This Week in AI: What Actually Matters For Your Business

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

Every Saturday I sit down and try to figure out what actually happened in AI this week that matters for the operator on the other side of this email. Not what was loudest. Not what got the most retweets. What you should actually do something about on Monday morning.

Some weeks the answer is "not much, keep your head down." Some weeks the answer is "drop everything, this changes the math." This week is in between. There were a handful of moves worth your attention, a handful of headlines that sounded big but mean nothing for you yet, and one thing that nobody is talking about that you should be paying close attention to. Let us walk through it.

What Actually Matters

Model pricing keeps falling, and the floor under your stack just dropped again.

This is the slow moving story of the year and it does not get the headlines it deserves. The cost of running serious model work continues to drop, on roughly a six month cycle, and the implications for small operators are massive. The work you could not afford to automate eighteen months ago because the per call cost was prohibitive is now economically obvious. Workflows that required careful prompt economy are now run wastefully because the marginal cost is approaching zero.

What you should do. Revisit any automation you abandoned six to twelve months ago because of cost. The math is different now. The transcription pipeline you killed because it was eating margin. The summarization layer you said no to because it was nice to have. The lead enrichment job you pulled because the per record cost did not pencil. Pull the spreadsheet back out. Re run the math on current pricing through whatever provider you use, including Galaxy.ai for unified access. Most of those projects are now obviously worth doing.

Agentic features are creeping into mainstream tools, quietly.

You have been hearing about agents for two years. Most of what got hyped early was demoware. What is happening now is less flashy but more important. The everyday tools you already pay for are quietly adding agentic features that actually work. Your CRM is running follow up sequences on its own. Your email client is drafting responses with full context awareness. Your calendar is scheduling and rescheduling meetings without you confirming each one. Your accounting tool is categorizing transactions and flagging anomalies before you ask.

This is not an "agents are coming" story. This is an "agents already arrived inside your existing software and you have not turned them on" story.

What you should do. Take an hour this weekend and audit the AI features in every tool you already pay for. Most of them have shipped agentic capabilities in the last quarter that you have not enabled. The number of operators paying for sophisticated AI features they do not know about is genuinely embarrassing.

Voice AI has crossed a quality threshold that changes phone work.

This is the one I think most operators are sleeping on. Voice models for both inbound and outbound phone work have gotten meaningfully better in the last six months, and the use cases that used to feel cringe are now genuinely useful. Appointment confirmation calls. Lead qualification. Customer service triage for common questions. Basic outbound research calls.

The economics are wild. A voice agent that can handle the call your front desk used to handle costs roughly the equivalent of one hour of a human's wage per month, and it works twenty four hours a day. For service businesses with high call volume, this is a genuinely strategic shift, not a nice to have.

What you should do. If you take more than ten phone calls a day for routine tasks, prototype a voice agent this month. Even if you decide not to deploy it, the act of building the prototype will tell you which parts of your phone workflow are actually high judgment work that needs you, and which parts are routine that should be automated.

What Sounded Big But Means Nothing For You Yet

Several major models had benchmark updates this week.

You will see these as headlines. "X model now scores Y on Z benchmark." Almost none of this matters for the operator running a small business. Benchmark deltas at the frontier rarely translate to noticeable differences in the everyday work you do. If your existing model handles your work well, you do not need to switch.

The exception. If you have a specific workflow that has been failing on your current model, occasionally a benchmark update is a signal that the failure case has been addressed. Worth re testing the workflow with a current frontier model if it has been more than three months since you tried.

Otherwise, ignore the leaderboard race. The leaderboard race is for people who use AI as a hobby, not for people who use it to run a business.

A few new specialty AI tools launched this week with massive press.

Every week there is at least one new AI tool that gets covered everywhere, raises a noisy round, and posts viral demos. Most of these tools will not exist in eighteen months. Some of them will. You cannot tell which from the launch coverage.

The rule I follow. Wait ninety days. If a tool that launched loudly is still being talked about by the people who actually use it, not the people who write about it, then it is worth a serious look. Most of them quietly disappear during the ninety day window.

This is also the rule for protecting your stack from the FOMO churn. You do not need to be on every new tool the day it launches. You need to be on the right tools that have proven they will be there next year.

