I’ve paid for eleven AI subscriptions in the past two years and cancelled eight of them. The reason was almost never that the tool was bad. It was that one or two headline features were genuinely excellent and everything else was padding I never opened twice. A features review is only useful if it draws that line clearly, and most of them don’t.
So here’s the split: the capabilities that changed how I work, the ones that looked great in a launch demo and did nothing after week two, and a short test you can run before handing over a card.
The features that actually earn their subscription
Real context limits, not advertised ones
Ask a tool to summarise a 60-page contract and you’ll find out quickly whether the context window is real. Vendors advertise 200,000 tokens; what matters is whether quality holds at page 45. In my testing, several tools that claimed huge windows started paraphrasing loosely and inventing clause numbers once a document passed roughly 40 dense pages. A model that handles 30 pages perfectly is more useful than one that fumbles 150.
File handling matters just as much. Drag-and-drop for PDF, XLSX and CSV with column headers preserved is a different product from a chat box that only accepts pasted text.
Output control: formats, structure, and exports
The tool is only half the job. Can it export to DOCX with headings intact? Can it give you a PPTX you can edit rather than a flat image? Can it output SRT captions at the right frame rate? These sound like small details until you spend 25 minutes copy-pasting a response into Google Docs and rebuilding the formatting by hand.
This is why I keep coming back to tools that produce something you can use immediately. If you’ve ever needed a diagram from a wall of text, Napkin AI turns a messy paragraph into a clean diagram in four steps, and the export options are the reason it stays in my rotation.
Integrations and where the tool lives
A browser tab you have to remember to open is a tool that gets used twice a week. An extension that sits in the sidebar of whatever you’re already doing gets used hourly. The same logic applies to code: JetBrains AI Assistant works best when it stays inline in the editor, suggesting a fix where the error actually appears rather than in a separate chat panel you have to describe the problem to.
Privacy and data retention
Check whether your inputs train the model. Check whether there’s a delete-my-history switch. For anything involving client work, contracts, or health information, this isn’t a nice-to-have. Lightweight options exist too: DuckDuckGo’s private AI chat keeps queries anonymous, which is often enough for quick research where you don’t want a permanent record attached to an account.
Media features that survive real footage
Video and audio features are the easiest to fake in a demo. The real test is a 12-minute recording with background noise, three speakers, and a stumble at minute four. Auto-cut silence, subtitle generation with decent punctuation, and the ability to fix one word without re-rendering everything are the features that matter. Kapwing’s AI editing setup handles that middle ground well, particularly for people who don’t want to learn a timeline editor.
Feature bloat you can safely ignore
Launch pages are long for a reason. Three categories consistently fail to justify their space:
- Autonomous “agent” modes. Chaining five steps unsupervised sounds great until step three goes sideways and you have to audit everything anyway. Supervised, single-step actions are more reliable.
- Template galleries with 500 entries. You’ll use four. The other 496 exist to fill a marketing screenshot.
- Voice and avatar generation. Fun for a week, rarely part of a paying workflow unless you’re producing training videos at volume.
None of these are worthless. They’re just not the reason to subscribe, and they shouldn’t be the reason you pay for a higher tier.
A 20-minute test that beats any feature list
Pick one real task you do weekly. Not a sample prompt from the onboarding flow, an actual piece of work with your own messy inputs. Then run it four ways:
- Feed it your ugliest document (scanned PDF, inconsistent formatting, a table with merged cells).
- Ask for output in the exact format you need, and time the cleanup.
- Interrupt mid-task with a correction and see whether it holds the change.
- Try it on a second device, since half of these tools behave differently in the mobile app.
That’s roughly 20 minutes and it tells you more than three hours of YouTube reviews. If the cleanup takes longer than doing it yourself, the feature is decoration.
What six months of paid testing kept showing me
Tools tend to win on one narrow strength and lose everywhere else. Writing assistants are a good example: the copy quality can be excellent while the pricing model quietly punishes you for using it. My notes on Jasper’s strengths, costs, and who should skip it land on exactly that tension. Strong brand-voice controls, real friction on the billing side.
After enough of these, the pattern became obvious: I was paying for two features per tool and subsidising six I ignored. That’s the whole argument for a feature-first review rather than a brand-first one. The logo on the tab matters far less than whether the thing produces a file you can send to a client.
Do the cost-per-feature maths before you upgrade
Most subscriptions are $20 to $40 a month, and the annual commit saves you around 17%. Fine. But work out what you’re actually buying. If a $30 plan gives you 15 features and you genuinely use three, you’re paying $10 a month per useful feature, or $360 a year for something a $12 tool might cover.
Run the numbers against hours saved instead of features unlocked. A tool that saves you four hours a month at a $40 hourly rate is worth $160. A feature that saves 10 minutes is worth about $6.67. Most of the shiny add-ons land in the second bucket, and once you see the number written down, the upgrade button gets a lot less tempting.
Buy the workflow, not the roadmap
Every tool on the market promises features that arrive “soon” and a roadmap that will supposedly unify everything. Ignore it. Judge what’s in front of you this week, on your actual files, and keep the number of subscriptions low enough that you can name every feature you’re paying for without checking a pricing page.
Two well-chosen tools with deep, reliable features will carry more of your work than six experimental ones with impressive landing pages. Cancel the rest, keep the two you’d notice immediately if they disappeared, and re-test the field in six months. The good tools get better, the filler gets cheaper, and your stack stays small enough to actually learn.