|
Welcome back to the Launch Key 🚀 |
This week I am testing something different - but first a little background. |
Early this year, Gary Frey asked me to be on the AI panel for BGW’s September CEO Convergence. And I warned him that the speed of AI might escape me by the time September got here. Plus, I’m no longer in the daily mix of running a business. How could I stay current and add to the discussion? |
In my attempt to keep up, I built a daily 5am email tech briefing creating an AI agent using Claude, Python, SQLite, launchd and Resend to aggregate key YouTube, RSS and X feeds. When Anthropic added improved design skills, I automated html and turned the daily Futurist into free git pages. Claude creates pre-market charts and delivers commentary with a technology optimist voice (think Marc Andreessen) who believes we are 3 years into the 30 year AI revolution. I write none of it. |
I added and deleted sources and as Summer wore on, I wanted to make sure I wasn’t blindly trusting Claude agents, while also seeing the bigger picture. I experimented with open source LLMs and built a few useful tools connected to Notion. I listened to pods as I walked the blind dog. And I enlisted Grok to do market research while reviewing the Futurist output, before creating the kind of brief SMB leaders might find useful. We’ll see if I can hold my own at the conference! |
The result is this Launch Key Summer 2026 AI update. I trained Grok on 150+ weeks of Launch Key issues before asking him to write this week’s Pull to Eject section. |
Stories will change tomorrow, but read the newsletter and see what you think. TLDR slides can be downloaded here.👇️ |
|
|
LaunchKey_AI_SMB_Q3_2026.pdf
254.54 KB • PDF File
|
| Download |
|
I edited it all in Beehiiv, but this AI Update is an AI product. |
Here’s how you can make this better: |
Tell me where it’s wrong Other sources that should be included? Forward it to friends Comment below: valuable? or already known? other? Email me if you want the detailed research brief
|
Let's get into it. |
|
|
|
|
|
|
Table of Contents |
|
|
|
|
Pull to Eject |
People are using more AI. |
The companies that sell it are making more money. The share of outfits that can point to earnings and say this came from the tool has barely moved. McKinsey’s August survey put “some EBIT impact” at 37 percent — same as last year. The group that can call the impact significant is still about 6 percent. |
Conviction is growing faster than the return. That is not a scandal. It is what happens when a junior employee gets a company card and nobody assigned them a job. |
|
|
Two different tools |
If you only remember one distinction from this quarter, make it this one. |
A chatbot answers. You type. It writes. You decide. Draft the email, summarize the contract, explain the clause you don’t want to read twice. Cost is usually small. Risk is a sentence you might paste without reading. A person is still in the loop. That is the tool most of you already have, and it is still the one that works. |
An agent acts. You give it a goal. It plans steps, opens your calendar, pokes the CRM, retries when it misses, and keeps going until the job is done or it makes a mess. Qualify the lead and book the call. Watch the inbox and handle the first-tier tickets. Process the week’s invoices. |
Cost is not small, because each job is not one reply. It is a pile of hidden steps. McKinsey’s work last month on agent economics is the useful version: in some service workflows the tokens — the metered pieces of text the vendor charges you for — are only 20 to 25 percent of the running cost. The rest is people checking the output. Risk is not a sloppy paragraph. Risk is a live action in a live system. |
The industry spent the summer talking as if those two things were the same product on a spectrum. They are not. One is a junior who writes memos. The other is a junior you handed the keys. |
That is why bills jumped for people who “just turned the agent on.” Unit prices fell. Usage exploded. OpenRouter-style traffic reports from the summer show agent workloads chewing through far more tokens than a human sitting in the same chat box. Anthropic’s run-rate kept climbing. The Futurist note on August 25 was the tell that matters for a medium size shop: the most expensive model is losing users to cheaper ones that are good enough. |
Capability is not the scarce resource. Judgment about when to spend it is. |
|
What actually worked |
The jobs that paid were boring on purpose. |
First-pass support. Appointment booking. Lead follow-up that used to die in the inbox. Invoice reminders. A first draft of a document someone still reads. When a small firm measures the old way first — how long it took, what it cost, how often it was wrong — those narrow loops still show real time back. Twenty to forty-five percent in the process, not in company profit. |
Salesforce’s late-August study (Free Knowledge) of agent deployments made the unfashionable point: the shops that got to a return first were not the ones that launched first. They were the ones that narrowed the job, cleaned the data that job needed, and decided in advance where a person steps in. Average time to “meaningful ROI” in that sample was about eight months. Professional services — late, careful, allergic to theater — got there faster than sectors that bought the hyped up demo. |
That should sound familiar. You have spent a career watching the first mover collect the press and the second mover collect the margin. |
A serious, narrow stack for a small operation still lives in the low hundreds of dollars a month if you refuse to put the premium model on every task. JPMorgan’s look at millions of small-business card records found most shops that pay for AI at all are in the neighborhood of forty dollars a month. Goldman’s 10,000 Small Businesses cohort said only 14 percent had it in the core of the operation. |
