Actually Useful #22
AI tools that actually work at work | Tuesday, May 19, 2026 | 4 min read
The AI tools you’ll actually use at work probably aren’t being built by companies trying to out-compete OpenAI. They’re being built by small teams solving one specific problem you have right now — and this week had a few worth knowing about.
Tool of the Issue
Trainer — Show AI what you do, and it learns to do it for you
Best for: Operations managers and admin teams tired of repetitive data entry
Trainer takes a different approach to building AI automation: instead of writing prompts or configuring code, you record your screen while doing a task, and Trainer uses that recording to train an AI agent to replicate it. The idea is genuinely appealing. Most people on ops and admin teams can’t write a prompt that reliably describes a multi-step workflow, but they absolutely can show someone how to do it. Screen recording as training input is a real insight — it meets non-technical users where they already are.
The trade-offs are real, though. Visual AI training is computationally intensive, and the gap between “works in a demo” and “works reliably on your actual messy data” is significant with this kind of tool. The product doesn’t yet publish accuracy rates or explain how it handles edge cases — the moments when your screen looks slightly different than the training recording. It’s also entering a space where established robotic process automation tools have years of reliability testing behind them.
That said, if execution is solid, this fills a genuine gap. Worth testing on one low-stakes repetitive task before rolling it out more broadly.
Usefulness score: 72/100
Kin Health — A notetaker built for patients, not just doctors
Best for: Healthcare operations teams and patient advocates managing care coordination
AI meeting notetakers are a proven category — tools like Otter and Fireflies have demonstrated that automatic transcription and summarization saves real time in real workflows. Kin Health applies that same logic to medical appointments, with a specific angle that makes it more interesting than a generic transcription tool: it’s built to share summaries with family members and caregivers, not just the patient.
That’s a meaningful distinction. Healthcare documentation is one of the most burdensome parts of clinical work, and the coordination problem — making sure a patient’s spouse or adult child actually understands what was discussed — is genuinely underserved. The $9M funding round signals that investors see a real market here.
The honest limitations: HIPAA compliance and medical liability are not small hurdles, and the available information doesn’t detail how Kin handles sensitive health data or what their regulatory pathway looks like. AI summary quality also varies considerably when medical jargon is involved. If you work in healthcare operations, it’s worth watching closely — but verify the compliance story before deploying it anywhere near patient data.
Usefulness score: 78/100
Workflow Win
The tools that will actually help your team aren’t competing with OpenAI — they’re built on top of it
Here’s a pattern worth paying attention to: OpenAI and Anthropic now capture 89% of AI startup revenues. At the same time, a steady stream of new products — Chert for customer texting, Haystack for code review, Trainer for agent training — are launching without any claim to building their own models. They’re all building on top of the same few APIs.
And then Anthropic went and acquired Stainless, a developer tools startup previously used by OpenAI, Google, and Cloudflare. Stainless builds tooling that makes it easier to build on top of AI models — not to compete with them. Even Anthropic, one of the two companies consolidating nearly all the revenue, is investing in reducing the friction of building on top of AI, not just building better AI.
The real competition in AI has shifted from “who builds the best model” to “who builds the most useful interface to it.”
This matters for your team in a few concrete ways.
First, the AI tools you’ll actually use at work won’t come from companies trying to out-train OpenAI. They’ll come from teams who deeply understand one specific workflow problem — medical notetaking, customer text responses, code review — and built something focused on that. The model underneath almost doesn’t matter anymore; what matters is whether the product fits your process.
Second, Anthropic’s Stainless acquisition tells you something about where the bottleneck actually is. It’s not model capability. It’s developer velocity and integration ease — how quickly teams can connect AI to the tools they already use. That friction is the problem being solved right now, and it’s why you’re seeing so many “last mile” tools launching.
Third — and this is the practical takeaway — when you’re evaluating an AI tool, stop asking “which AI does it use?” and start asking “does it solve my specific problem reliably?” Because within a year or two, most of these tools will be running on the same handful of models anyway. The differentiation will be entirely in the workflow layer.
If you’re on a non-technical team trying to figure out where to start with AI: look for tools built for your specific job function, not tools built to be impressive. The impressive ones are already commoditizing. The useful ones are just getting started.
Skip This One
Sony’s PlayStation-to-PC port strategy — a useful reminder about expansion initiatives
A high-profile multi-year initiative that couldn’t justify itself when the numbers came in
This one isn’t an AI tool, but it’s worth a moment of your time if you’re thinking about technology expansion initiatives of any kind. Sony spent several years porting PlayStation exclusives to PC, built real goodwill in the process, and then reversed course — announcing at the executive level that the strategy wasn’t delivering sufficient value to continue.
The warning signs were quiet: a shift from expansion to consolidation, an ROI conversation that apparently didn’t go the way the initiative’s advocates hoped, and a departure from a stated multi-platform vision. None of this was visible until the reversal was announced.
The lesson for teams evaluating new tools or platforms: distribution and expansion strategies need to be validated against actual business outcomes on a regular basis, not just at launch. Market enthusiasm and user goodwill are real things — but they’re not the same as a sustainable return. If you’re mid-way through an AI tooling rollout that felt right six months ago, it’s worth asking whether the outcomes are matching the original case you made for it.
Revisit when: Sony clarifies what metrics would change their calculus, or when a competitor demonstrates a PC port model that works financially.
What’s one repetitive task your team does every week that you’ve never thought to automate — and why not?
Tested by AI. Approved by people who just want to get work done.
Know someone who said “we should be using AI more” in a meeting last week? Forward this.
On Substack? A quick Recommend helps more teams find us — takes 2 seconds.
Want the skeptic’s take on what’s real vs. hype?
Overfitted cuts through AI noise with data and zero patience for buzzwords. Mon/Wed/Fri.
Build your own?
The Autonomous AI Newsletter Pipeline Guide: GumRoad or Dodo
How we source, classify, synthesize, and publish at scale.