<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Elegant Wolves Labs on Medium]]></title>
        <description><![CDATA[Stories by Elegant Wolves Labs on Medium]]></description>
        <link>https://medium.com/@elegant-wolves-labs?source=rss-399d456db99f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*BsVins2UK1BH0Kfoh-ChBw.png</url>
            <title>Stories by Elegant Wolves Labs on Medium</title>
            <link>https://medium.com/@elegant-wolves-labs?source=rss-399d456db99f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 22:58:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@elegant-wolves-labs/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The next caller might be someone’s AI]]></title>
            <link>https://elegant-wolves-labs.medium.com/the-next-caller-might-be-someones-ai-46aecfcdb5d7?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/46aecfcdb5d7</guid>
            <category><![CDATA[local-business]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[customer-service]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[google-maps]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 16:06:51 GMT</pubDate>
            <atom:updated>2026-11-19T16:06:51.590Z</atom:updated>
            <content:encoded><![CDATA[<h4>Google Maps and Meta’s Muse now phone local businesses for their customers. Here’s what I’d set up before the first one reaches your front desk.</h4><figure><img alt="An auto mechanic in blue overalls talking on a mobile phone in a repair workshop, with a car raised on a lift behind him." src="https://cdn-images-1.medium.com/max/1024/1*ts9IV2h2YamZEzZ-PLkbZA.jpeg" /></figure><p>Picture a Saturday morning at a small auto shop. Two bays are full, the coffee is cold, and the phone rings.</p><p>The voice on the line is polite and a little too even. It says it’s calling on behalf of a customer who wants an oil change and a tire rotation, ideally before noon today, and asks whether there’s an opening and roughly what it would cost.</p><p>The person at the counter has a second to decide what this is. A robocall? A scam? A competitor fishing for prices? Most people in that seat would do what they’ve done for years with strange calls. They’d hang up.</p><p>As of this week, that reflex costs real jobs. The caller isn’t selling anything. It’s a customer’s assistant, and on the other end of it is a person with a car, a free morning and a credit card.</p><h3>What changed in the last ten days</h3><p>Two announcements landed close together, and I think they belong in the same conversation.</p><p>On November 12, Meta opened phone calls in Muse, its consumer AI agent, to paid subscribers in the US. <a href="https://techcrunch.com/2026/11/12/muse-can-now-place-calls-to-u-s-businesses-if-you-pay/">TechCrunch reported</a> that Meta pitches it for the calls people would rather not make, such as insurance, cable, the DMV or a restaurant without online reservations. According to <a href="https://www.meta.com/help/subscriptions/1021145227643680/">Meta’s help page</a>, the $20 a month plan includes 10 phone calls a day and the $100 plan includes 20. Meta says Muse has more than 10 million users. Not all of them pay, but the ones who do now have a patient assistant willing to sit on hold for them. <a href="https://www.engadget.com/2315525/metas-muse-can-now-make-calls-on-your-behalf-if-you-have-a-paid-subscription/">Engadget described</a> one early user whose agent called a car dealership, waited on hold for about 15 minutes and came back with the next open service slot.</p><p>Four days later, on November 16, Google added appointment finding to Ask Maps, the chat built into Google Maps. In <a href="https://blog.google/products-and-platforms/products/maps/ask-maps-holiday-travel-features-2026/">Google’s words</a>, Ask Maps can now find beauty, pet care, auto and home repair appointments online and call local businesses for you. Their example request is “Find me a 4 p.m. nail appointment near me.” It’s launching first in the US, on Android and iOS, just ahead of the holidays.</p><p>Google has been placing calls like this for a while through Search, and Pixel owners already have a feature that calls businesses for them. What’s new is the reach. Google Maps is the app people already open when they want a nearby shop. Putting an agent there, one that phones around for open slots, moves this from a curiosity to something a nail salon in a mid-size town will hear on an ordinary Tuesday.</p><h3>This isn’t a robocall, even if it sounds like one</h3><p>The most useful line I read this week was buried near the end of the TechCrunch piece. Many businesses and receptionists hang up on calls from AI agents as soon as they realize there’s no person on the line.</p><p>I understand why. Small businesses get buried in spam calls, and the instinct to protect the front desk is healthy. But the incentives here are the opposite of a robocall. A robocall wants something from you. This caller wants to give you business. Hanging up doesn’t stop the agent. It moves on to the next shop on its list, and the customer never learns your name.</p><p>It also helps to know that these calls won’t all sound alike. Google’s <a href="https://support.google.com/business/answer/7690269">Business Profile help page</a> says its automated calls are recorded and that the caller is Google acting for a customer. In the Muse call Engadget described, the agent said it was calling for someone and that it was transcribing the call, without spelling out that it was an AI. As the voices get better, your staff won’t reliably be able to tell. So “hang up on anything that sounds like a bot” stops working as a rule. You need a better one.</p><h3>How these agents decide who to call</h3><p>Before you think about the phone, think about how the agent picked your number in the first place.</p><p>Google hasn’t said how Ask Maps chooses which businesses to check. <a href="https://www.searchenginejournal.com/google-ask-maps-appointments-business-calls/597678/">Search Engine Journal</a> noted that when AI calling first launched in Search last year, a Google spokesperson said it picked businesses based on traditional local search rankings. If that carries over, your Business Profile still does the heavy lifting. Your category, hours, reviews, services and booking link are the agent’s whole picture of you.</p><p>The same help page says that if you work with an online booking provider, Google may book through that partner instead of calling. So there are two ways in. If your open slots are visible online, the agent might not need to call at all. If they aren’t, the phone is the fallback, and your front desk becomes the booking system.</p><p>We wrote recently on the EWL blog about <a href="https://elegantwolves.com/ai-agents-local-booking-forms/">AI agents booking appointments through website forms</a>, and how they get stuck on date pickers and buried required fields. The phone version of that problem is simpler and harder at once. Nobody has to fix a form, but somebody has to pick up and answer well.</p><h3>The setting most owners don’t know they have</h3><p>Here’s the concrete part. If you have a verified Google Business Profile, go to More, then Business Profile settings, then Advanced settings. Look for a section called “Google automated calls and text messages.” Inside it is a toggle labeled “Bookings and inquiries from customers.”</p><p>Google’s help page says these calls are free for both sides, that Google limits repeat calls for the same information and avoids late night and early morning hours, and that the feature works in every US state except Indiana, Louisiana, Minnesota, Montana and Nebraska. You can also tell a Google caller to stop calling during the call itself.</p><p>One honest caveat. Search Engine Journal pointed out that Google’s Ask Maps announcement doesn’t say whether Ask Maps calls follow this toggle. I’d still set it on purpose, because it’s the only documented control, and I’d check again in a few weeks as Google fills in the details.</p><p>That toggle covers Google only. There’s no equivalent switch for Muse, or for the next agent that launches, since those calls come in as ordinary calls. For everything outside Google, the answer lives with whoever picks up the phone.</p><h3>Should you turn it on or off?</h3><p>My default for most appointment businesses is to leave it on and get ready for it. The case for that is simple. These callers come with a customer who has already decided to book something today. For a groomer, a detailer or a handyman, that’s about as warm as a lead gets.</p><p>There are good reasons to turn it off, though, and I wouldn’t argue with an owner who did.</p><p>If your work needs a real conversation before you can quote it, an automated “how much for a water heater?” call wastes everybody’s time. A plumber can’t give an honest price without asking questions the agent likely can’t answer. If you’re already fully booked through the holidays, extra calls are noise. And if your front desk is one person who also does the work, every interruption has a cost.</p><p>I’d also watch for price shopping at scale. An agent with ten calls a day and infinite patience can call five shops in an afternoon, ask each the same question and report back the cheapest. That’s fine for a customer. For you it means your phone quote is now a line in a comparison table, read without your tone of voice or your explanation of why you cost a bit more. If that worries you, the fix is to decide what you say on the phone about price, not to pretend the calls aren’t happening.</p><h3>A short policy for calls from AI</h3><p>Here’s what I’d write down and tape next to the phone. Keep it to one page. If it takes longer than a minute to read, nobody will follow it.</p><p>First, the default is to treat the caller like a customer’s assistant, because that’s what it is. Be polite, be brief and answer the question it asked. Don’t hang up because the voice sounds off.</p><p>Second, confirm with the human, not the agent. Take the booking, then confirm it the way you’d confirm any appointment, by text or email to the customer’s own contact. If the agent can’t give you a name and a phone number or email for the real person, you don’t have a booking yet. You have a pencil hold, and you should treat it that way.</p><p>Third, decide what never happens on these calls. For most shops, that list is short. No card numbers or deposits over the phone with an automated caller. No quotes on work you haven’t seen. No changes to an existing customer’s account, their address or who’s allowed to pick up a car or a pet. If a request touches money or someone’s private details, the answer is “we’ll confirm directly with the customer.”</p><p>Fourth, write down what you said. When a person calls, they remember the conversation too. When an agent calls, the customer gets a summary written by the agent. If your staff quoted $89 and the summary says $79, you want a note in your calendar or booking system so the conversation at the counter goes smoothly.</p><p>Fifth, give price answers you’re happy to see in a comparison. “Our standard oil change starts at this price and most cars land around that” works better than a single number with a dozen asterisks you never say out loud.</p><p>None of this is complicated. The goal is a rule your newest employee can follow on a busy Saturday without asking you.</p><h3>What I’d check this month</h3><p>If you run a beauty, pet care, auto or home repair business in the US, these are the things I’d do before December gets busy.</p><p>Set the Google toggle on purpose, and make a note of the date so you remember you chose it.</p><p>Fix the facts the agent will read. Holiday hours, services, price ranges and your booking link should all match what your staff says on the phone. An agent that reads “open until 6” and calls at 5:30 to book will create a confused customer if you close at 5.