You do not need forty AI tools. You need about three, and the reason has nothing to do with features. It is that a business run by one or two people has a fixed amount of attention, and every tool you add spends some of it permanently: a login, a billing line, a place where work now lives, a thing to remember to check.
Every roundup of the best AI tools for business ignores that cost, because roundups are paid by the tool. So here is a shorter list, scored on the only question that decides whether a tool survives past week three: does the output ship without you rewriting it?
- Five jobs are worth paying for in a small business: drafting copy, answering the inbox, reading reviews in bulk, moving data between apps, and cleaning up photos, a job with a boundary worth knowing before you cross it, since a generated catalogue image can only honestly do some of the jobs a listing needs.
- The scoring rule that matters is the rewrite rate, not the feature list. A tool whose output you rewrite every time has a negative return however cheap it is.
- Checked on 30 July 2026: Claude Pro is 17 dollars a month billed annually or 20 monthly, Notion Business is 20 dollars a seat, and Zapier Professional starts at 19.99 a month annually for 750 tasks.
- Zapier bills by task, not by seat, which is the only pricing model on this list that punishes you for growing rather than for adding people.
- Notion charges AI agents in credits on top of the seat price, 10 dollars per 1,000 monthly credits, so the sticker price is not the bill.
- The US Census Bureau measures AI use across fifteen separate business functions, which is a useful reminder that a business counted as an AI user is often doing exactly one thing with it.
- Skip anything whose output you cannot check in under a minute, and skip the fourth tool entirely until one of the first three has earned its place.
How many AI tools can a small business actually keep?
Three, in our experience of building for merchants, and this is a judgement rather than a measured statistic, so treat it as a starting hypothesis you can test in a fortnight. The mechanism is simple enough. A tool only pays back if you reach for it without thinking. Reaching for it without thinking requires having used it enough times that it became a habit. Habits form on things you touch several times a week. There are not many jobs in a small business you touch several times a week.
Add a fourth tool and something predictable happens: it competes for the same slot as one of the first three, you use both at half the frequency, neither becomes a habit, and six weeks later you are paying for four subscriptions and using one. This is the actual failure mode of AI adoption in small businesses, and it is not a failure of the tools.
The Census Bureau's business survey supports the shape of this from the other direction. Its breakdown of AI use across firm sizes measures adoption across fifteen distinct business functions, from finance to customer service to marketing. A firm that answers yes to one of those fifteen counts as an AI user. When you read that a fifth of businesses use AI, that fifth is mostly firms doing one thing. Depth, not breadth, is the normal pattern even among the companies that look ahead of you.
What is the right way to score an AI tool?
By rewrite rate. Run the tool on ten real pieces of work and count how many outputs you shipped as written. Nothing else on a pricing page predicts whether you will still be paying in three months.
On copy specifically the rewrite rate splits sharply by page type, which we break down in the piece on how much editing AI written shop copy needs before it ships. The reason this beats the usual criteria is that it collapses several things into one number you can measure in a fortnight. A tool with a great model and a bad interface has a high rewrite rate because you never reach for it. A tool that produces fluent text about a product it knows nothing about has a high rewrite rate because you have to correct the facts. A tool that saves you eight minutes of typing and costs you twelve minutes of verifying has a high rewrite rate by definition. One metric, all three problems.
Count rewrites, not minutes. Minutes saved is the number people get wrong about themselves, sometimes by forty percentage points, which we go through in detail in the piece on which hour of a founder's week AI genuinely gives back. Rewrites are countable and unflattering, which is what makes them useful.
