When a patient's home monitor flags a dangerous reading, how quickly does someone call them? At a patient monitoring company called Cadence, it used to take almost 2 hours. Now an AI voice assistant makes that call in 3.5 minutes. Cadence built an agent, which means software that carries out a task for you, on ElevenLabs' voice platform. When a reading is out of range, it calls the patient, asks the symptom questions, then hands over to a nurse or closes the alert.
Why it matters to you
Imagine you run a building repair firm. An alarm trips at night, and the agent calls the tenant within minutes, asks 3 set questions, and wakes an engineer only if needed. Speed also gets people to pick up. Cadence says answer rates on critical calls rose from 41% to 60%.
The risks
Cadence assumes the agent will sometimes be wrong, so it designed for that: a 'not sure' counts as a yes, and the alert stays open. It hands over to a nurse whenever in doubt, and a second check reads the call transcript before an alert is closed. The agent also sees only the health details it needs for that one call.
Try it safely
Start small: pick one repeatable call your team makes from a clear script, like a reminder or a first check. Keep a person listening to real calls early on, and turn every mistake into a new test.
Annie's take This is a clear win for voice agents, because speed changes the outcome while people still make the calls that matter. The rule worth copying: when the agent isn't sure, a person takes over.
Ask an AI chatbot how sure it is, and it'll happily give you a number, say 90%. Should you believe it? A popular post on Hacker News this week argues you shouldn't, and explains what to do instead. The author, Amrut Joshi, says a chatbot writes its confidence number the same way it writes everything else: as the most likely next words. It isn't reading a gauge. Ask again, or ask 'are you sure?', and the number often changes.
Why it matters to you
For teams using AI on customer support, that's a real trap: replies marked as very sure might go out with no human check. A wrong answer comes with a number that looks just as precise as a right one, so the score gives a false sense of safety.
The risks
The danger is trusting a number that nobody has ever checked against what really happened. And the fix takes some record keeping: you note every answer, and whether it turned out right.
Try it safely
Here's a quick test: ask your AI tool the same question 3 times, with its confidence each time, and watch the number move. The better way, the author says, is to score confidence from things you can see, like whether it found a matching help article and whether 2 answers agree. Then check every week. When your rule said 90%, was it right about 90% of the time?
Annie's take Never let a chatbot mark its own homework, because its confidence is just more words. Trust the numbers you've checked against real results.
What if AI tidied every product photo the moment you listed it, and those listings did better? At its GTC Berlin event, NVIDIA shared how European retailers are using AI, and a few of the numbers are worth a look. Photoroom's AI cleans up about 40,000 product photos a day for the resale site Vestiaire Collective, across 70 countries. NVIDIA says many of these retailers run open models, AI models a company can adapt and run on its own systems, to keep costs down.
Why it matters to you
For a small business selling online, the photo result is the one to watch: those listings got 3% to 7% more views each. Shopify says a tuned open model now powers part of Sidekick, its assistant for merchants, at 90% lower running cost and 7 times faster. And a virtual try-on service says one fashion label's sales rose 56% after it added fit previews.
The risks
These are NVIDIA's own customer stories, so expect the best results, not the average. Most of this needs a tech team or a partner to set up. It isn't something you switch on this afternoon. And an AI-edited photo must still match the real item, or you'll pay for it in returns.
Try it safely
One way to start: put your weakest product photos through an AI photo editor and compare their views for a month. If your shop runs on Shopify, Sidekick is the easy place to see this kind of AI at work.
Annie's take Copy the photo idea first, because it's cheap to test and easy to measure in views and sales. Treat the bigger numbers as vendor claims until your own shop shows them.
An AI agent that sorts your team's problems before anyone logs on sounds great. But how often does it agree with your experts? The engineering software company Rescale has shared its answer: 68% of the time, up from 44% when it started. Rescale built an agent, which means software that carries out a task for you, to sort the test failures its systems hit every night. It labels each failure, say a code bug or a flaky test, pulls the logs, and writes a report with suggested fixes for an engineer to approve.
Why it matters to you
Say you're a manager whose team starts each day sorting alerts, tickets or complaints before the real work begins. For teams like that, Rescale's result is worth noting: 20% less time with releases held up while failures were being sorted. It went from first idea to a working trial in 5 days, in a workshop with AWS.
The risks
68% agreement still means about 1 in 3 calls differs from a person's, so an engineer checks every report. It learns from past failure records, so without good records, an agent like this has little to work with.
Try it safely
Try this: list the sorting job your team repeats most, and write down the categories a person uses today. Then measure any agent against your best person's choices before you let it route anything, the way Rescale did.
Annie's take The honest 68% is why I like this story, because the team measured agreement instead of promising magic. Let the AI sort, and let people decide.
Ask an AI for a chart today, and you usually get a flat picture. What if you could click it, and ask about the part you clicked? Microsoft Research has released Flint-interactive, which is built for exactly that. It's a language that lets an AI assistant build charts you can explore: click, filter, zoom and highlight. Select part of one chart, and the table or text next to it can update too.
Why it matters to you
Picture a sales manager with a quarterly report: she clicks one region, and the other charts and the summary follow. Microsoft says assistants make these interactive charts more reliably this way, with fewer rounds of fixing. And what you select carries into your next question, so the follow-up is about the data you picked.
The risks
It's a research release, and adding it to your chat app means following a technical setup guide. A chart you can click can still mislead, so check the numbers behind anything you share.
Try it safely
If you're curious, Microsoft has an interaction gallery with examples, and guides for chat users and app builders. A good first step: ask whoever builds your reports to try one monthly report with it.
Annie's take This is early but heading the right way, because a report should answer your next question, not just your first. Worth watching rather than switching to today.
Status
Research
For
Managers who live in reports
Cost
Not stated
Useful to you
The Friday edition
The week in AI, in 1 email.
The 5 stories that mattered, 1 tool worth trying and 1 safety tip. Free, every Friday morning, New York time.