</p><p>Call yourself. Ask Ask Maps, or whatever assistant you use, to find an appointment at your own shop, and listen to what happens from both sides. You’ll learn more in ten minutes than from any article, including this one.</p><p>Brief your team in one sentence. “Some callers this season will be AI assistants booking for real customers. Help them, then confirm with the customer directly.”</p><p>Then, a few weeks in, look back. How many of these calls did you get? Did they turn into appointments, or no-shows? Were they mostly for one service? That’s real information about what people want from you, and you’ll get it before most of your competitors start paying attention.</p><h3>If you already use an AI receptionist</h3><p>A lot of small businesses have quietly added AI phone answering this year. If that’s you, there’s an odd new situation to plan for. Your AI will answer calls from someone else’s AI.</p><p>That will work fine for simple questions, and it will go badly in ways that are hard to see. Two agents can politely agree on a booking that neither human wanted. They can loop on a menu. They can each assume the other one confirmed the details.</p><p>So keep a person on the confirmation step. Make sure your system sends a confirmation to the real customer and flags any booking made by an automated caller so someone glances at it. And make sure your own assistant says it’s automated and hands off to a person when the request is unusual. That’s good manners with humans, and it’s even more useful when the caller is a machine taking notes.</p><h3>The bigger picture, briefly</h3><p>I don’t want to overstate this. Most people will keep booking haircuts the way they always have, and plenty of agent calls will fail in clumsy ways for a while. But the direction is clear. The big platforms want their assistants to do errands, and for local businesses, errands mean phone calls and appointment slots.</p><p>The businesses that do well here won’t be the ones with the fanciest tools. They’ll be the ones whose public information is right, whose staff knows what to do when an odd call comes in, and who still put a person in front of the commitment. That’s mostly ordinary good operations, with one new kind of caller to plan for.</p><p>Thanks for reading. If you’ve already had one of these calls at your business, I’d like to hear how it went, the good and the strange. And if you want more practical notes like this on AI for local businesses, there’s more at <a href="https://elegantwolves.com/">elegantwolves.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46aecfcdb5d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The alarm that went to an inbox]]></title>
            <link>https://elegant-wolves-labs.medium.com/the-alarm-that-went-to-an-inbox-d30959d460d6?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d30959d460d6</guid>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business-continuity]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 16:09:20 GMT</pubDate>
            <atom:updated>2026-11-17T16:09:20.236Z</atom:updated>
            <content:encoded><![CDATA[<h4>Last week’s Microsoft 365 outage lasted almost a day. For most of us that meant a quiet inbox. At one hotel in Germany, it meant a safety warning nobody saw. Here’s how I’d find the alarms in your own business that only live in email.</h4><figure><img alt="A cafe owner in an apron taking an order on a corded phone at the counter and writing it in a notebook." src="https://cdn-images-1.medium.com/max/1024/1*auCk5G4Z8qqAZb0xwpFeSw.jpeg" /></figure><p>Last Thursday, a little before lunch on the East Coast, a lot of inboxes went quiet at once.</p><p>At first most people assumed it was them. They refreshed Outlook, restarted a laptop, asked the person at the next desk. Then the group texts started. Is your email down? Is Teams working for you? By early afternoon it was clear this wasn’t one office. Microsoft 365 was having one of the worst days in its history.</p><p>For most businesses the outage was an annoyance with a cost attached. Orders sat unread, meetings moved to phone calls, and somebody spent Friday morning digging through a backlog. Irritating, and survivable.</p><p>One story from that day has stayed with me, though, and I think it’s worth a full essay. It’s about a pool.</p><h3>What happened at the hotel</h3><p>On Friday, November 13, the automatic chlorine dosing system at Hotel Ling Bao, part of the Phantasialand theme park in Brühl near Cologne, malfunctioned and pumped far too much chlorine into the pool. The hotel had a monitoring system watching for exactly this kind of fault. It noticed the problem and raised the alarm.</p><p>By email.</p><p>German public broadcaster <a href="https://www1.wdr.de/nrw/rheinland/rhein-erft-kreis/technischer-defekt-chlor-phantasialand-1-100.html">WDR reported</a> that the warning emails never reached staff because Outlook was down. Nobody knew anything was wrong until guests in the pool area started having trouble breathing. According to <a href="https://www.theregister.com/saas/2026/11/16/microsoft-outlook-outage-blamed-after-15-injured-in-hotel-pool-chlorine-incident/5310087">The Register</a>, the local fire department sent around 40 people to the scene and checked 15 guests, most for irritated airways. Four went to the hospital and one stayed in for observation. The pool was closed, the water replaced, and it later reopened.</p><p>A cybersecurity architect WDR spoke with, Stephan Skrobisch, didn’t blame Microsoft for this part. He put it on the way the system was set up. Nobody had added a second way to raise the alarm, something as simple as a sound in the building. His point, roughly translated, was that a warning system meant to protect people shouldn’t depend on one email inbox.</p><p>I agree with him, and I’d go further. Almost every business I know has some version of that pool alarm. Usually the stakes are money, customers or a deadline rather than chlorine, and the warning sits in an inbox, waiting for a day when the inbox doesn’t work.</p><h3>What actually broke at Microsoft</h3><p>It helps to understand the outage itself, because the details explain why “it’ll be back in an hour” was the wrong assumption.</p><p>Microsoft’s <a href="https://next.ink/wp-content/uploads/2026/11/PostIncidentReport.pdf">preliminary post-incident report</a> puts the outage at just under 24 hours, from 16:20 UTC on November 12 to 16:10 UTC on November 13. That’s roughly 11:20 AM Thursday to 11:10 AM Friday Eastern time. Exchange Online took the worst of it, and the list of affected services ran long, including Teams, SharePoint, OneDrive, Copilot, Planner and Intune.</p><p><a href="https://arstechnica.com/gadgets/2026/11/microsoft-details-how-three-problems-combined-for-a-global-outage/">Ars Technica’s summary</a> describes three problems lining up. A fix rolled out to Microsoft’s network gateway servers back in July carried a hidden defect, one that only showed up under heavy load on newer machines with more processor cores. Microsoft’s load testing ran mostly on weaker hardware, so the tests never triggered it. The change ran worldwide for more than three months with no trouble.</p><p>Then, about 15 hours before the outage, Teams traffic moved over to the same gateway that serves Exchange Online. That’s a normal resilience step, the kind of failover that’s supposed to keep things running. At the same time, nearly 14% of the gateway’s capacity was out of service because of an unrelated issue from the day before. Each of those on its own would have been fine. Together, during the US business peak, they woke up the July bug.</p><p>Servers locked up. Traffic shifted to healthy servers, which then locked up too, and in less than two hours more than 80% of the gateway fleet stopped responding. Restarting didn’t help, because the servers came back to huge queues of waiting requests and fell over again. Rolling back didn’t help either, because the bug had been in place since summer. Engineers had to patch servers and bring them back gradually. Microsoft also stretched delivery timeouts so delayed messages wouldn’t bounce, which is why a lot of Thursday’s email showed up late on Friday instead of disappearing.</p><p>The full report is due November 20. I wrote a shorter piece on the engineering side of this pattern, about why <a href="https://elegantwolves.com/failover-testing-during-maintenance/">failover testing has to include the maintenance window</a>, and I won’t repeat it here.</p><p>What matters for a business owner is simpler. This wasn’t a careless company or a cheap service. It was one of the most heavily engineered systems in the world, with redundancy built in, and it still went down for most of a working day. If that can happen to Microsoft, the honest planning assumption is that any single service you rely on will have its bad day too. The question is what your business does during that day.</p><h3>The quiet failure is the dangerous one</h3><p>Here’s what I think the pool story shows that the headline numbers don’t.</p><p>When email goes down, you notice the email you were waiting for. You notice you can’t send the proposal. Those are loud failures, and loud failures get handled. People pick up the phone.</p><p>The failures that hurt are the quiet ones. A message that was supposed to arrive and didn’t, where nobody was waiting for it, so nobody noticed it was missing. The pool alarm is the extreme case. Nobody at the hotel was sitting there thinking, “Odd, no chlorine warnings today.”</p><p>Most businesses have a surprising number of these silent messages. Some examples I see over and over:</p><ul><li>A website contact form that emails new leads to one inbox</li><li>Payment failure notices from the card processor</li><li>Alerts from a freezer, a server room, or a security system</li><li>Online booking requests and cancellations</li><li>Low stock or reorder warnings from inventory software</li><li>“Your domain or certificate is about to expire” notices</li></ul><p>None of these feel like alarms when you set them up. They feel like notifications. But each one represents a decision somebody made years ago, usually without much thought, about how your business finds out something went wrong.</p><p>There’s a newer layer too, and it’s growing fast. A lot of small and mid-size businesses now have AI tools and automations reading their email and acting on it. An assistant that sorts incoming requests, a workflow that turns an order email into a job ticket, an agent that drafts replies. Copilot was on the list of affected Microsoft services last week. When the inbox stops, every automation downstream of it stops too, and none of them complain. They simply have nothing to do. If you’ve wired AI into your inbox, your dependency on that inbox got deeper, even if it doesn’t feel that way.</p><h3>How I’d find your pool alarms</h3><p>You don’t need a consultant or a disaster recovery binder for this. You need about an hour and an honest conversation with whoever handles the day-to-day.</p><p><strong>Start with one question.</strong> Sit down and ask, “What tells us when something goes wrong?” Skip the list of tools you pay for and focus on how news of a problem actually reaches a person. Write down every answer, even the boring ones. You’re looking for anything that arrives by email, chat or app notification and that someone is expected to act on.</p><p><strong>Then ask the 24-hour question about each one.</strong> If this message didn’t arrive for a full day, and nobody knew it was missing, what would happen? Last week gives you a real, recent example to point at, so it’s not a hypothetical anymore.</p><p>Most items will land in “annoying but fine.” A newsletter signup can wait a day. A few will land somewhere worse. A lead goes cold. A refrigerated delivery gets left on a dock. A client’s appointment request sits unanswered and they book somewhere else. And once in a while, an item lands in “someone could get hurt” or “we could lose a lot of money.” Those are your pool alarms.</p><p><strong>Give the top tier a second path that doesn’t share the first one’s weak point.