The shortlist
Five jobs, one tool class each. Prices are stated only where we opened the vendor's own pricing page on 30 July 2026 and read the figure there; where we could not verify a price today, the row says so rather than guessing.
| Job it replaces | What it costs | What it cannot do | Who should not bother |
|---|---|---|---|
| Drafting copy: product pages, emails, replies, alt text | Claude Pro at 17 dollars a month billed annually, 20 monthly; team seats 20 annually, 25 monthly | Know anything about your product it was not told, including specs, materials and stock | Anyone writing fewer than a handful of new pieces a month. The habit will not form |
| Answering the inbox: the questions you have answered two hundred times | Price not verified today, and the market moves monthly. Check the vendor page before you budget | See your order system unless you connect it, or handle the exception politely | Shops with low ticket volume but high ticket value. The exceptions are the whole job |
| Reading reviews and tickets in bulk | Usually included in the assistant subscription above, so no separate line | Weight the rare expensive complaint properly against the common cheap one | Anyone under about fifty reviews. Read them yourself, it is faster and better |
| Moving data between apps: orders to accounting, forms to a sheet | Zapier Free at 100 tasks a month; Professional from 19.99 monthly billed annually for 750 tasks; Team from 69 annually per month at 2,000 tasks | Fix a broken process. It will run a bad workflow faster and more reliably | Anyone whose apps already integrate natively. Check that first, it is often free |
| Internal notes, docs and search across your own material | Notion Plus at 10 dollars a seat, Business at 20, plus 10 dollars per 1,000 monthly credits for agents | Be the source of truth for stock or orders. It is a notebook, not a system | One-person shops whose notes fit in one document. You are the search function |
| Cleaning up catalogue photos: backgrounds, crops, consistency | Price not verified today. Most bundle it into a design subscription rather than charging separately | Change what the product looks like without creating a returns problem | Anyone selling items where texture or exact colour decides the purchase |
What do the assistants actually cost per month?
Less than the listicles imply, and structured differently than you would expect. Anthropic's published plan table lists Pro at 17 dollars a month if you pay for the year up front and 20 if you pay monthly, with team seats at 20 a seat annually or 25 monthly, and a heavier tier from 100 a month for people who hit usage limits. The pattern across the category is the same: roughly 20 dollars a month per person for the general assistant tier, with the real variable being how much usage you get before you are throttled rather than which model you can reach.
For a one-person business that means the assistant is not the expensive part of your stack. It costs about what one paid ad campaign costs in a slow week. The expensive part is everything that bills by volume.
Where the bill surprises people
Two pricing models on this list behave completely differently from a seat, and both are worth understanding before you commit.
Zapier bills by task. Its pricing table puts the free tier at 100 tasks a month and the Professional tier at 19.99 a month billed annually for 750 tasks, rising through 1,500 and 2,000 task tiers, with Team starting at 69 a month billed annually for 2,000 tasks. A task is one step of one automation running once. So the moment your order volume doubles, your automation bill moves with it, and a workflow with six steps consumes six tasks per order rather than one. This is the single most common billing surprise for merchants automating order handling: people budget for the plan and get billed for the growth.
Notion charges for agents in credits on top of the seat. Its plan page puts Plus at 10 dollars a seat and Business at 20, and prices custom agents at 10 dollars per 1,000 monthly credits after a free trial. Seat plus consumption is becoming the standard shape for AI features inside existing software, and it means the sticker price on the plan is a floor rather than a bill.
Google Workspace tiers work a third way again: rather than a separate AI subscription, its plan comparison gates Gemini access by tier, with the entry plan carrying limited access and the middle tiers adding Gemini inside Gmail, Docs, Sheets and Meet, plus a separate expanded-access add-on. If you already pay for email and documents, part of what you are shopping for may already be switched off in your own account.
What about the AI already bundled into software you pay for?
Check it before you buy anything new, because bundled AI has a structural advantage the standalone tools cannot match: it sits where the work already is. The rewrite rate on a drafting feature inside your email client beats the same quality of output in a separate tab, because the separate tab requires you to remember it, switch to it, paste context into it, and paste the result back. Four steps that each leak time and none of which appear on a pricing page.
This is why the Google Workspace tiering matters more than it looks. If your business already pays for email and documents on a middle tier, a meaningful part of the drafting job may already be sitting unswitched-on in your account. The same is true of most accounting software, most email marketing platforms and most helpdesks in 2026: the AI feature arrived in an update you did not read. An hour spent auditing what you already own is the highest-return hour in this whole exercise, and it is the one nobody sells you.
