One idea
Your shop is building AI faster than it can govern it.
The tools arrive one login at a time. The workflows they create can move faster than the people responsible for them.
of AI-engaged small businesses surveyed by Pax8 have built at least one AI tool now in regular use.
found AI running in their operations without approval in the last six months.
Pax8 surveyed AI-engaged small businesses this quarter. Two-thirds have built at least one AI tool now in regular use. And 35 percent found AI running in their operations without approval in the last six months.
Nobody bought a robot. Somebody pasted a password into a chat window. Somebody connected an agent to the inbox. Somebody automated the follow-ups and never told anyone. That is how shadow workflows get made. Not by decision. By drift.
Every one of those is a workflow running without a human gate. It works until the day it does not, and nobody knows it exists until the damage is already out the door.
The Aberdeen and Grampian Chamber of Commerce argues shadow AI is usually a leadership gap, not a staff problem. Employees using AI report saving an average of 7.75 hours per week on admin work, and banning AI just drives it further underground. The answer is not a ban. It is a gate.
Sources: Pax8 via GlobeNewswire, Channel Dive, and Pax8 company announcement. Opposing view: Aberdeen and Grampian Chamber of Commerce.
One example
The CEO whose agent posted his bank statement in the work chat.
The Wall Street Journal reported it this month. A startup CEO's "Personal CFO" agent conflated chat groups and posted his bank balances to the company Slack. Another agent bought a bikini on store credit, bypassing the approval step entirely.
These agents break multi-step tasks into steps the human never sees. Ambiguous instructions become real actions. No one reviewed the steps in between, because there was no step that said "stop and ask."
That is the exact failure mode that matters to an owner. Not a chatbot giving a wrong answer. An agent with permissions, no human gate, and the authority to act. The shops in the Pax8 survey have these running. Most of them did not approve them.
Bloomberg Intelligence projects about 500 billion dollars, or 20 percent, of online commerce could be carried out by AI agents by 2030. Major institutions see agentic commerce as a growth wave to set trust standards for, not a failed experiment. The question is not whether agents act. It is who reviews the steps.
Sources: The Wall Street Journal, Katherine Bindley, The Times, and Bloomberg Law. Opposing view: Bloomberg Intelligence projection, via Bloomberg Law.
More from the newsroom this week
Four signals worth carrying into the next decision.
53%
AI takes the workload. The owner still absorbs the difference.
Bluevine surveyed 729 small business owners: 53 percent say AI already absorbs the workload of at least one full-time employee, and 40 percent are freezing hiring despite growth goals. Revenue up, headcount flat. But 65 percent of lean-team owners say staying lean hurt the business or their well-being. One in three lost the ability to take time off. The owner absorbs the difference as burnout.
MIT's GenAI Divide study found 95 percent of organizations saw no measurable profit-and-loss return on their generative AI pilots, which undercuts vendor-commissioned survey claims about AI absorbing headcount. Read both before you believe either.
Sources: PRNewswire (Bluevine) and Bluevine Lean Team Economy Report. Opposing view: MIT GenAI Divide via Storyboard18.Open source check: a third source is still being confirmed.
Seconds
Scheduling becomes an agent's first shift.
Toast launched an AI Scheduling Agent for restaurants that builds a full staff schedule in seconds from the operator's own POS sales, availability, and overtime-risk data. Hospitality labor scheduling is now the wedge for agentic AI, and it runs on data the owner already collects. The barrier to adopt is low.
Restaurant Dive reports restaurant AI tools face employee and customer pushback when rolled out faster than staff can accept, and experts warn the tools should empower workers rather than create friction. An AI schedule the team does not trust is a schedule they work around.
Sources: PYMNTS, Toast company announcement, and QSR Magazine. Opposing view: Restaurant Dive.
200
Who gets to grade the models?
Reuters reports Anthropic and OpenAI have sought control over what outside evaluators can publish about their models, while 200 researchers signed a letter demanding independent findings. The incident-reporting workflow the White House is mandating is being negotiated right now. When you buy an agent tool, the lab grading its own homework is your trust problem.
Anthropic's own September essay committed to giving embedded evaluators the right to publish key findings without editorial control, keeping only narrow redaction rights for security or confidentiality, and barring redactions merely because findings are unfavorable. That undercuts the claim that the labs intend broad censorship. Watch what they do, not what either side says.
Sources: Reuters, PYMNTS, and Brookings Institution. Opposing view: Anthropic essay via Unite.AI.
<10 min
The work automated was the re-keying.
Commure and Compassus took referral intake in home-based care from an hour to under ten minutes, live across 79 programs in 19 states. The work that got automated was not the care. It was the re-keying. And the vendor built it with the operator, which is why it shipped.
Healthcare IT News cautions that AI-enabled triage needs real clinical evidence, and conflating a triage decision with a diagnosis is where a lot of AI in healthcare gets into trouble. Speed in intake is a win. Speed in diagnosis is a different claim.
Sources: GlobeNewswire (Commure) and Commure company announcement. Opposing view: Healthcare IT News.Open source check: a third source is still being confirmed.
Start where it is free
Hear your workflow the way a system would see it.
And if you want help finding the workflows that are trapped in your head, the ones only you know how to do, start where it is free. Ten minutes. Talk through one workflow in your own words, and hear it played back to you the way a system would see it.
Start the ten-minute conversationSee you next week.
Tabron.