</strong> This is where people usually get it wrong. Sending the alert to two email addresses on the same provider isn’t a second path. Sending it to Teams instead of Outlook wasn’t a second path last week either. The backup has to fail for different reasons than the primary.</p><p>Practically, that might mean a text message, an automated phone call, a light or buzzer on site, or a notification from a different provider entirely. For safety equipment, the right answer is often a local alarm that makes noise in the building and doesn’t need the internet at all. For a high-value lead form, it might mean the form also texts the owner’s phone.</p><p><strong>Watch for silence, not only for alarms.</strong> This idea is old in software and underused everywhere else. Instead of only sending a message when something breaks, have the system check in on a schedule when everything is fine. If the check-in doesn’t show up, that absence is the alert. A monitoring service sends you a text when it hasn’t heard from your system in an hour. It catches the email outage, the dead sensor, the unplugged router and the expired account all at once, because it isn’t trusting any of them to speak up.</p><h3>Write the outage card before you need it</h3><p>The second half of last Thursday, for most businesses, was improvisation. People figured out on the spot how to reach customers, who had the vendor’s phone number, and whether to tell clients anything.</p><p>A one-page outage card fixes most of that. Keep it somewhere that doesn’t depend on the systems it covers, which means printed and pinned up, or saved on phones. Mine would cover a handful of things:</p><ul><li>Who decides we’re in “outage mode” and tells everyone</li><li>How staff reach each other if email and chat are both down (a group text usually works)</li><li>How we reach our most important customers, with numbers written down</li><li>Where to check status for our main providers</li><li>What we tell customers, in two or three plain sentences written ahead of time</li><li>Who goes through the backlog when things come back, so nothing gets lost in the flood</li></ul><p>That last item matters more than it sounds. When Microsoft’s delayed mail started arriving on Friday, plenty of people faced an inbox full of day-old messages mixed with new ones. The ones most likely to slip are exactly the ones that needed attention on Thursday.</p><h3>Test it the way Microsoft didn’t</h3><p>The most useful detail in Microsoft’s report, to me, isn’t the bug. It’s the testing gap. Their load tests ran on hardware that didn’t match what they ran in production, so the tests passed and the problem hid for months.</p><p>Small businesses do a version of this all the time. The backup alert gets set up, someone sends one test message, it arrives, and nobody looks at it again for three years. By then the phone number belongs to an employee who left, the text service account lapsed, or the sensor’s battery died.</p><p>So test the real thing, on the real setup, on a schedule. Once a quarter is plenty for most businesses. Trigger the actual alarm and see who gets it. Have someone walk through the outage card without email for an hour. It feels a little silly the first time. It’s much less silly than finding out during the real thing that the backup went to a disconnected number.</p><h3>Where I wouldn’t spend money</h3><p>I want to be fair about the tradeoffs, because the internet’s advice after every outage tends to be “build redundancy for everything,” and that’s bad advice for a 12-person company.</p><p>I wouldn’t move off Microsoft or Google over one bad day. Every major provider has had big outages, and switching costs are real. I also wouldn’t pay for a second full email system as a standby for most businesses. It’s expensive to maintain and people forget how to use it by the time it’s needed.</p><p>I’d spend the effort on the small number of messages that pass the 24-hour test badly. For most businesses that’s somewhere between two and six things. Giving each of them an independent second path and a silence check is usually cheap. A text alerting service, a simple uptime monitor, a buzzer on a piece of equipment. The money goes where the risk actually sits.</p><p>And if you’re adding AI automations to your workflows right now, which a lot of businesses are, ask the same question of each one before it goes live. What happens if this stops working quietly for a day? Who would notice? If the answer is “nobody,” that’s worth fixing while the setup is still fresh in your head.</p><h3>The question worth asking this week</h3><p>When we design software, one of the first things we ask about any alert is what happens when the alert itself doesn’t arrive. It’s an unglamorous question, and it rarely comes up in a sales demo. But last week gave everyone a free rehearsal, and the hotel in Brühl showed what the stakes look like when the answer is “nothing happens.”</p><p>So here’s my small ask. Sometime this week, while Thursday is still fresh, sit down with your team and ask what tells you when something goes wrong. Find the one or two answers that live only in an inbox. Give them a second way to reach a human.</p><p>That’s it. No binder, no big project. Just making sure the important warnings have more than one way to reach someone.</p><p>Thanks for reading. If you want more of these notes on practical software and AI for real businesses, we write regularly at <a href="https://elegantwolves.com/">elegantwolves.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d30959d460d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The ad sitting next to your health question]]></title>
            <link>https://elegant-wolves-labs.medium.com/the-ad-sitting-next-to-your-health-question-fdc6679d08a1?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdc6679d08a1</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[digital-health]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[patient-safety]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 16:14:04 GMT</pubDate>
            <atom:updated>2026-11-12T16:14:04.681Z</atom:updated>
            <content:encoded><![CDATA[<p>A new study found a clearly labeled drug ad shifted which treatment AI models picked, while the answers still looked medically sound. Ads are arriving in chatbots now, so here’s what I’d watch and what I’d build differently.</p><figure><img alt="A man at a wooden table holding a smartphone, with a glass of water and a few pills in front of him." src="https://cdn-images-1.medium.com/max/1024/1*EFl0tGlY6qvvzLiCEi7fcA.jpeg" /></figure><p>Picture a man at his kitchen table late on a weeknight. His doctor mentioned two possible medicines for his blood pressure, and he didn’t catch the difference. The office is closed. So he opens a chatbot on his phone, types out his situation, and asks which one makes more sense for someone like him.</p><p>The answer comes back calm and well organized. The chatbot names one of the two, gives a few sensible reasons, and suggests he confirm with his doctor. Nothing looks wrong.</p><p>Now change one detail. Somewhere in what the model read before answering, there was an ad for one of those two drugs. Clearly labeled. Would the answer have changed?</p><p>A study published this week says yes, more often than you’d hope. I keep thinking about this one because the problem is quiet. Nobody gets an obviously bad answer. They get a reasonable one with a thumb on the scale.</p><h3>What the study found</h3><p>Researchers at the Icahn School of Medicine at Mount Sinai published the work on November 11 in The Lancet Digital Health. Mount Sinai posted a <a href="https://www.mountsinai.org/about/newsroom/2026/drug-ads-can-influence-ai-treatment-recommendations-study-finds">plain-language summary</a>.</p><p>The setup was simple, which is what makes the result convincing. They took physician-written clinical cases and gave them to 12 AI models from the big AI companies. Sometimes the case arrived alone. Sometimes a real, clearly labeled drug advertisement sat right before it. The patient details and the treatment choices stayed the same. Only the ad changed.</p><p>Across 33 scenarios and more than 258,000 responses, the advertised drug got picked 47.6% of the time when the ad was present, compared with 35% without. In four cases, a model went from never choosing the advertised drug to choosing it every single time once the ad showed up.</p><p>Two more findings stuck with me. In a subset of free-text answers, the models mentioned the ad only 29.9% of the time, while they repeated claims from the ad more often than that. So the ad shaped the reasoning without the answer saying so.</p><p>The other finding is the reassuring one, and I want to be fair here. The ads did not clearly push models toward unsupported supplements or clearly worse drugs. The shift happened among reasonable options. The researchers also didn’t study whether any of this changed real prescriptions, costs, or health outcomes. An earlier preprint version of the paper also reported the size of the effect varied a lot from one model maker to another.</p><p>The first author, Dr. Mahmud Omar, put the core worry well. “What I find quite interesting, and even surprising, was how reasonable the answers often appeared,” he said. The model’s preference among several good options moved, and because the final answer still looked medically sound, the influence “could be easy to miss.”</p><h3>Why accuracy checks miss this</h3><p>Most of the ways we test AI answers ask one question. Is the answer right?</p><p>This study shows why that question isn’t enough. When two treatments are both appropriate, either answer passes. A reviewer grading for correctness gives full marks both times. The bias lives in the choice between good options, and a correctness check doesn’t look there.</p><p>The senior author, Dr. Girish Nadkarni, said, “An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer.”</p><p>That line applies far outside medicine. Plenty of everyday questions have more than one acceptable answer. Which brand of water heater. Which accounting tool. Which urgent care nearby. Which supplier to call. A model nudged toward one of several fine options will sail through every quality check you throw at the answer, because none of the options are wrong.</p><p>So the useful test is a comparison, not a grade. Run the same question with and without the commercial content and see whether the answer moves. That’s one of the steps the researchers themselves suggest for AI developers and health systems.</p><h3>The timing matters</h3><p>If chatbots were ad-free, this would be an interesting lab result and not much more. They’re not.</p><p>WIRED reported on November 5 more than 100,000 customers have advertised in ChatGPT since ads debuted there in February, and Google had recently expanded ads in its AI Mode to 10 more countries. One company tracking ChatGPT answers every day estimated about one in seven answers in its sample carried a sponsored ad on November 10. The tracker says its numbers are indicative, not a full census, so I’d treat that as a rough picture.</p><p>Meanwhile, a lot of people bring health questions to these tools. A KFF poll from earlier this year found about a third of U.S. adults had used AI chatbots for health information or advice in the past year. A later KFF poll found 29% now use them for health information monthly, up from roughly 17% in June 2024.</p><p>I want to be careful about one thing here, because it’s easy to blur. In the study, the ad was placed inside what the model read before answering. In consumer chatbots today, ads show up as sponsored placements near or within the answer, and the study didn’t test any live product. I’m not claiming any chatbot feeds ads into its model.