The counterargument is real and worth stating. Bundled AI is usually a weaker model with less control, and on a job where quality decides the outcome, the standalone tool wins clearly. The rule that falls out is not complicated: use the bundled version for volume work where good enough is genuinely good enough, and pay for a standalone assistant for the handful of pieces that carry your name.
Why is reading reviews the most underrated job on the list?
Because it is the one job on the shortlist where the machine is better than you rather than faster than you, and the only one whose output you cannot produce by hand at all.
You can write forty product descriptions. It will take a weekend and they will be better than a model's. You cannot read three hundred reviews and hold the distribution of complaints in your head. Nobody can. Summarising a large pile of unstructured text into recurring themes is a task where human working memory is the binding constraint, so the comparison is not model against you, it is model against nothing.
The caveat from the table still holds and it is a sharp one. Frequency and cost are different axes, and a summary weights by frequency. Nine people mentioning slow delivery and one person describing a product that arrived unsafe will come back as one theme and one outlier, when the outlier is the one that ends up costing you. Read the summary for what to fix at scale, then read the one-star reviews individually. That takes twenty minutes and it is not optional.
Which of these should you skip?
Most of them, most of the time. Three specific skips are worth naming because they cost real money, and the full list of jobs where no tool is good enough yet is longer still.
Skip the forecasting tool until your data qualifies. Two years of clean history per line and few zero sales weeks, or a reorder point does the job better. The evidence for that is in our piece on what a demand forecast needs before it beats a reorder point.
Skip the automation tool until the process is right. Automation is a multiplier on a process, and multipliers work in both directions. If your order handling has a manual check in it because that check catches mistakes, automating around it removes the check and keeps the mistakes. Fix the process on paper first, then automate the version that works.
Skip the inbox tool if your volume is low and your tickets are valuable. A shop handling six enquiries a day, each worth a few hundred, is a shop where every enquiry is an exception. The category is built for the opposite shape.
Skip anything you cannot check in under a minute. This filter removes more bad purchases than any feature comparison. If verifying the output takes longer than producing it yourself, the tool has not saved you work, it has converted writing into auditing and added a subscription.
Why does the fortieth tool make you slower?
Because attention is the scarce input, not software. Each tool carries a fixed overhead that never goes away: remembering it exists, remembering which of your two overlapping tools does this particular job, keeping its content in sync with the other place the same content lives, and the small tax of deciding.
The deciding tax is the one people underestimate. When one tool does a job, using it is not a decision. When three tools could do it, every instance of that job now starts with a choice, and choices are what make a Tuesday feel long. A stack of three tools you never think about beats a stack of twelve you evaluate each time, even if the twelve are individually better.
There is a version of this that applies to your website rather than your desk. Merchants often end up with a site, a separate admin panel, a spreadsheet that is really the stock system, and a fourth thing for the blog, then hire someone to keep the four in sync. That is the same sprawl problem wearing different clothes, and it is why we build the storefront and its back office as one thing: an AI generated shop with its admin included removes a category from the stack rather than adding one to it. The credit costs for a build are worth checking against what four subscriptions and a freelancer come to.
The two week test
Pick the one job on the shortlist you repeat most. Take the cheapest credible tool for it, not the best reviewed one. Run it on ten real pieces of work over two weeks, and write down after each whether you shipped the output as written or rewrote it.
Seven or more shipped unedited: keep it, and only then consider a second. Four to six: keep it but expect it to stay a drafting aid rather than a time saving. Three or fewer: cancel it, and do not blame yourself or the model, because that ratio usually means the job needed knowledge the tool did not have rather than words it could not write.
That test costs you two weeks and one subscription. The alternative, which is what the forty-tool lists produce, costs a quarter and five subscriptions and ends with the same three tools you would have picked anyway. If you want the broader argument about where AI genuinely lands in a small business week, the companion piece on taking one project from prompt to deployed shows the same one-at-a-time discipline applied to building rather than buying.