Everyone is selling "AI agents" this year. Here is what they genuinely do well inside a small business today, where they fall over, and the five jobs worth handing to one first.
"AI agent" has become the phrase of the year. Every tool you already pay for has announced one, every consultant has a deck about them, and the promise is always the same: software that does not just answer questions but goes and does the work. Some of that is real and some of it is marketing wearing a new badge. Having built and run agents inside client businesses for the past year, we think the honest picture is more useful than either the hype or the dismissal — so here it is.
First, the definition that matters. A chatbot answers. An automation follows a fixed recipe: when a form is submitted, send this email, create this row. An agent sits between the two — it is given a goal and a set of tools (your inbox, your CRM, your store, a browser, a calendar) and it decides, step by step, which tools to use to reach the goal, checking its own progress as it goes. That freedom to decide is what makes agents powerful on messy, variable work, and it is also exactly why they need boundaries. Everything below follows from that trade.
What agents do well today is narrower than the demos suggest and more valuable than the sceptics admit. They are excellent at work that is repetitive in shape but different in detail every time: reading an enquiry, deciding what it is about, pulling the right information from your systems, and drafting a specific response. They are excellent at chasing — follow-ups, reminders, missing information, unpaid invoices — because they never forget and never feel awkward. And they are excellent at first-pass triage: sorting, tagging, summarising and routing, so a person only sees the things that need a person.
Here are the five jobs we hand to an agent first, in the order that pays back fastest for most businesses. One: lead qualification and reply. A new enquiry arrives; the agent reads it, checks it against what you actually sell, looks up whether the sender has contacted you before, drafts a tailored reply with the right next step, and either sends it or holds it for approval depending on the rules you set. Two: follow-up sequences that adapt. Not a fixed drip of three emails, but a follow-up that reads the reply, notices "we're deciding next month", and comes back next month. Three: inbox and ticket triage — every message categorised, urgent ones escalated, routine ones answered from your own documentation. Four: order and delivery updates for e-commerce, pulling the real status from your store and carrier and answering "where is my order" without a human. Five: reporting — the weekly summary of leads, sales, ad spend and site performance, written in plain English and sent to you on Monday morning, with anything unusual flagged.
Where agents fall over is just as important to know. They are poor at anything where being wrong once is expensive and hard to reverse: sending money, agreeing prices, making promises to customers, deleting data. They struggle when your own information is inconsistent — if your price list on the website disagrees with the one in your proposals, the agent will pick one and be confidently wrong. And they degrade when given too broad a goal; "grow the business" is not a task, "reply to every unanswered enquiry from this week" is. Almost every agent failure we have seen traces back to one of those three: an irreversible action, contradictory source data, or a goal with no edges.
So the way to run one safely is not complicated, but it is non-negotiable. Give the agent read access widely and write access narrowly. Put an approval step on anything that leaves the building — the first month, every outgoing message gets a human glance, and you loosen that only for the categories it has proven itself on. Log everything it does, in a place you actually look. Fix the source data before you connect it, because the agent will amplify whatever mess it finds. And measure it against a real number — response time, hours saved, leads answered within an hour — not against a feeling that it seems clever.
On cost, be realistic in both directions. A well-scoped agent for one of the five jobs above is typically a few weeks of setup, then a modest monthly cost in model usage and upkeep. That is not free, but it is a fraction of the salary it replaces part of — and, more to the point, it does the work at 11pm on a Sunday when the enquiry actually arrives. The businesses that see the return are the ones that start with one narrow job, prove it, and expand; the ones that try to automate everything at once usually end up trusting none of it.
If you take one thing from this: an agent is a junior colleague who works instantly, never sleeps, never gets bored, and has no judgement of their own yet. Treat it that way — clear tasks, good source material, a review step, and growing trust — and it will earn its place within a month. If you would like a straight assessment of which of your own processes are worth handing over first, that is exactly the conversation we start every automation engagement with.
E-commerceA growing share of shoppers now ask ChatGPT, Gemini or Perplexity what to buy — and let it compare, shortlist and even check out for them. Here is what your store needs to be one of the options it picks.
SEOGoogle AI Overviews, ChatGPT and Perplexity now answer the question before anyone clicks. Here is what actually gets a business named in those answers — and the seven changes to make first.
Software DevelopmentSpreadsheets everywhere, double data entry, tools that almost fit? Ten reliable signs that custom software development would pay for itself.