</p><p>But the WIRED piece describes something worth noticing. One ad network said it generally receives a user’s entire conversation up to that point to choose which ad to show. So the conversation flows into the ad system. The study tells us what happens when the flow goes the other way, when promotional text ends up in front of the model. And there are more ordinary ways for this to happen than you’d think.</p><h3>Where commercial text sneaks in</h3><p>Builders should pay attention here, not only the big chatbot companies.</p><p>Models read a lot of material before they answer. Web pages pulled in by a search step. Documents in a knowledge base. Product listings from a partner feed. Emails a user forwards. A page someone pastes in to ask, “Is this legit?”</p><p>Plenty of that material is promotional. A manufacturer’s brochure is an ad with nicer fonts. A vendor’s comparison page ranks the vendor first. A sponsored listing looks a lot like a regular one once turned into plain text.</p><p>Here’s a scenario I find realistic. A small clinic sets up a patient-facing assistant to answer common questions after hours. To give it something to work from, someone loads the document library with patient handouts. Some came from the clinic’s own staff. Some came from drug and device reps, because those handouts are free and well designed. Nobody thinks of them as ads. They’re “the materials.”</p><p>Six months later, the assistant is answering “which of these two options is better for me” questions. Based on what this study found, you’d want to know whether the rep materials are tilting those answers. And you won’t find out by grading answers for accuracy, because the answers will look fine.</p><p>Same pattern outside healthcare. A home services company builds a quote assistant fed by supplier catalogs. A retailer’s shopping helper reads product pages written by the brands. An internal tool for a property manager pulls in contractor websites. Each one mixes neutral information with commercial content, and the model treats all of this as context.</p><h3>What we’d do if we were building this</h3><p>At EWL we build software and AI tools for real operations, and we lean toward boring, checkable setups. Here’s how I’d approach a feature like that clinic assistant.</p><p>First, I’d sort the sources before the model ever sees them. Some content is advice-grade, written or approved by people accountable to the patient or customer. Some content is commercial. The two shouldn’t share a bucket. For anything resembling a recommendation, the assistant should draw from the advice-grade pile only, and commercial material should be either excluded or clearly marked as coming from a manufacturer when it’s shown to the person.</p><p>Second, I’d run paired tests. Take a set of real questions where more than one answer is acceptable. Run them with the commercial content available, then without. Count how often the answer changes and in which direction. If one brand keeps winning only when its brochure is in the library, you’ve learned something no accuracy score would show you. We wrote last month about <a href="https://elegantwolves.com/ai-model-benchmarks-own-tests/">testing models on your own work instead of trusting leaderboards</a>, and this is the same instinct applied to influence.</p><p>Third, I’d keep a record of what the model saw. When an answer looks a little off, the first question is always “what was in front of it?” If you log which documents went into each answer, a person reviews a sample every week and spot patterns. Without that record, you’re guessing.</p><p>Fourth, I’d teach the assistant to show its options. For questions with several reasonable answers, the honest reply usually names them and explains the tradeoffs, then points to the person who should decide. A doctor, a pharmacist, an estimator, a manager. That’s slower than one confident pick. It’s also closer to the truth.</p><p>None of this is exotic. This is the same discipline you’d apply to a human employee who reads vendor pitches all day. You’d want them to know which papers were sales material, and you’d want a manager checking their recommendations now and then.</p><h3>The tradeoffs are real</h3><p>I don’t think the answer is “ads are evil, ban them.” Running these models costs a lot of money, and ads are one way to pay for free access. A lot of the people asking health questions at midnight are the ones who can’t easily get an appointment, and a free tool helps them.</p><p>Disclosure isn’t a full fix either. The ads in this study were clearly labeled, and the models were still influenced. Labeling helps the human reader. It doesn’t seem to protect the model’s reasoning by itself. So “we label our ads” is a good start for a chatbot company, and not the end of the conversation.</p><p>There’s also a cost to stripping out all commercial content. Manufacturer materials often contain accurate, useful details, like dosing instructions or warranty terms. Throwing them away makes the tool less helpful. That’s why I’d rather separate and test than delete everything. You keep the useful facts and watch for the tilt.</p><p>For builders, the honest tradeoff is a bit more work up front. Sorting sources, writing paired tests, and keeping logs takes time a demo doesn’t need. But a demo doesn’t run for two years with a growing document library and nobody checking what got added.</p><h3>For the rest of us, as users</h3><p>If you’re the person at the kitchen table, a few habits help. Ask the question again in a fresh conversation and see if the answer holds. Ask directly what the other reasonable options are and why someone might choose each. Notice when an answer leans hard on one brand without explaining the alternatives. And for anything about your health, bring the answer to your doctor or pharmacist as a starting point for the conversation, not the decision.</p><p>That last one isn’t new advice. The study gives this habit a sharper reason.</p><h3>A small test worth running</h3><p>If you run a practice, an agency, or any business with an AI assistant reading outside material, here’s one thing to try this month. Pick ten questions where more than one answer is fine. Run them with your full document library, then again with the vendor and manufacturer content removed. Write down what changed.</p><p>You might find nothing. That would be a good afternoon. Or you might find your assistant has a favorite brand you never chose.</p><p>Thanks for reading. If you want more on practical AI with people kept in the loop, we write about it regularly at <a href="https://elegantwolves.com/">elegantwolves.com</a>. Take care, and I’ll see you here next week.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdc6679d08a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your software bill is turning into a meter]]></title>
            <link>https://elegant-wolves-labs.medium.com/your-software-bill-is-turning-into-a-meter-496f407ee3b1?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/496f407ee3b1</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[saas]]></category>
            <category><![CDATA[pricing]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 16:10:36 GMT</pubDate>
            <atom:updated>2026-11-10T16:40:23.957Z</atom:updated>
            <content:encoded><![CDATA[<h4>In the last two weeks, a handful of tools small businesses use every day switched their AI features to credits. Here’s what changed, where the surprise bills come from, and the hour of setup I’d do this month.</h4><figure><img alt="A woman reviewing a document at home with a phone and laptop nearby" src="https://cdn-images-1.medium.com/max/1024/1*JEYCADyzMPtt-Xdn3qFZbg.jpeg" /></figure><p>I spent a good part of last week reading pricing pages and help docs, which nobody does for fun. I went in expecting price increases. What I found was quieter than that, and I think more important for anyone who runs a small business.</p><p>Almost nobody raised the price of a seat. The monthly number you’d see on your card statement stayed the same. What changed was what that number buys. More and more of the AI work inside everyday tools now runs on credits, with a monthly allowance, a meter, and a setting somewhere that decides what happens when the meter hits zero.</p><p>If you’ve ever had a phone plan with a data cap, you already understand the model. The difference is that most owners don’t know which of their tools now work this way, who in the company holds the switch, or what the switch is set to.</p><p>So this week I want to walk through what moved, why vendors are doing it, where the surprises come from, and what we’d do about it if we were sitting in your chair.</p><h3>What moved in the last two weeks</h3><p>Notion is the cleanest example because the change landed on a single date. On November 2, existing Business and Enterprise workspaces moved to a monthly AI credit system. Each member now gets 1,500 Standard credits, used when Notion picks the model for you, and 500 Select credits, used when you pick a model yourself or run a custom agent you set up. The two pools are separate. They don’t roll over, and you can’t move credits from one person to another. According to Notion’s own help page, once a member uses up a pool, the work that depends on it pauses until the monthly reset. A workspace owner can turn on a setting that lets people keep going on purchased workspace credits, and that setting is off by default.</p><p>One detail is worth circling on your calendar. Notion says AI meeting notes will start drawing from the Standard allowance on December 1, at roughly 10 credits for a one hour meeting. If your team records a lot of meetings, that’s a new draw on the same pool people use for writing and search.</p><p>Microsoft made a similar move for smaller companies. As CRN reported, new Microsoft 365 Copilot Business licenses bought through Microsoft’s partner program now come with usage-based billing on by default, starting November 2. Everyday tasks, like summarizing a document or turning a meeting transcript into notes, stay covered by the per-user license. Heavier agent work runs on Copilot Credits at one cent each. Microsoft says nothing gets billed until an administrator creates a spending policy, and admins can set budgets by company, group or user.</p><p>Then there’s the long tail. PricingSaaS, which tracks changes to software pricing pages, logged four vendors adding paid AI credit meters in the week to November 2 without touching their seat prices. Two of them sell to the kind of businesses we talk to every week. ServiceM8, a job management app for trades, added a $39 Office Agent Pro add-on. Shopmonkey, which runs auto repair shops, added a $134 Concierge add-on. Sprout Social added a $28 per seat tier that comes with 1,000 AI credits a month.</p><p>None of this is a scandal. It’s a pattern. And once you see it, you start finding it everywhere.</p><h3>Why vendors are doing this</h3><p>I want to be fair to the people selling software, partly because we build software too.</p><p>A flat seat price worked when every user cost the vendor about the same to serve. AI broke that. Summarizing a two-paragraph email costs a vendor almost nothing. Letting an agent read fifty documents, call three other systems, retry twice, and write a report costs a lot more, and the gap between those two jobs keeps growing as agents get more capable. Microsoft said this out loud in its announcement. Its credits are charged based on the model used, how long the task runs, how much context the agent needs, and how many actions it takes.</p><p>So vendors are splitting the bill. Light, predictable use stays in the seat. Heavy, unpredictable use gets metered. From their side, that’s the only way to keep offering the expensive features without raising the price on everyone.</p><p>From your side, it means your software costs now behave like a utility for part of the month and like a subscription for the rest. That’s fine. You just have to manage the utility part like a utility.