The Story Nobody Is Talking About

Here is the underreported one.

The gap between businesses that have integrated AI into their actual operations and businesses that have a ChatGPT subscription they use occasionally is widening fast, and it is starting to show up in basic financial metrics like cost per customer acquired, time to fulfillment, and gross margin per employee. The integrated businesses are pulling away. The occasional users are not.

This is not a future problem. It is a current quarter problem.

If you are still using AI as "I will paste something into ChatGPT when I think of it," your competitors who have built actual workflows are running at a different cost structure than you, and they will be able to either undercut you on price or invest the difference in better customer experience. Either way, you lose ground.

The fix is not to use AI more. The fix is to integrate AI into specific workflows that show up on your P&L. Customer service response time. Lead qualification cost. Content production output per dollar. Sales follow up coverage. Pick one. Build the workflow. Measure the before and after. Repeat with the next bottleneck.

This is the actual race that is being run right now. Not the model race. The integration race. And the operators who treat this as urgent are the ones who will be standing in two years.

Tools Worth Knowing About This Week

Three tools came across my desk this week that earned a closer look. Not endorsements. Just things worth being aware of.

A new browser based agent layer that lets you describe a task and have it executed across your existing web tools. The use case I have been testing is "summarize this week's revenue across these four dashboards and post the summary to my team chat." Works reliably. Worth a look if you spend time aggregating data across multiple web tools.

An updated version of the meeting note category with notably better speaker identification and action item extraction. Fathom remains my pick for most operators because of price and reliability, but the category is competitive enough now that it is worth re evaluating annually.

An automation platform feature update that makes building multi step AI workflows meaningfully easier than it was six months ago. Worth checking if you have been on the fence about whether to learn Make.com or one of the alternatives. The barrier to entry on this category keeps dropping.

What I Did This Week That You Should Steal

A few things I tested in my own business this week that earned their place.

Switched my morning brief workflow to use a different model for the synthesis step. Result, slightly better summaries with less hedging. Cost difference, negligible. Time saved per brief, maybe ninety seconds. That adds up.

Built a quick prompt for converting meeting transcripts into proposal drafts. When a sales call ends and the prospect wants a proposal, the transcript plus a templated prompt produces a first draft proposal in roughly two minutes that I edit in another fifteen. Used to take me ninety minutes from scratch.

Killed three subscriptions that did not earn their place this quarter. Recovered roughly seventy dollars a month. Reinvested fifty of it into upgrading a tool that has been carrying real weight. Net thirty dollars a month back to the business, plus a sharper stack.

The pattern there is the same as always. Run the audit. Cut the dead weight. Reinvest in the strongest tools. Repeat.

The Quiet Mistake Most Operators Are Making

One more pattern from this week worth naming.

I had four separate conversations this week with operators who said some version of the following. "I am waiting until things settle down before I commit to a stack. The space is moving so fast I do not want to invest in something that is going to be obsolete in six months."

I understand the instinct. It is wrong, but I understand it.

Things are not going to settle down. Not this year. Probably not next. The category is in active expansion and the rate of new tool launches and feature updates is, if anything, accelerating. Waiting for the dust to settle is waiting for a thing that is not going to happen on the timeline you want.

The operators who are winning are not the ones who picked the perfect tools. They are the ones who picked good enough tools, built workflows around them, and ate the cost of occasionally migrating when something genuinely better came along. The migration cost is real but small. The cost of waiting is enormous and invisible, which is why most people underestimate it.

If you have been sitting on the sidelines waiting for clarity, the clarity is this. Pick the wrappers and platforms with the largest current user bases and the strongest funding runways. Build your workflows on those. Re evaluate annually. Move on. The hand wringing about which tool will win the long game is wasted energy that could be spent compounding output starting today.

The Action Items For Your Monday

If you do nothing else this week, do these three things.

Audit one tool in your stack you have not touched in thirty days. Cancel it or commit to using it.

Turn on one AI feature in a tool you already pay for that you have been ignoring. Use it for one week. Decide if it stays on.

Pick one workflow that is currently manual and that costs you more than two hours a week. Sketch out what it would take to automate it. Even if you do not build it this week, the sketch will tell you something useful.

That is the week. See you Sunday for the strategic piece.

Until tomorrow,

Jordan