreadsheets with different naming conventions will produce confidently wrong answers, faster than a human would.

The cost question nobody explains

Agents run on tokens, and this catches almost everyone out the first time. Every message the agent reads, every tool call it makes, every step of "thinking" it does before acting, costs tokens, and the cost scales with complexity, not just volume. A simple one-step agent answering a fixed question might cost pennies per run. A multi-step agent that reads an email, checks a database, drafts a reply, and logs it in your CRM can cost several times that per interaction, because it's making multiple calls behind the scenes. I've seen a client's monthly bill jump from around 40 pounds to over 300 pounds simply because the agent started making three extra "checking" calls per request after a prompt update, nobody noticed for six weeks. If you want the full mechanics of why this happens and how to keep it under control, I've laid it out in why AI agents need tokens and what it means for your costs. Budget for monitoring, not just for building.

A five-step check before you build anything

• 1. Write down the task exactly as it happens now, every branch and exception, even the annoying edge cases. If you can't write it in a page, it's not ready for an agent yet.
• 2. Count the monthly volume honestly. Pull real numbers from your inbox, CRM, or booking system, don't guess.
• 3. Decide the ceiling of authority. What can it decide alone, and what must it flag to a human? Put a number on it, like "can offer a discount up to 10 percent, anything above needs approval."
• 4. Build the smallest version first. One task, one tool connection, tested on real (not hypothetical) enquiries for two weeks before you expand it.
• 5. Check its work weekly for the first month, then monthly after that. Agents drift as your business changes; nobody is watching if you're not.

None of this requires a developer or a six-month project. I've walked through the actual build process, tool by tool, for people doing it themselves in how to build AI agents from scratch without a developer, and if you'd rather have someone map your specific tasks against the volume and cost math above before you commit any budget, that's exactly the starting conversation I have with clients through my AI implementation coaching work.

The pace of what's available changes fast enough that it's worth checking in regularly rather than deciding once and forgetting it, which is why I keep a running note of what's shifted for small businesses each week in posts like this week's AI news roundup.

Frequently asked questions

How many AI agents should a small business have?

Most small businesses that use them well run two to four narrow agents, each handling one clear task like support triage or lead qualification, rather than one broad assistant trying to do everything. Narrow agents are easier to test, cheaper to run, and easier to spot when something goes wrong.

Do I need a developer to set up an AI agent?

No, for most common business tasks like email handling, booking, or CRM updates, you can build a working agent using no-code platforms and existing tool integrations, usually within a few days rather than months. A developer becomes useful once you need custom integrations with older or bespoke internal systems.

What's a realistic monthly cost for running an AI agent?

A simple single-step agent handling a few hundred interactions a month typically runs somewhere between 20 and 100 pounds in token costs, depending on the model used. A more complex multi-step agent checking multiple systems per interaction can run several hundred pounds, so it's worth monitoring usage weekly for the first month rather than assuming the initial estimate will hold.

What's the biggest reason AI agent projects fail?

The process being automated wasn't clearly defined before the build started. When the underlying task is inconsistent or lives only in one person's head, the agent inherits that inconsistency and executes it at speed, which usually shows up as errors reaching customers rather than staying internal.

Sources worth reading

• Anthropic's own documentation
• the NIST AI Risk Management Framework