</p><h3>Where the surprise comes from</h3><p>Pax8, a cloud marketplace that sells to small businesses through IT providers, surveyed 1,150 business leaders at companies with 5 to 499 employees for its Q3 report in October. Among those that paid for AI in the past year, 49% got at least one bill or usage charge higher than they expected. In the U.S. the figure was 57%, and a quarter of U.S. respondents got more than one. Pax8 has an interest in this topic, so read the numbers with that in mind, but they line up with everything else published in the past few weeks.</p><p>In my experience, the surprise rarely comes from the price itself. It comes from four other places.</p><p>The first is defaults. Somebody set up the tool a year ago, clicked through the onboarding, and moved on. Now the tool has a setting nobody has looked at, and its default decides whether you get a pause or a bill. Notion’s default is to pause. Microsoft’s is to bill nothing until someone writes a policy. Other tools let top-ups run on a saved card. You won’t know which you have until you look.</p><p>The second is that the heavy users aren’t who you’d guess. Chris Regan, a Microsoft partner in New Jersey, told CRN that some of the people hitting their quotas are users “we would have never expected in a million years.” He described people who are a week into the month and already asking for more. That rings true to me. The person who leans on AI hardest is often the office manager or the coordinator who found one workflow that saves her an hour a day, not the person who talks about AI in meetings.</p><p>The third is work that runs when nobody’s watching. A summary you ask for costs what it costs, once. An agent that runs every time a form comes in, or a meeting ends, or a file changes, keeps spending while you sleep. Pax8 found usage-based charges were the hardest for small businesses to forecast, and they grow as companies build more agents and workflows. A writer at ThinkFast who looks at monthly closes put it plainly. “A seat license shows up the same every month. A usage charge shows up when someone’s workflow ran all weekend.”</p><p>The fourth is overlap. Pax8 found U.S. small businesses paying for AI through usage pricing, flat subscriptions, and per-seat pricing in roughly equal shares, often all three at once. So the same task, say drafting a customer reply, might be covered by your email tool, your CRM, your help desk, and a chatbot subscription somebody expensed. You’re paying four meters to do one job.</p><h3>A pause or a bill, pick on purpose</h3><p>The most useful thing I took from reading all these help pages is that every metered tool forces you to answer one question, whether you know you’re answering it or not. When we run out, do we stop, or do we keep paying?</p><p>Neither answer is right for every tool.</p><p>Think about a small home services company with a dispatcher who uses an AI assistant to draft job summaries and reply to customers. If that assistant pauses on the 20th of the month, the dispatcher writes the replies herself for ten days. Annoying, but nobody gets hurt and nothing breaks. A pause is the right default there, and the cap is a useful signal that maybe you need a bigger plan.</p><p>Now say the same company has an automation that reads incoming service requests and sorts them into the schedule overnight. If that pauses, requests sit unread until someone notices on Monday. Customers wait. For that workflow, you probably want it to keep running past the allowance, with a hard ceiling on what it can spend and an alert when it crosses halfway.</p><p>The point is to make the decision per tool, deliberately, and write it down. Most businesses never make it. They inherit whatever the vendor chose.</p><h3>What we’d do this month</h3><p>None of this needs a consultant or a new piece of software. It needs about an hour, a recent card statement, and someone with admin access to your tools. Here’s how I’d spend that hour.</p><p>Start with the statement, not your memory. Pull the last two or three months of software charges and write down every tool, what it costs, and when it renews. The tool you forgot about is the one most likely to have changed its terms.</p><p>Next to each tool, mark how its AI features bill. Flat inside the seat, credits that reset monthly, or metered by use. You’ll find this on the vendor’s pricing page or in a help article usually titled something like “AI credits” or “usage limits.” If you can’t tell, that’s a finding too.</p><p>For every tool with credits or a meter, find the setting that decides what happens at zero. Decide whether it should pause or continue, using the question above. If it continues, set a hard monthly cap, not just an email alert. An alert tells you the money is gone. A cap stops it from going.</p><p>Give each meter an owner. One person whose name is next to the tool, who gets the alerts and who decides when to raise a limit. In a five-person company that might be you for everything, and that’s fine. The failure mode is when the alerts go to an inbox nobody reads.</p><p>Check usage halfway through the month for the next two months. You’ll learn more from two mid-month checks than from any forecast. You’ll see who uses what, which features burn credits fastest, and whether your allowance is a comfortable fit or a cliff.</p><p>Then do the harder thing, which is to judge each AI feature by what one finished piece of work costs you, not by its monthly price. That includes your people’s time. Pax8 found 55% of small business decision-makers spend three or more hours a week reviewing, editing or fact-checking AI output. If a cheaper model needs three tries and twenty minutes of cleanup, it isn’t cheaper. If a pricier one gets a job right the first time and saves your coordinator an hour, the credits are a bargain.</p><p>Last, look for the duplicates. If three tools all offer to draft emails, pick one and turn the others off. Fewer meters means fewer surprises.</p><h3>The tradeoffs I’d be honest about</h3><p>There’s a version of this advice that turns into cost-cutting for its own sake, and I’d push back on that. If your team found AI features that save real time, the worst outcome is clamping down so hard that people stop using them, or worse, start paying for their own tools on a personal card where you can’t see what data goes where. Pax8 found 35% of AI-engaged businesses discovered tools, workflows or agents running without approval in the past six months. Tight caps with no way to ask for more push people in that direction.</p><p>So I’d set the caps generously at first, watch for a month, and tighten only where usage doesn’t line up with value. Make it easy to ask for more. Microsoft’s admin settings let a user who hits a limit see a warning, request more, or see nothing at all. Pick the option that keeps the conversation open.</p><p>There’s also a real argument for consolidation. A single tool with a bigger allowance is often easier to manage than five tools with small ones, even if the sticker price is higher. Fewer places to check, fewer defaults to get wrong.</p><p>And some of this will change again soon. Pricing is moving fast. PricingSaaS logged one vendor, Funnelstory, switching its enterprise agent billing from consumption to outcome-based pricing just this week, where you pay per result instead of per unit of compute. Pax8 found 69% of small businesses would be more likely to expand AI if they could pay for results at a similar cost. I wouldn’t plan around that yet, but I’d expect more vendors to try it.</p><h3>A note for anyone building software</h3><p>We build software at Elegant Wolves Labs, so I read all of this from both sides of the counter.</p><p>If you’re putting a meter on your product, the meter is part of the product. People don’t mind paying for what they use. They mind finding out after the fact. The vendors handling this well show users what used their credits, warn them before a pause instead of after, explain in a sentence why a task cost what it did, and make the “what happens at zero” setting easy to find and easy to change. The ones handling it badly bury all of that three menus deep and let the first bill do the explaining.</p><p>That’s the same standard we try to hold for anything that acts on someone’s behalf. Show what happened. Say why. Give people a way to preview or undo. A credit system that does those things builds trust. One that doesn’t becomes the reason a customer starts shopping around.</p><h3>Before your next bill</h3><p>If you only do one thing after reading this, open the admin settings of the tool your team uses most and find out what happens when the AI allowance runs out. It’ll take ten minutes, and you’ll know more about your software costs than most companies your size.</p><p>The meters aren’t going away. Handled with a little attention, they’re a fair trade. Ignored, they’re the line on next month’s statement nobody can explain.</p><p>Thanks for reading. We write about practical AI, calm software, and the ordinary work of running a business with both at <a href="https://elegantwolves.com/">elegantwolves.com</a>. And if a surprise AI charge has already caught you, I’d like to hear what tripped you up. I’ll be back here on Thursday.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=496f407ee3b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Writing the code isn’t the scarce skill anymore]]></title>
            <link>https://elegant-wolves-labs.medium.com/writing-the-code-isnt-the-scarce-skill-anymore-c97ef9c48fc1?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/c97ef9c48fc1</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 16:29:37 GMT</pubDate>
            <atom:updated>2026-11-10T16:38:58.806Z</atom:updated>
            <content:encoded><![CDATA[<h3>Agents already produce a large share of work code. The hard part is directing them, reviewing what comes back, and deciding what ships.</h3><figure><img alt="Three colleagues reviewing work together on a laptop" src="https://cdn-images-1.medium.com/max/1024/1*hrGlyQSuNfJ0rvDN25pBMw.jpeg" /></figure><p>Last month I watched a senior engineer open three agent sessions before lunch. One was drafting a migration. One was writing tests for a bug she already understood. The third was summarizing an old pull request so she could explain a design choice to a new hire. She typed almost no code herself. She spent the morning reading diffs, rejecting half of them, rewriting one prompt, and merging two clean changes.</p><p>By noon she looked tired in a different way than she used to. Not from fighting a compiler. From deciding, over and over, what was good enough to keep.</p><p>I keep running into that scene. And this week a pile of fresh survey data made it hard to treat as a personal quirk. For a lot of developers, writing code by hand is no longer the center of the job. The scarce work is directing agents, reviewing their output, and deciding what ships.</p><h3>Half the code is already coming from agents</h3><p><a href="https://blog.jetbrains.com/research/2026/08/how-much-code-do-developers-really-let-agents-write/">JetBrains’ Developer Ecosystem Survey 2026</a>, run from May through July with more than 15,000 professional developers, asked a blunt question. What share of the code you produced last month for work was fully agent-generated, written with some AI help, or written fully by you?</p><p>On average, people reported roughly 47% fully generated by agents, about 38% written with AI assistance, and about 27% written fully by hand. JetBrains notes those averages can sum past 100% because of how the answer buckets work, so I am reporting the numbers the way they published them rather than forcing them to add up neatly.</p><p>The distribution is what sticks with me. Over half of developers now write less than 20% of their code manually. About one in five write none of it without AI. Around 22% generate more than 80% with agents. Senior developers are among the first to hand coding over. About a quarter of seniors put more than 80% of their work code through agents.