Most of you are not late. |
|
The bill, without the poetry |
A token is a scrap of text. A page is a few hundred. A chatbot answer is a few thousand. An agent that plans, looks things up, calls your software, and tries again can burn tens of thousands on one job. |
Vendors will keep cutting the price per scrap. That is not the same as a smaller invoice. More people, more steps, more retries. CloudZero’s summer panel had AI at about 2.6 percent of the median cloud bill and still compounding month to month. |
About one in five companies in McKinsey’s survey said operating cost was already constraining how much they let people use the tools. KPMG’s pulse earlier in the season found a large share of organizations narrowing or pausing agent rollouts once the run cost showed up. Gartner is still on the record that a chunk of agent projects get cancelled by 2027 for cost, foggy value, or missing controls. Read that as a market learning to say no. Not as a reason to freeze. |
The number that matters is not tokens. It is cost per finished job, including the person who checks the work. Ticket actually closed. Invoice actually posted. Appointment actually on the calendar. If you cannot name the job, you are not ready to give the tool a goal. |
|
|
|
What got cheaper that you can use |
The other story of the quarter is that “good enough” got good enough. |
Open-weight models — files you can run without sending every prompt to a vendor — kept arriving. Meta’s Muse Glimmer under a clean Apache license. Alibaba’s Qwen line shipping mixture-of-experts models that keep only a sliver of the network “on” per request, which is a fancy way of saying faster and cheaper for the same quality band. Tencent dropped a large open model at the end of August. You do not need to run any of this on a machine in the closet. You need to know the option exists so a salesperson cannot tell you the frontier model is the only grown-up choice. |
Routing is the adult habit: cheap model for volume, expensive model for the 10 or 20 percent that is actually hard. Self-hosting only starts to win after you have real daily volume. Most of you should stay on a subscription and a cheap default until one workflow is truly busy. |
The Futurist days in late August also kept surfacing the same warning in different clothes. Agent “auto” modes that can be talked out of their own guardrails. Skill files that leak. Fake authors contaminating paper trails. A children’s studio telling artists to “keep a human in the loop,” which is a phrase that means nothing unless someone can say whose loop and what they check. Trust is lagging production speed. |
Handmade “no AI” signs showed up as a sales pitch. That is the other side of cheap generation. Average output is no longer scarce. Taste, a name people trust, and the judgment to throw work away — those still are. |
|
What to ignore |
Ignore the weekly model name. Ignore anyone who cannot tell you the job, the baseline, and the kill date. Ignore the board slide that treats “we have AI” as a strategy. Ignore the implication that you are behind because you are not running a swarm of agents. Large companies scaling agents moved. Smaller ones, in McKinsey’s cut, were flat. That is not a moral failing. That is a market that has not yet made the economics obvious at your scale. |
Ignore, also, the idea that the tool replaces the part a client is paying you for. The tool does exactly what it was designed to do. It will complete the task you pointed it at, including the wrong one, with confidence. The risk is not that it takes your chair. The risk is that you stop practicing the part you’d notice was off if the stick got handed back. |
|
What to do this week |
Pick one process you already understand. Time it. Cost it. Write down how often it is wrong. |
Then decide whether you need an answer or an action. If you need an answer, keep the chatbot and read the draft. If you need an action, give the agent one job, one system it can touch, and a person who can pull the plug. |
Do not buy a second tool until the first one has a number. |
That’s the quarter. The models will have new names in December. The distinction will not. |
|
|
Now go launch something 🚀 |
|
|
| ❝ |
| |
Much to the surprise of the builders of the first digital computers, programs written for them usually did not work. |
| |
| Rodney Brooks |
|
|
|
|
Free Knowledge |
What 2,000 leaders said about winning with agentic AI. The finding is the issue: being first is not the advantage. Clean data for that job, a narrow scope, and a human stop before launch beat model choice. Professional services — late, careful — hit a return faster (~6.5 months vs ~8 on average). |
|
|
|
|
Recommendations |
📕 MGMT Playbook : Practical management insights straight to your inbox every Wednesday. |
🏢 Retirepreneur : Your guide to turning decades of experience into profitable consulting, coaching, or course businesses, without risking your retirement security. |
🧾 Redefining Retirement : Too Young to Retire, Too Old To Take Orders? Yeah, Me Too. |
|
|
Visual Crapshoot |
|
|
|
|
|
|
Share Launch Key | The Newsletter for Late-Career Entrepreneurs |
Thanks for referring the Launch Key newsletter! 🚀 |
You currently have 0 referrals, only 1 away from receiving Free tools to launch your new idea or digital product. Instant PDF download with 1 referral. |
|
|
|
|
Or copy and paste this link to others: https://www.thelaunchkey.com/subscribe?ref=MRjCg0GwZI |
|
|
|
|
|
Launch Key readers – thank you for your support and feedback. I appreciate each and every one of you as I work to build something you value. Remember, if there's anything you'd like to share — a recommendation, a story idea, or just a note to say hi, hit the reply button and fire away. |
| |
| ~ Rob |
|
|
