What if every caller on a property listing got an answer at 2am, in their own language? Immobiliare.it, Italy's largest property site, says its AI phone assistants now qualify 3.3 times as many leads. Immobiliare.it built an agent, which means software that carries out a task for you, on ElevenLabs' voice platform. It answers questions on each listing day or night, collects the caller's details, and sums up every call so the estate agent can see how keen the lead is.
Why it matters to you
Say you run a lettings office with more calls than people to answer them. For a small business like that, an agent picks up every call and passes on the warm leads. Callers also shared more by voice: the company says people were 74% more likely to leave a phone number for a follow-up. People looking for a home also felt easier giving personal details by voice than by typing, so the leads got better. And more than 70% of private sellers chose to switch it on.
The risks
This is ElevenLabs telling its own customer's story, so read the numbers as a best case. And with more than 1.2 million listings, a wrong answer repeated at that scale adds up fast. Tell callers they're talking to an AI, and decide who can see the personal details they share. Immobiliare.it also worked towards an outside safety standard for voice agents, and that's worth asking any vendor about.
Try it safely
Start small: give an agent one listing, or one product line, and let it answer the common questions only. Read the call summaries every week, and fix the answers it gets wrong.
Annie's take This is a smart way to grow, because they started with one agent, proved it worked, then added support, search and outbound calls. One small agent you can measure beats a big one you can't.
How long does it take your team to find out why something went wrong? At Bayer Crop Science, it often took over an hour. Now it has AI agents that answer about 80% of staff questions without an engineer stepping in. Bayer's plant breeders run hundreds of data models, and explaining a result meant digging through logs, dashboards and documents. So Bayer built 22 specialist helpers, each for one area, plus a lead agent that sends every question to the right one.
Why it matters to you
If you run a business with lots of systems, think of all the 'where is this' and 'why did that break' questions your experts field every day. For teams like that, Bayer says a diagnosis that took over an hour of hunting through logs now takes under 10 minutes. That's time your best people get back for the work only they can do.
The risks
Bayer found a newer AI model wasn't a simple swap: it broke how the agents used their tools, so the team kept the older one. Another trap: failed tasks were being counted as successes, because the system still replied when the job went wrong. Each agent also gets only the tools it needs, which keeps it focused. And this was a big build, with AWS engineers in meetings 3 times a week.
Try it safely
Here's a quick test: list the 10 questions your experts answer most often, and where each answer lives. If most answers sit in a few documents, a simple assistant may be enough before you build a whole team of agents.
Annie's take This is one of the more useful agent stories, because it measures what matters, questions answered and time saved. Start with one helper, and only split it up when it grows too big.
Most voice assistants turn your words into text, and lose your tone along the way. A small research lab at Block says its new model, Dessa, learned to hold a conversation from sound alone. Dessa was trained only on recordings of people talking, with no text at all, and Block says it needed hundreds of times fewer words than text-based AI. It listens and talks at the same time, so it can laugh mid-sentence, cut in, or say 'got it' to show it's following.
Why it matters to you
For a business, the promise is a phone assistant that hears when a customer is upset, not just what they said. In a busy contact centre, that could mean fewer customers hanging up angry. Picture a support line where a frustrated sigh changes how the call goes, the way it would with a good member of staff. Block also points to safety: today's voice AI can't tell when someone on the line is in serious pain.
The risks
It's a research release. You can listen to recordings, but you can't use it, and only a small group of safety researchers gets early access. A voice this human raises a fair worry: people may not realise they're talking to an AI. And its makers say there are risks they haven't found yet, which is why they've asked outside experts to look.
Try it safely
Nothing to set up yet. Listen to the sample conversations on Block's Dessa page, and notice the pauses and the timing. Then call your own support line the way a customer would, and compare how natural it feels.
Annie's take This is early research, but it points the right way, because tone carries meaning that text throws away. Watch it, but don't plan around it yet.
Where are your customers, and which delivery routes cost you the most? A map often answers that faster than a spreadsheet. Amazon's Quick Sight dashboards can now draw single locations and whole routes on a map, not just regions. Quick Sight is Amazon's tool for business dashboards, part of Amazon Quick, which bundles in AI agents that help with research and reports. You upload a map file that marks each store or traces each route, then link it to your sales or delivery data.
Why it matters to you
Imagine you run a few bakeries across a city: each shop is a dot, coloured by sales, so the slow one stands out at once. Or a delivery firm: every route is a line, coloured by any field in your data, such as the driver or the cost. Amazon's own example maps the flights leaving New York's JFK airport on a single day.
The risks
It's not magic: someone has to prepare the map file and keep it matched to your data. Files are capped at 100 megabytes, and each layer shows one kind of shape at a time, so busy maps take planning. And a map can mislead: a big dot looks important even when the number behind it is small.
Try it safely
One way to start: if your team already uses Quick Sight, map your top locations and look for the odd one out. If you don't, try a map in the tool you already have before you pay for anything new.
Annie's take This is a small update with a real payoff, because location patterns hide in tables and jump out on a map. Just make sure the numbers behind the dots are right.
Do you type the same instructions into your AI tool every day? OpenAI's coding assistant now lets you save them. The latest update to Codex adds numbered slots for reusable prompts and commands, with a keyboard menu to pick one. Codex is an agent, which means software that carries out a task for you, and this one helps people write and fix code. You can also switch the model, how hard it thinks, and its permissions with one short command, instead of digging through settings.
Why it matters to you
Here's where it lands for anyone who uses AI at work: your best prompts are worth keeping, so stop retyping them. For teams that repeat work, think of a finance team running the same month-end check, or a marketer with a favourite brief: one saved prompt, the same quality every time. The update also hides passwords and keys in some of the commands it runs, so they're less likely to leak.
The risks
That hiding only covers supported commands, so never paste a password into any AI tool and assume it's safe. Coding agents can change real files, so give them the fewest permissions that get the job done. Saved prompts can also go stale, so check them when your tools or rules change. And this is a tool for people who are comfortable typing commands, not a point-and-click app.
Try it safely
A good first step: save your 3 best prompts in a shared document today, whatever AI tool you use. If your developers use Codex, ask them to save a preset for the job they repeat most, and to review each change before they approve it.
Annie's take Saved prompts sound small, but they're how a team gets the same good result twice, because a great prompt shouldn't live in one person's head. Start your prompt library now.
Status
Shipped
For
Developers, and anyone who reuses prompts
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.