</p><p>A companion <a href="https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/">JetBrains post on adoption</a> fills in the tools side. Ninety percent of professional developers use AI coding agents at work at least weekly. Sixty-eight percent use them daily. Claude Code sits around 39% adoption at work. GitHub Copilot has slipped to about 21%.</p><p><a href="https://survey.stackoverflow.co/2026/ai">Stack Overflow’s 2026 Developer Survey</a> tells a similar story from another angle. Coding assistants and agents are the dominant AI use case at 66%, a hair ahead of general chatbots at 63%. Among people who use these tools, 73% do so daily, and nearly a third of daily users spend four or more hours with them.</p><p>This is not a hobby layer on top of the old job. For many teams, it is the job.</p><h3>Review is where the work moved</h3><p>If agents write a large share of the code, somebody still has to decide whether it belongs in the product. That somebody is usually a person who used to spend most of the day writing.</p><p>An <a href="https://www.adaptavist.com/blog/what-ai-means-for-the-future-of-software-engineering">Adaptavist report published today</a>, based on a survey of 1,000 people in software roles across the UK, US, Germany, Spain, and the Netherlands, puts sharp numbers on the shift. Eighty-three percent say AI made them more productive. Eighty-one percent say it made them better at their jobs. At the same time, many find the work less enjoyable and more stressful. Sixty percent say they are being asked to do more with fewer resources.</p><p>Two findings feel especially useful if you run a small engineering team.</p><p>Fifty-two percent admit they use AI more often than they need to, because their organization treats AI usage as a performance signal. And 82% regularly deploy multiple AI agents for a single task, while 68% say managing agents is becoming an explicit part of their daily role.</p><p>So the productivity story is real, and so is the pressure story. People are faster. They are also spending more of the day supervising tools, reviewing output, and proving they used the tools enough. That last part is a quiet tax. Measuring “how much AI did you use” is not the same as measuring whether the software got better.</p><p>Scott Hanselman, VP of Developer Community at Microsoft, put the trust question in a line I keep quoting from the Stack Overflow survey. He talks about treating AI as an exoskeleton, like a Tony Stark augmentation, instead of an empty Ultron with nobody inside. The Stack Overflow numbers back the caution. Forty-eight percent trust AI when they can easily validate the answers. Only 16% extend that trust to most tasks that are not important work decisions. Human-in-the-loop is still the default mode, not a temporary phase.</p><p>For a builder, the practical translation is simple. The agent drafts. A person still owns the merge button. If nobody owns the merge button, you do not have a team with agents. You have a team with a lot of plausible-looking code and no one accountable for it.</p><h3>The junior learning problem is not abstract</h3><p>There is a cost hiding under the speed.</p><p>In the Adaptavist survey, 63% of developers say they learn less through hands-on coding because of AI. Forty-three percent say they have fewer chances to upskill. Seventy-one percent of experienced developers and 68% of early-career developers worry juniors over-rely on automation.</p><p>I believe them. The old path into engineering was messy on purpose. You wrote the wrong loop. You chased a null for an afternoon. You read someone else’s ugly function and slowly learned why it was ugly. Agents short-circuit a lot of that friction. The first draft arrives looking finished. A junior who accepts it too fast never builds the muscle that would have caught the subtle bug next month.</p><p>That does not mean juniors should avoid agents. It means a team that cares about next year’s seniors has to protect some hands-on work on purpose. Leave a few problems that a person is expected to solve without an agent first. Ask for a short note on what the agent changed and why those changes are safe. Pair a junior with a senior on reviews of agent-heavy pull requests, not only on greenfield features.</p><p>If every ticket becomes “prompt, accept, merge,” you get throughput this quarter and a thinner bench the year after.</p><h3>Treat AI-assisted pull requests as evidence</h3><p>Simon Willison wrote something useful on November 2 in his <a href="https://simonwillison.net/2026/nov/2/llm/">llm 0.37 release notes</a>. Open source maintainers complain about floods of AI-assisted pull requests. His project has seen plenty of them. He is not finding them as painful as the complaints suggest, partly because he uses models to help with triage.</p><p>He frames those pull requests as bug reports with attached examples. Sometimes he lands them. Sometimes he writes his own fix. The release included work from twelve authors, almost all AI-assisted, and the review burden was lower than he expected.</p><p>I like that frame because it lowers the temperature without lowering the standard. An AI-assisted pull request is not a finished patch you are obligated to accept. It is evidence. Evidence of a bug. Evidence of a possible approach. Evidence that someone spent tokens looking. Your job is to decide what that evidence is worth.</p><p>The same habit helps inside a company. When a teammate opens a pull request that an agent drafted, ask what outcome they checked. Did the tests pass on a clean machine? Did the migration run against a copy of production data? Did the UI still match the design the customer approved? Chat summaries are not proof. The system of record is proof.</p><p>We wrote about a related idea on our blog in <a href="https://elegantwolves.com/ai-agent-changes-check-the-record/">verify AI agent changes against the system of record</a>. An agent can report success while a trigger or side effect changes far more than the chat claims. The lesson travels well to code review. Trust the diff, the tests, the logs, and the data. Do not trust the agent’s confident paragraph about what it did.</p><p>Sometimes the useful move is to turn a failed agent attempt into a regression check, so the next run cannot repeat the same mistake quietly. That is ordinary engineering discipline with a new source of failures.</p><h3>What a small studio should change this month</h3><p>Elegant Wolves Labs is a small custom software and AI studio. We build practical systems with human oversight, not demos that look clever in a slide deck. Here is what I would change first if a five-to-fifteen person engineering team asked me what to do with all of the above.</p><p>Name a reviewer for agent-heavy work. Not a committee. A person. When a pull request is mostly agent-generated, that person knows they are responsible for the judgment call, not only for a polite thumbs-up.</p><p>Stop measuring AI usage as a performance score. Measure outcomes. Did the change reduce support tickets? Did the feature ship without a rollback? Did the on-call week get quieter? If half your team is using agents more than they need to because someone is counting prompts, you taught them the wrong lesson.</p><p>Protect a slice of hands-on work for juniors and for seniors who are mentoring them. Pick a weekly window or a class of tickets where the first attempt is human. Agents can still help after the first honest struggle. The struggle is part of the training.</p><p>Make outcome checks the default close for an agent task. Before merge, compare the claimed change with the record the software keeps. Database rows. Log lines. Screenshots. Contract tests. Whatever matches the risk. A chat window saying “done” is a claim, not a receipt.</p><p>Write down which decisions still need a person. Billing rules. Permission changes. Anything that deletes or rewrites customer data. Anything a regulator or a customer would ask about later. Agents can draft the code. A named human still decides whether it goes out.</p><p>None of that requires a new platform. It requires clearer ownership and a little less romance about speed.</p><h3>The tradeoff you should name out loud</h3><p>There is a real tradeoff here, and pretending otherwise helps nobody.</p><p>If you slow down to review carefully, you ship fewer agent drafts this week. If you protect hands-on learning time, some tickets take longer. If you refuse to treat AI usage as a metric, a vendor slide or an internal dashboard will look less impressive.</p><p>I would take those costs. A team that merges fast and cannot explain its own code is not productive. It is accumulating debt with better grammar. A team that reviews well, teaches juniors on purpose, and checks outcomes against the record will still use agents heavily. They will also sleep better when something breaks on a Friday.</p><p>The surveys this fall are not saying “stop using agents.” They are saying the job changed shape. Productivity went up. Enjoyment and learning are under pressure. Managing multiple agents is becoming part of the daily role. Trust stays conditional, tied to whether a person can validate the answer.</p><p>That is a workable future if you staff for it. Hire and promote for judgment, communication, and taste in diffs, not only for raw typing speed. Give seniors time to review. Give juniors time to struggle. Give every agent task a person who owns the result.</p><h3>Pick one habit</h3><p>I still like writing code. I also spend more of my week reading code agents wrote, asking better questions of them, and throwing out work that looked fine until I checked the record. The craft did not disappear. It moved.</p><p>If you lead a small product team or a studio like ours, pick one habit this week. Name the reviewer for your next agent-heavy pull request. Or pick one junior ticket that stays hands-on on purpose. Or stop counting prompts and start counting outcomes.</p><p>Small moves beat big speeches.</p><p>Thanks for reading. We write about practical AI, calm software, and the ordinary work of building systems people can trust at <a href="https://elegantwolves.com/">elegantwolves.com</a>. I will be back here on Tuesday.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c97ef9c48fc1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Saying yes to an AI tool is the easy part]]></title>
            <link>https://elegant-wolves-labs.medium.com/saying-yes-to-an-ai-tool-is-the-easy-part-029ab43af7a9?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/029ab43af7a9</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[operations]]></category>
            <category><![CDATA[health-technology]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 16:13:52 GMT</pubDate>
            <atom:updated>2026-11-10T16:37:37.309Z</atom:updated>
