Last spring, a 40-person marketing agency walked me through its software bill. Eleven AI subscriptions, roughly $2,300 a month, and the operations lead could name a clear owner and a clear use case for four of them. The other seven had arrived as free trials that converted quietly, and nobody had audited them since.
That pattern shows up in almost every business I talk to. The tools are genuinely good. The marketing around them describes a capability and skips the limit, and the limit is exactly where your money and your team’s hours go. So here is an honest review of AI tools for business, built around the limitations rather than the launch videos.
The 80% problem: output that looks finished and isn’t
A general-purpose assistant will get a deliverable to roughly 80% in about four minutes. The last 20% takes forty. A first-draft proposal, a competitor summary, a job spec: the drafting is fast and the verification is not, and the verification is where a junior employee’s afternoon disappears.
Narrow tools behave differently. A dedicated meeting notetaker like Otter.ai, which transcribes and labels speakers, doesn’t attempt your strategy document. It captures the call, timestamps everything and pulls out action items. Because the scope is small, the failure modes are small and easy to predict. That trade, narrow and reliable over broad and unpredictable, is the single biggest thing most buying decisions get wrong.
Where a wrong answer actually costs you
Nobody minds if an AI invents a fictional backstory. People mind when it puts a statistic that doesn’t exist into a client deck, or cites a regulation with a confident-sounding article number. Language models generate plausible text, not verified text, and the tone never shifts between the two. In my own testing, the errors cluster around numbers, dates, names and anything jurisdiction-specific. The prose stays fluent throughout, which is what makes it dangerous.
Practical fix: strip every claim that carries a figure or a legal reference and check it at source. Budget real time for that. If the task is customer-facing, budget twice.
Integration friction is where adoption quietly dies
The demo connects to everything. Your setup is different: a CRM nobody has cleaned since 2021, permissions set at folder level, an SSO policy that blocks new apps by default. A tool that can’t read your SharePoint, or won’t respect folder-level access, gets flagged by IT in week two and abandoned by week six.
Ask three questions before anyone builds a workflow on top of it. Does it support SSO and SCIM provisioning? Can it inherit existing permissions rather than duplicating them into a second system? And what happens to generated content when an employee leaves? Boring questions. They decide whether the licence is still in use at month nine.
Pricing models that punish you for using the product
Seat-based, credit-based and per-minute pricing all sound reasonable on the pricing page. Run your own volume through them before you sign.
Take a transcription or voiceover tool billed by the minute. A 45-minute standup, five days a week, across four teams is roughly 3,600 minutes a month. At around $0.20 a minute that lands near $720, before anyone records a single client call. The same exercise with a credit system is harder to model, because credits expire, top-ups apply at a different rate, and the “unlimited” tier usually carries a fair-use clause that arrives later as an email from an account manager.
Worth checking too: what the annual discount locks you into, whether overage is metered or hard-blocked, and whether removing a seat mid-term frees the licence or just the person.
Data handling, and the questions vendors answer slowly
Does the tool train on your inputs? By default, several free tiers do. Does it retain prompts after deletion? Which sub-processors touch your data, and where are they based? If you handle customer records in the UK or EU, you need the data processing agreement, the sub-processor list and the transfer mechanism in writing, not paraphrased on a sales call.
Two defaults to change on day one: switch off model training for your workspace, and set a retention limit. Neither takes more than a few clicks, and both are painful to fix retrospectively.
What these tools are genuinely good at
The useful framing isn’t “is this tool good” but “is this task tolerant of a wrong first draft.” These are:
- Repetitive drafting with a human check. Briefs, meeting agendas, job adverts, first-pass email sequences.
- Structural conversion. A written brief into a 12-slide outline takes minutes with something like SlidesAI turning text into Google Slides, and the output gives a designer something to react to instead of a blank page.
- Transcription and recall. Searching a six-month-old call for the exact sentence where a client changed their mind is a real time saving, not a novelty.
- Draft voice and video. Scratch voiceover for an internal explainer via WellSaid Labs, or rough cuts from long recordings using Kapwing’s AI video editing, both work well when the deliverable is internal and speed beats polish.
- Summarising volume. Fifty support tickets into themes, or a 90-minute recording into eight bullet points.
Every item on that list has a human reviewing output before it goes anywhere public. That isn’t caution. It’s the workflow.
How to test a tool before you commit
Run twenty real tasks, not a demo
Pull twenty tasks from last month, including the messy ones: the client with the unusual naming convention, the document with three conflicting versions, the spreadsheet with two blank columns. Score pass or fail on accuracy only, and ignore how good the interface feels. A tool that fails four out of twenty on your actual work will fail more once volume picks up.
See what happens when it’s wrong
Look for an audit trail, an export that isn’t a PDF screenshot, and a way to correct bad output so the correction persists. Tools with no memory of a fix will make the same mistake forever, and your team will quietly build workarounds around them instead of telling you.
Model cost at three times current volume
Pilots look cheap because they’re small. Price the tool at the usage you expect in twelve months, once a handful of extra people discover it and start routing work through it.
The limits that won’t be fixed next quarter
Some constraints are structural rather than temporary. A model doesn’t know your pricing exceptions, why one account gets priority handling, or which of your two conflicting policies is the real one. Context windows are bigger than they were a year ago, but memory between sessions is still unreliable, and confidence is not calibrated to correctness.
There’s an organisational limit nobody budgets for either. Someone has to own each tool, review its output quality and decide when it stops earning its place. Without that, subscriptions accumulate, the average quality of work drifts down, and the invoice climbs anyway.
The businesses getting real value from AI right now don’t have the longest tool list. They can tell you, for every subscription, which person’s week it shortens and by roughly how much. If nobody can answer that in a single sentence, you already know which one to cancel.