Eleven new AI tools landed in my inbox before 9am last Tuesday. Four promised to replace my entire workflow. Two turned out to be the same product with a different logo. One was genuinely good: a small transcription model that runs on my laptop and never sends an audio file to a server.
That is roughly the ratio. One in ten earns a bookmark. The rest are a landing page, a waitlist and a founder posting “we’re live” into four different Slack groups.
Finding new AI tools isn’t the hard part. Directory sites list more than twenty thousand of them. The hard part is working out, in under half an hour, whether the thing in front of you will still be worth using in six weeks.
The launch cycle has outrun the review cycle
Most reviews get written on day three, when the tool is at its best. Servers are empty, founders answer support tickets personally, nothing has broken yet. By week six the free tier has been trimmed, the pricing page has three new tiers, and the feature that sold you on it has moved behind the top plan.
That gap is why launch coverage works better as a list of candidates than as a buyer’s guide. You still need a filter of your own. The categories below are where the current batch has genuinely improved.
Three kinds of tool that earned a permanent tab
Agents that finish a job instead of describing it
For a long time an “AI agent” meant a chatbot with a to-do list attached. The current generation behaves differently in one measurable way: hand it a goal, leave for twenty minutes, come back to an actual file.
Last month I gave one a simple brief. Find forty suppliers in a niche, with contact details, website and a rough shipping estimate, and put it in a spreadsheet. It returned thirty-seven rows. Two links were dead, the formatting was a mess, and one supplier had shut down in 2023. Cleaning it up took eight minutes. Doing the job by hand would have eaten most of an afternoon.
That’s the honest bar now. Not “did it get everything right” but “was fixing its mistakes faster than starting from zero.” Agents also live or die on what they remember between sessions, and I tested four ways to give an AI assistant a memory: the priciest option came last.
Small models that run on your own machine
An 8-billion-parameter model, quantised to around 4GB, runs comfortably on a laptop with 16GB of RAM. On an M2 MacBook it produces 30 to 40 tokens a second, which is faster than most people read. It won’t draft your strategy document. It will summarise a 60-page PDF, tag five hundred support tickets, rewrite product descriptions and translate a contract. Offline, with nothing leaving the machine.
For anyone handling client data, that last part is the whole argument. No data processing agreement, no clause nine about using your inputs to improve the service.
The quality gap is real and you will notice it. Local models are weaker at long reasoning chains and worse at fiddly instructions. Stop treating them as a cut-price frontier model and point them at narrow, repetitive jobs instead.
Image generators that hold a style across a set
The interesting shift here isn’t resolution, it’s consistency. Reference images and style locking mean you can generate twelve visuals for one blog post and have them look like they came from the same art director rather than twelve strangers. There’s a walkthrough on building a cohesive blog image set with a free AI image generator if you want the prompt structure that makes it work.
Red flags in almost every weak launch
After enough of these, the tells get easy. Three or more of the following and I close the tab.
- The demo is a video, not a product. No free tier, no sandbox, no way to touch it before a call with sales.
- Every screenshot shows a generic marketing task. “Write a blog post about productivity” tells you nothing about what the tool is good at.
- Pricing is in credits. Credits are deliberately awkward to convert into a monthly cost, and that is usually the point.
- No export. If your work can’t leave in a standard format, you’re not a customer, you’re inventory.
- The team page is empty. Or it lists four advisors and one anonymous founder.
A twenty-minute test before you pay
Run this first. It’s deliberately boring.
- Bring real work, not a sample task. An actual messy email thread or half-built spreadsheet.
- Give it one job and time yourself fixing the output.
- Repeat with slightly different inputs. Tools that only work once are prompt-shaped, not product-shaped.
- Try to get your data out. Export to CSV, copy the raw output, download the file.
- Find the free tier’s limit, then halve it. That’s closer to what you’ll get after a month.
Twenty minutes separates a tool that saves an hour a week from one that costs you twenty minutes every time you open it.
Credits, seats and the pricing line worth reading twice
Nearly every new AI tool now prices in credits instead of seats. The entry tier looks cheap at eight or ten dollars. The number that matters is cost per finished task, and that isn’t on the pricing page. It’s in the help docs, under something like “how credits are consumed,” where you discover a long document costs fourteen credits and your plan includes six hundred a month.
Do that arithmetic before you commit. A breakdown of which AI tools are actually worth paying for in 2026 puts real monthly figures against the popular options, which beats another feature comparison table.
Five tools that fit together beat fifty that don’t
The most common failure isn’t picking a bad tool. It’s picking eleven good ones that don’t talk to each other. You become the integration layer: copying output from one window, pasting it into another, reformatting by hand three times a day. That’s not a workflow, it’s a commute.
A setup that works tends to look dull. One model for drafting, one for research, one for images, one place where everything lands. If you want a structure to copy, this guide to building an AI tool stack that actually saves you time is a reasonable blueprint.
Where the next batch is heading
Two things are worth watching over the next six months. The first is on-device agents: models small enough to sit on a phone or laptop with access to your calendar, files and messages, handling small tasks without a round trip to a data centre. Apple, Google and a cluster of startups are all pointing the same way. If that lands properly, the “which subscription” question gets a lot less interesting, because the assistant already lives where your work does.
The second is evaluation. Right now you judge a new AI tool on vibes and a timed test. A growing number of products ship with scored test sets, so you can see that a transcription model hits 94% word accuracy on your accent and 89% on a colleague’s. That number is worth more than any benchmark on a landing page.
Neither is here yet. Until they are, the filter stays the same: try it on your own work, time the cleanup, and ask whether you’d still open it in a month once the novelty wears off.