            <content:encoded><![CDATA[<h4>Health systems got good at approving AI. The harder work starts the month after, and the lessons travel well beyond hospitals.</h4><figure><img alt="doctor and a nurse reviewing a laptop together in a medical office" src="https://cdn-images-1.medium.com/max/1024/1*QJIYcSgl-oOZhOBdZoDLOQ.jpeg" /></figure><p>A few weeks ago I heard a line from a health system leader that I haven’t been able to shake. Speaking on a webinar about AI governance, a panelist from Hartford HealthCare said hospitals tend to test AI like a new drug and then monitor it like a fax machine.</p><p>I laughed, and then I winced, because I’ve seen the same pattern in plenty of places that have nothing to do with medicine. A team spends weeks deciding whether to adopt a tool. There are demos, a security review, a pilot, a meeting where everyone signs off. Then the tool goes live, and the attention moves on to the next decision. Six months later, nobody remembers who was supposed to be watching.</p><p>This week I want to stay with that gap. Healthcare is where it shows up most clearly right now, so most of my examples come from there. But if you run a business that buys or builds software with a model inside, I think you’ll recognize your own team in this.</p><h3>The tool you approved isn’t the tool you’re running</h3><p>When you approve an AI tool, you’re approving a snapshot. You saw how it behaved on a certain day, with a certain model behind it, on a certain set of examples.</p><p>Then the world keeps moving. The vendor ships updates. The underlying model gets swapped for a newer one, sometimes without much notice. Your own data shifts, too. Patients change, customers change, and the way your staff writes notes or enters orders changes a little every quarter.</p><p>Hartford HealthCare described an example I keep coming back to. A routine vendor upgrade quietly put a large language model behind software the hospital had approved years earlier as a simple, predictable application. Nobody filed a request for a new AI tool. It arrived inside an update to an old one.</p><p>A JAMA Health Forum essay from mid-October gave this situation a name I like, an accountability inversion. The vendor controls how the model is designed and updated. The hospital carries the responsibility for safety. And the hospital often has no view into training data or version history. So the party holding the risk has the least visibility into what changed.</p><p>You don’t need to work in a hospital to feel that. If your scheduling app, your CRM or your accounting software added an AI feature this year, ask yourself a plain question. Would you know if the model behind it changed next month?</p><h3>How drift shows up</h3><p>When people talk about AI monitoring, it sounds like dashboards and alerts. In practice, the first warning usually looks a lot more human.</p><p>A Healthcare Technology News piece from late October, summarizing reporting from Becker’s, put it well. Health systems have become disciplined about vetting AI before go-live. After go-live, ownership gets fuzzy. Contracts often don’t say who watches for degradation, how often, or who acts when something breaks. So the first sign of trouble tends to be a frontline clinician complaining, not an alert firing.</p><p>Picture an ordinary Tuesday. A nurse notices the visit summaries feel a little off. Not wrong enough to stop work, only off. She mentions it to a colleague, or she starts fixing them by hand and stops trusting the tool. Weeks pass before anyone with authority hears about it, and by then the workaround has become part of the job.</p><p>The same thing happens in a small business. A front-desk employee starts double-checking every AI-drafted reply before sending it, because two of them went out with the wrong hours last month. Nobody logged those two. Nobody measured the extra ten minutes a day. The tool is technically still “working,” and the business is quietly paying for it twice.</p><p>Drift rarely announces itself. It shows up as small friction that people absorb because they’re busy and the work has to get done.</p><h3>What one health system does about it</h3><p>The most useful example I’ve read this fall comes from Inova. According to Becker’s, the health system has 92 AI features running in production and is preparing to oversee as many as 500. That’s a lot of moving parts. So they review performance on a schedule.</p><p>In their latest quarterly review, three tools had slipping performance scores. All three now have decommissioning plans.</p><p>I find that detail more reassuring than any vendor claim I’ve seen this year. Someone looked. Someone had a number to compare against. And when the number dropped, they made a decision instead of hoping the next update would fix it. Retiring a tool is a normal, healthy outcome of watching it. A team that never retires anything usually isn’t looking hard.</p><p>Hartford HealthCare takes a similar line from a different angle. When a vendor changes the model behind a tool, the tool goes back through governance intake, the same way a new tool would. And they measure performance against limits written into their own policy, not the vendor’s own safety claims.</p><p>That last bit matters. A vendor’s test results tell you how the tool performs on the vendor’s examples. Your limits tell you what’s acceptable in your building, with your patients or customers, given your work. Those are different questions, and only you can answer the second one.</p><h3>Checking facts and checking judgment are different jobs</h3><p>There’s a second thread here that I think gets lost when people say “monitor your AI.”</p><p>Some outputs have one right answer. A dose, a unit, a date, a phone number, a price. Other outputs are judgment calls. Which tests a patient needs, how to word a delicate message, whether a customer complaint deserves a refund.</p><p>Late last month Doximity published a study on dosing that shows why this split matters. They tested nine language models released in 2026 on more than 8,000 dosing questions modeled on real physician queries. At least 2.3% of the doses came back wrong, and a portion of those errors were serious. The most common miss was failing to adjust a dose for things like age or kidney function, even when the question included that information. Newer drugs had much higher odds of an error, which makes sense when you think about how these models learn.</p><p>What struck me was their fix. They didn’t ask a second model to grade the first one. They pull each dosing claim out of the answer and compare it against a structured drug database built on FDA labeling. The same claim gets the same result every time, and the clinician sees where the answer came from. I wrote about this in more detail on our blog in <a href="https://elegantwolves.com/ai-dosing-errors-rule-checks/">a post on AI dosing errors and plain rule checks</a>.</p><p>For a business owner, the takeaway is practical. If an AI tool produces facts with one right answer, check those facts against a source with simple rules. Prices against your price list. Hours against your actual hours. Appointment times against the calendar. That kind of check is boring, cheap and deterministic, and boring is what you want there.</p><p>For judgment calls, monitoring looks different. You’re looking for patterns over time. Are people overriding the suggestion more often than last quarter? Are customers pushing back more? Is the tool drifting toward a certain kind of answer, like suggesting more tests or offering bigger discounts? Those signals come from people and from simple counts, not from a rule engine.</p><h3>What I’d put in place before go-live</h3><p>If I were rolling out an AI feature tomorrow, in a clinic or a lawn care company or a small accounting firm, here’s what I’d want settled before launch day. None of it is fancy.</p><p>First, one named owner per tool after launch. Not a committee, a person. And that person gets actual time on the calendar to look at how the tool is doing. An owner with no hours set aside is a name on a slide.</p><p>Second, a written floor. Before go-live, decide what “good enough” looks like in numbers you can track. Say the AI-drafted replies should need a human edit less than one time in five, or the intake summaries should to match the source record on every medication and allergy. Pick something simple you’ll measure, and write down what happens when the tool falls below it.</p><p>Third, a rollback you’ve tried. Know how to turn the feature off and go back to the old way without a scramble. If you’ve never tested it, you don’t have a rollback, you have a hope.</p><p>Fourth, contract language about model changes. Ask your vendor to tell you when they swap the model behind a feature or add AI to something that didn’t have it. Ask before signing, because that’s when you have the most say. Few vendors volunteer this kind of notice on their own.</p><p>Fifth, a short path from the front line to the owner. The nurse who notices summaries feel off, or the receptionist who catches the wrong hours, needs an easy way to say so, and a reason to believe someone will read it. A shared channel, a form with three fields, a standing question in the weekly huddle. Whatever fits your team.</p><p>That’s the whole list. Five things, and none of them requires an AI expert on staff.</p><h3>The tradeoffs are real</h3><p>I don’t want to pretend this is free. Every hour spent watching a tool is an hour not spent on something else, and for a ten-person business that’s a real cost.</p><p>So I’d scale the watching to the stakes. A tool that drafts social posts for your review needs a light touch. If it gets worse, you’ll notice, and the cost of a bad draft is small. A tool that touches medications, money, or anything a customer relies on needs a written floor and a regular look, even if that look is fifteen minutes a month.</p><p>There’s also a tradeoff in how much you lean on the vendor. Some vendors offer their own monitoring dashboards, and those help. But a vendor’s dashboard measures what the vendor decided to measure. I’d use it as one input and keep at least one check of my own, tied to how the work goes in my business.</p><p>And sometimes the honest answer is to not adopt the tool yet. If nobody owns it, there’s no floor, and turning it off would be a scramble, you’re not ready to rely on it. Waiting a quarter while you sort that out is a perfectly good decision.</p><h3>The same shape everywhere</h3><p>Healthcare is ahead on this because the stakes force it. When a tool suggests the wrong dose, people notice fast, and regulators and journals start asking questions. Most industries don’t have that pressure, so the gap stays hidden longer.</p><p>But the shape of the problem is the same everywhere. AI features are showing up inside software people already use, often through routine updates. The tool keeps changing after you approve it. And the first person to notice a problem is usually someone close to the work, not someone with authority to fix it.</p><p>At Elegant Wolves Labs, this is how we think about the software we build and run. AI is useful when it makes a real task easier, and it needs clear boundaries and a person who’s paying attention after launch. The launch is the beginning of the job, not the end of it.</p><p>So here’s my one question for you this week. Pick the AI tool your team relies on most. Who’s watching it right now, and when did they last look?</p><p>If you have a good answer, I’d love to hear how you set it up. If you don’t, that’s a fine place to start, and a much easier fix than it sounds.</p><p>Thanks for reading. We write about practical AI, health care operations and calm software at <a href="https://elegantwolves.com/">elegantwolves.com</a>, and I’ll be back here on Thursday.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=029ab43af7a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Everyone uses AI now. Almost nobody has built it in.]]></title>
            <link>https://elegant-wolves-labs.medium.com/everyone-uses-ai-now-almost-nobody-has-built-it-in-5b8295860fe6?source=rss-399d456db99f------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b8295860fe6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[operations]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Elegant Wolves Labs]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 15:12:10 GMT</pubDate>
            <atom:updated>2026-11-10T16:36:43.938Z</atom:updated>
