Most articles about AI agents for small business are written by the people selling the platform. This one is written by an agency that put its own marketing on them first, kept notes, and is willing to say which parts are still bad.
We run Fallback Labs on nine brains and five teams. That is the honest scope of what follows: one small business, its own marketing, and a measured trail that starts in July. Everything here either happened to us or comes out of our own Search Console. Where we do not know, we say so.
What is an AI agent for a small business?
An AI agent owns one job from start to finish and does it without being asked each time. That is the whole distinction worth caring about. A chat window waits for your prompt and forgets you between sessions, while an agent keeps hold of what your business is and acts on its own schedule.
The part buyers underestimate is that second half. Software that does not know your products, your prices and your customers will produce confident, generic work at speed, which is worse than producing nothing at all, because generic work still costs you the time it takes to notice. Ours are given the company's own details before they are given a task, and that ordering does most of the heavy lifting.
| Asked to write a blog post | A chat window | An agent with context |
|---|---|---|
| Starting point | A blank prompt every time | Knows your products, prices and past posts |
| Topic choice | Whatever you suggest | Picks by search demand and difficulty |
| Consistency | Forgets the approach between sessions | Holds the same rules each time |
| Your job | Re-explain the business, then edit | Check it, then publish or send it back |
What did we actually replace?
Two of ours earn their place without argument. The SEO system covers the job that monthly research subscriptions are sold to do. The blog system replaced the copywriters we used to hire per article.
The SEO half is the less obvious win, so it is worth explaining. Research tools of that kind are good at showing you numbers and bad at deciding anything. They would report that a phrase had a certain volume and a certain difficulty, and then stop, leaving the judgement to whoever happened to open the dashboard that week. Ours makes the call and records why, which means the reasoning survives after the person who made it has moved on to something else.
The blog one is the cleaner story, so take that one. Writing an article used to mean briefing a freelancer and waiting, or opening a chat window and explaining the business again from scratch. The chat window could write perfectly good English. What it could not do was hold the plan: it would lose track of the approach it was supposed to take and hand back something readable and unfocused, and we would spend the saved time fixing the aim rather than the prose.
What changed is that our SEO and blog systems talk to each other. The blog side does not guess at a topic. It is told which phrase to target, how competitive that phrase is, how many people search it each month, and which smaller related phrases belong in the same piece. That handover is the actual product. Neither half is worth much alone, which is why buying a writing tool by itself so often disappoints.
The chat window could write. It could not decide what was worth writing.
Does it actually work? Here is our own data
Our impressions in Google grew roughly tenfold across three weeks in July, from about twelve a day to about a hundred and thirty. That is our own Search Console rather than an estimate, and it covers a young site with very little history behind it.
Now the other half, which is the part you will not usually be shown. Clicks did not follow. They went four, then five, then three, then zero across those same weeks, twelve for the month. Being seen is not the same as being chosen. We are early, the pages are new, and we would rather tell you that than crop the chart.
What works, and what still does not?
Writing, code, research and anything driven by data or an API work well enough to hand over. Pictures and video do not. That split has held for us through every experiment we have run, and it has not narrowed much this year.
The trouble with visuals is not that the output is unusable. It is that the output is subtly wrong in ways you catch after publishing, when a hand has too many fingers or a product looks almost but not quite like yours. We build video systems internally and we deliberately do not sell them, because we are not willing to put our name on that output for a paying client yet. When that changes, we will say so.
Dependable today
Not yet
There is a subtler failure worth knowing about, because it runs against instinct. Rules that are too specific make the output worse. Constrain the software tightly enough and it starts satisfying the letter of your instruction while missing the point of it, in the way a very literal new hire might. A clear job plus real detail about your business beats a long list of prohibitions.
Which one should a small business set up first?
Start with the one that reads, not the one that publishes. Something that has taken in your whole business first gives everything you add later a true starting point. Ours is the Website Brain, and knowing what we know now, we would set it up first again.
After that we would add SEO and blogs together, in that order, because that pair produced the clearest result for us and because each one checks the other. Publishing without the research half is how you end up with a tidy article that nobody was looking for.
We would not begin with anything that posts in public. Your published output is the most visible thing you own, and it is the worst possible place to find out what ungrounded software sounds like.
What does it cost?
Ours are $200 per brain, one time, and they keep running if you stop paying us. Managed is $1,250 a month at the introductory rate, renewing at $2,500 a month, which covers all nine brains and their teams with a person steering the work.
We print those numbers because almost nobody in this category does. That is worth pausing on when you shop. If a provider will not show you a price before a call, the price is a variable, and you are one of the inputs to it.
| Set it up once | Run it with you | |
|---|---|---|
| Price | $200 per brain, one time | $1,250 a month, renewing at $2,500 |
| Who operates it | You do | We do, with a person steering |
| Ongoing fee | None | Monthly |
| Best when | Someone on your side has time to run it | You would rather buy the outcome |
When are AI agents the wrong answer?
If your bottleneck is deciding what to sell, or who to sell it to, this will help you produce more of the wrong thing faster. Agents are good at execution once the direction is settled. They are not a substitute for settling it, and no amount of context fixes an unclear offer.
They are also wrong if nobody is going to read the output. Early on you check everything, and that is not a defect in the tool, it is the entry cost. We are still in that mode for parts of our own setup: running the systems, reading what comes back, and automating a step further only where the results have earned it. Anyone promising you a business that runs itself from week one is selling the demo rather than the thing.
If you want the longer version of how this splits into nine separate jobs, we wrote that up in our guide to AI marketing agents, and every brain is listed with its price on our build page.
Questions people actually ask
Are AI agents worth it for a business with no technical team?
For writing, research and data work, yes, provided somebody reads the output before it ships. Setup is closer to a conversation than a project. What you cannot skip is the reviewing, at least early on, because a confident wrong answer looks exactly like a correct one until you check it.
What can AI agents not do well yet?
Anything that depends on taste. Picture and video generation remain the weak spot, and they fail in ways you notice only after publishing. We hold our own video work back from sale for that reason. Writing, code, research and anything driven by data or an API are far more dependable.
How long before an AI agent saves you time?
Expect to spend more time in the first weeks than you save, because you are checking everything. The payback arrives when you stop reading every line and start spot-checking instead. That switch is a judgement call rather than a date, and rushing it is how bad work gets published.
Do AI agents replace employees or contractors?
In our case they replaced contractors we hired per piece, not salaried staff. The work that moved was repeatable and brief-driven. Anything needing accountability, taste or a client relationship stayed with people, and we do not expect that to change soon.
Can you write tighter instructions to make an AI agent more accurate?
Only up to a point. We found that rules which are too specific start producing worse results, not better ones, because the software contorts itself to satisfy the letter of the instruction. Give it a clear job and real detail about your business, then correct what comes back.
How do you judge an AI agent vendor?
Ask for a price before a call, ask what happens to your setup if you stop paying, and ask what their software is bad at. A vendor with no answer to the third question either has not used it seriously or is not going to tell you the truth about it.
