The average mid-sized company now pays for somewhere between ten and twenty AI-powered subscriptions. Ask three different people in the building which ones they are and you’ll get three different answers. Somewhere in that stack is a $29-a-month transcription app that two people tried in February and nobody has opened since.
That’s the awkward truth about shopping for AI tools for business. Buying isn’t the hard part. Choosing is. Every vendor promises the same outcomes, the demos are built to impress, and the free trial quietly becomes an annual contract around day fourteen. The two things that actually matter, whether the tool saves real hours and whether anyone still uses it in month three, are the two things the sales page never addresses.
So here’s a way to work through it that doesn’t cost you a quarter of experimentation.
Start with the job, not the software
Spend twenty minutes listing the tasks that eat the most time in your business. Not the most strategic ones. The ones that repeat. Rewriting job specs, chasing invoice approvals, drafting the same proposal with different client names, answering the same eight support questions, reconciling a bank feed against a ledger.
Then pick one and put it through a filter: high volume, low judgement, and a clear definition of what “done” looks like. A task that happens forty times a week and takes four minutes each time is worth automating. A task that happens twice a month and requires you to weigh five competing priorities is not, no matter how good the demo looked.
If you’re a small team without a technical person, this discipline matters even more. The same advice that applies to solo operators applies at scale: start with one job, not twenty apps, and let the results earn the next purchase.
Where these tools genuinely earn their subscription
Support triage and first response
A support desk handling 1,500 tickets a month will typically find that 30 to 40 percent fall into a dozen repeat categories. A tool that drafts replies for those, with a human approving before sending, cuts handling time per ticket from six minutes to two. That’s roughly 60 hours a month back. The number to watch isn’t deflection rate, it’s how often an agent rewrites the draft from scratch.
Sales admin and follow-up
Meeting notes, CRM updates, and the “just circling back” email are the least favourite part of most sales roles. Tools that transcribe a call and produce a summary with next steps are mature and cheap. The win is small per rep, maybe three hours a week, but it compounds across a team of eight.
Back-office document handling
Invoice data extraction, contract clause flagging, expense categorisation. Unglamorous, measurable, and usually the fastest payback in the whole building. If someone is retyping numbers from a PDF into a system, that job is basically solved by commodity software at this point.
Content and marketing, with a caveat
Drafting, repurposing, and editing are legitimate uses. Publishing raw output is not. The teams getting value treat the model as a first-draft machine and keep a human editing pass that actually changes the copy. Companies that skipped that step have spent the last two years cleaning up the damage.
The cost maths most buyers skip
List price is rarely the real price. A $40-per-seat tool for twelve people is $5,760 a year before anyone has configured it, trained the team, or built the integrations that make it useful. Add an implementation consultant and a part-time admin and you’re at double. Now ask whether the hours saved clear that bar within two quarters. If the answer is “probably, once we get it set up,” that’s a no.
Five checks before you sign anything
- Trial it with your own mess. Vendor demo data is clean and your data is not. Feed it three real invoices, a genuine support ticket, and your ugliest spreadsheet.
- Read the usage pricing. Per-seat pricing is honest. Per-credit pricing can triple your bill the month something goes viral.
- Ask where your data goes. Specifically: is it used for training, how long is it retained, and can you turn both off. Get it in writing, not on a call.
- Check the exit. Can you export everything in a usable format? A tool that holds your data hostage has effectively set its own renewal price.
- Name one owner. Every tool needs a person who checks the usage dashboard monthly and kills it if the numbers are flat. Unnamed tools never get cancelled.
It also helps to benchmark against what other people are actually paying for. Roundups like the top AI tools in 2026 and what’s actually worth paying for are useful mostly for spotting the categories where you’re overpaying, not for finding a magic new app.
When off-the-shelf stops being enough
At some point the workflow you care about doesn’t map to any product. Maybe it’s quoting, maybe it’s compliance review, maybe it’s a handoff between three systems that no vendor will ever care about. That’s when the conversation shifts from buying to building.
Building well is less about the model and more about structure. The teams that succeed tend to build a scaffold that can evolve rather than hard-coding a prompt and hoping the underlying model never changes. If agents are going to touch real systems, you’ll also want a control plane for those agents before they touch anything that sends money or emails.
Budget for this honestly. A custom internal tool that saves 200 hours a year is worth building. One that saves ten is a hobby.
The people question nobody budgets for
Tools don’t implement themselves, and the person who owns them matters more than the tool itself. A capable operator with a mid-tier product will outperform a mediocre process with the best software on the market every single time. The framing that helps here is treating AI as a hire rather than a utility: it needs a job description, a manager, and a review cycle.
For anything beyond straightforward SaaS, you’ll also need to know what kind of technical person you’re actually looking for, because “AI engineer” can mean four different jobs with four different salaries.
Pick one, give it thirty days, and pick a number first
Choose the single highest-volume task on your list. Decide in advance what success looks like in numbers: tickets handled per agent per day, hours saved per week, error rate on extracted invoices. Then run it for thirty days with one owner and one dashboard.
At day thirty, one of three things happens. The number moved and you keep it. The number didn’t move and you cancel. Or nobody used it, which is the same as the second outcome but tells you something more useful about your rollout. Either way you’ve learned more in a month than a year of vendor demos would teach you, and you’ve spent a fraction of what a company-wide licence would have cost.
That loop, run four or five times a year, will build a stack of AI tools for your business that each have a reason to exist. It’s slower than buying everything at once. It’s also the only version that survives contact with a budget review.