            <content:encoded><![CDATA[<h4>Small businesses crossed the adoption line fast. The harder, more useful work is turning one chat window into one dependable process.</h4><figure><img alt="wo colleagues talking over a laptop at a small office desk" src="https://cdn-images-1.medium.com/max/1024/1*8-3kdIYIVkIpsZI7jcN67A.jpeg" /></figure><p>Ask a small business owner whether they use AI and you will almost always get a yes now. Ask them what changed in the business because of it, and the answer gets quieter.</p><p>Usually it goes something like this. “I use it to write emails.” “My office manager uses it to clean up the newsletter.” “I asked it to help me think through pricing once and it was pretty good.” All of that is real, and none of it is nothing. But if you listen closely, every one of those answers is about a person having a slightly better afternoon. Very few of them are about the business running differently on a Tuesday when nobody is thinking about AI at all.</p><p>I have been chewing on that difference for a couple of weeks, mostly because a new report put a clean number on it.</p><h3>The number that caught my attention</h3><p>In October, Goldman Sachs’ 10,000 Small Businesses program published a report called <em>The AI Economy Is a Small Business Economy</em>. It is based on a survey of 2,473 of the program’s alumni, run in August. Two figures sit right next to each other on the first pages, and they tell the whole story.</p><p>Eighty-six percent of those owners said they use AI in their business. Two years ago it was 56 percent. That is a fast climb by any standard.</p><p>And 15 percent said AI is fully integrated into their core operations.</p><p>Most of the rest are somewhere in between: using it in some areas, or still experimenting without much consistency. The report says it plainly, and I think it is the most honest line in the whole thing: high adoption does not mean deep integration.</p><p>I wrote a short version of this idea on our blog earlier this week (<a href="https://elegantwolves.com/small-business-ai-integration-gap/">small business AI integration is the gap that matters now</a>). Here I want to slow down and get practical, because “integrate AI” is the kind of advice that sounds useful and gives you nothing to do on Monday morning.</p><h3>Using AI and building it in are different jobs</h3><p>Here is the simplest way I know to describe the difference.</p><p><strong>Using AI</strong> is a person opening a chat window to get help with a task. The person decides when to use it, what to give it, and whether the answer is good enough. When that person is on vacation, the AI part of the business goes on vacation too.</p><p><strong>Building AI in</strong> means a specific, recurring job in the business now runs a little differently every single time, whether or not anyone remembers to open a tab. The inputs come from somewhere predictable. The output lands somewhere predictable. Someone owns the result and knows what to do when it is wrong.</p><p>The first one is a skill. The second one is a process. Small businesses have gotten very good at the first, very quickly. The second is where the actual savings live, and it is also where most of the friction is.</p><p>That friction is not about the AI being too weak. In my experience, the model is rarely the bottleneck for small business work. The bottleneck is all the unglamorous stuff around it: where the data lives, who is allowed to see it, what “good” looks like, and what happens on the day the output is confidently wrong.</p><h3>A tale of two invoices</h3><p>Let me make this concrete with a made-up but very ordinary example.</p><p>Picture a twelve-person plumbing company. The owner’s partner handles the books. Every week she spends a few hours matching supplier invoices against job tickets, flagging the ones where the parts billed don’t match the parts used, and chasing down the odd duplicate.</p><p><strong>Version one, using AI:</strong> She pastes a couple of confusing invoices into a chatbot and asks it to explain the line items. It helps. She saves maybe twenty minutes, on the weeks she remembers to do it. The process itself hasn’t changed. If she is out sick, nothing happens.</p><p><strong>Version two, building it in:</strong> Supplier invoices that arrive by email get collected in one place. Each one is read, the line items are pulled out, and they are compared against the job tickets for that week. Anything that matches cleanly is marked as matched. Anything that doesn’t gets put on a short list with the reason (“billed for two water heaters, ticket shows one”). She reviews the short list on Friday morning instead of the whole pile.</p><p>Notice what version two needed. It needed a decision about where invoices land. It needed the job tickets to be somewhere a system could read them. It needed a rule for what counts as a match. It needed a person who reviews the exceptions, and a way to tell whether the thing is actually catching problems or just producing a nice-looking list. None of those are AI questions. They are operations questions that AI happens to make worth answering.</p><p>That is the gap. And it explains why so many owners feel like they “use AI” but can’t point to what it changed.</p><h3>Why the gap is so stubborn</h3><p>When I talk to owners about this, the same handful of reasons come up. None of them are about intelligence or willingness. They are about how small businesses actually run.</p><p><strong>The process lives in someone’s head.</strong> In a lot of small companies, the way a recurring task gets done is “however Maria does it.” That works fine until you try to hand any part of it to software, because software needs the steps written down. Writing them down is often the most valuable hour in the whole project, AI or not.</p><p><strong>The data is scattered.</strong> Job notes in one app, invoices in email, customer history in a spreadsheet somebody made in 2019. A chat window can work with whatever you paste in. A built-in process needs to know where to look, every time.</p><p><strong>Nobody owns the result.</strong> When a person uses AI to draft something, they own the draft. When a process uses AI quietly in the background, ownership gets fuzzy fast. If no one is clearly responsible for checking it, mistakes pile up silently until a customer notices.</p><p><strong>The cost question feels unanswerable.</strong> The Goldman report says more than three quarters of the alumni would like more training and resources. A separate YouGov study for Meta, covering small business owners across Asia Pacific, found the same shape: most already use AI, and most say they need more guidance to get the full benefit, especially clearer information on cost and return and on data privacy. Those are integration worries. People aren’t asking “should I try this?” They are asking “what will it really cost me to depend on this?”</p><p><strong>Trust has to be earned, not assumed.</strong> It is one thing to let AI help you write a message you read before sending. It is another to let it touch billing, scheduling, or anything a customer sees. Owners are right to be careful here. The careful ones just need a way to move forward without betting the business on it.</p><h3>What I would actually do</h3><p>If you run a small team and want to cross from “we use it” to “it is part of how we work,” here is the approach I keep coming back to. It is boring on purpose.</p><h3>1. Pick one recurring task that already costs real hours</h3><p>Not the most exciting task. The most tedious one that happens every week and that somebody quietly dreads. Invoice matching, sorting inbound requests, preparing a weekly cash snapshot, writing up job notes after a visit, answering the same five booking questions. Recurring matters more than impressive, because a task that happens fifty times a year is where small savings add up and where you will learn fastest.</p><p>It is worth noticing that the Goldman survey found 42 percent of those owners now use AI for financial analysis and business planning, more than double the 2024 rate. That kind of work is a good fit for this approach, because it is repetitive, it rewards consistency, and a person already reviews it anyway.</p><h3>2. Write down how it is done today</h3><p>Before any tool shows up, write the steps on one page. Where does the information come from? What does the person look at? What decisions do they make? What does “done” look like? You will probably find a step or two that nobody needed in the first place. That alone is a win.</p><h3>3. Change only that one task</h3><p>Resist the urge to “AI-enable the business.” Change the one process, and leave everything around it alone for now. Small, contained changes are easy to watch, easy to fix, and easy to undo. Big, sweeping ones are how you end up with three half-working tools and a team that has quietly gone back to the old way.</p><h3>4. Keep a person in the loop until it gets boring</h3><p>For the first stretch, someone reviews every output. Not forever, just until the mistakes become rare and predictable. Then you can move to reviewing the exceptions only, like the Friday short list in the plumbing example. The goal is not to remove people. It is to move people from doing the routine part to checking the parts that actually need judgment.</p><h3>5. Name the data it can touch, and nothing else</h3><p>Decide exactly which information the process is allowed to read and where its output goes. Write it down. This is the part owners worry about most, and it is the part that gets skipped most often when things are moving fast. A narrow, named scope is easier to explain to your team, to your customers, and to yourself.</p><h3>6. Decide how you will know it worked</h3><p>Before you start, pick one simple measure. Hours per week on the task. Number of mismatched invoices caught. Response time on booking questions. Check it after a month. If it saved time, great. If it just moved the work from one person to another, that is worth knowing too, and it is a perfectly good reason to stop.</p><h3>The tradeoffs, honestly</h3><p>I don’t want to make this sound easier than it is, so here are the honest costs.</p><p>Building something in takes more upfront time than opening a chat window. You will spend a few hours, maybe a few days, on decisions that feel like they have nothing to do with AI. That is normal, and it is where most of the lasting value comes from.</p><p>It also creates something you have to maintain. A process that depends on a supplier’s invoice format will break the day that supplier changes their template. Someone needs to notice and fix it. If nobody on your team wants that job, that is important information before you start, not after.</p><p>And sometimes the right answer is to not integrate at all. Some tasks are rare enough, or sensitive enough, or simple enough, that a person with a good chat tool is the best setup you will ever need. That is fine. The point isn’t to put AI everywhere. It is to put it where a repeatable process clearly beats a repeated favor.</p><h3>What this means if you build software</h3><p>A quick note for the builders reading this, because I think the gap is a message for us too.</p><p>If 86 percent of owners already use AI and only 15 percent have it built into their operations, then the useful work is not another clever chat interface. It is the plumbing: connecting to the places small businesses already keep their information, making review easy, showing clearly what the system did and why, and making it simple to undo when it gets something wrong. The flashy part is mostly solved. The trustworthy part is where people are still waiting.</p><p>That is the kind of software we like building at Elegant Wolves Labs, for what it is worth. Calm, specific, and honest about where a human should stay involved.</p><h3>The question I would leave you with</h3><p>Here is a small exercise. Think about your business on a normal week. Find the one task that happens over and over, eats more hours than it should, and already has someone checking the result. Now imagine it running a little more on its own by spring, with that person reviewing only the parts that genuinely need them.</p><p>If you can picture that clearly, you already know where to start. If you can’t, the first step is just writing down how that task works today. Honestly, that step is underrated, and it costs nothing.</p><p>Thanks for reading. If you want more practical, unhyped notes on AI, software, and running a small operation, we write about this most weeks at <a href="https://elegantwolves.com/">elegantwolves.com</a>. I’d love to hear which task you’d pick.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b8295860fe6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>