Field notes

Straight answers on automation

Buyer's guides for people spending real money. What vendors actually sell, what it costs, what breaks later, and the times the honest answer is to not automate it at all.

Security

Your AI Agent Has Write Access. Does Anything Have an Undo Button?

Most small businesses give an AI agent the ability to send emails, update records, or edit files without ever asking whether any of it can be undone. It usually can't — not by the AI tool itself. What real recovery exists in the systems you already run, and where to put a checkpoint before you grant write access.

2 October 2026 10 min read Read the checklist
Cost control

Free ChatGPT, Free Gemini, Free Copilot: What Actually Happens to What You Paste In

Somebody on your team is already using the free tier of a major AI tool for real work. The paid business version of that same tool promises your content is never used for training. The free one, by default, usually doesn't.

1 October 2026 10 min read Read the comparison
How it actually works

Why Your AI Gave Two Customers Two Different Answers

Two people ask the same AI the same question and get two different answers. That's not a malfunction — it's sampling, the same mechanism that makes AI writing feel natural instead of robotic. The question that actually matters is where you can afford it and where you can't.

30 September 2026 9 min read Read it
Buyer's guide

The SLA Language That Sounds Reassuring and Promises Almost Nothing

An AI vendor's service-level agreement is written to be skimmed. Read closely, the uptime guarantee, the maintenance exclusion, the credit cap, and the beta disclaimer usually commit the vendor to far less than the paragraph sounds like it promises.

29 September 2026 10 min read Read the checklist
Field notes

What Actually Happens to the Job AI Was Supposed to Replace

Nobody gets walked out the door because a model can draft an email. What actually happens, according to 2025-2026 labor research, is messier: the task moves, a new job appears to check it, the door narrows for whoever would have learned the role by doing it, and the skills that get paid for shift.

28 September 2026 10 min read Read the research
Regulation

The FTC Is Suing Companies Over AI Marketing Claims. Here's the Pattern

DoNotPay's “AI Lawyer.” Workado's 98% accuracy claim that was really 53%. Growth Cave's $48.6 million settlement over an AI that didn't automate what it promised. Real FTC cases, one recurring pattern, and a self-check before your own sales page becomes the next one.

27 September 2026 10 min read Read the cases
Security

The Voice on the Phone Is No Longer Proof of Anything

A deepfake video call talked a finance employee into moving $25.6 million in a single day — using AI-generated faces and voices of executives he already knew. The technology that made it work is cheaper now, not less common. The low-friction protocol that still stops it.

26 September 2026 9 min read Read it
Cost control

When the Cheaper AI Vendor Costs More by Month Six

The monthly price on an AI vendor's pricing page is the easiest number to find and the least predictive of what the tool actually costs you. A meaningful share of small-business AI vendors won't exist in their current form in two years — how to price that risk before you sign, not after.

25 September 2026 10 min read Read the framework
How it actually works

What an AI Agent Can Actually See Once You Connect It

The consent screen says “view and manage your mail.” What it grants is a standing credential that keeps working until you revoke it — and research suggests most connected AI tools hold far more access than their actual job requires.

24 September 2026 10 min read Read the audit
Buyer's guide

How to Actually Read an AI Vendor's SOC 2 Report

“We're SOC 2 compliant” is not a magic phrase and it is not a pass/fail test. It's an audit report with a scope, a date, and often a list of exceptions — and almost nobody who accepts the badge ever reads the document behind it.

23 September 2026 10 min read Read the guide
Field notes

Why AI Pilots Stall After the Demo

The demo worked. Everyone agreed it was worth doing. Three months later the tool is still technically live, and almost nobody opens it. That gap is not rare — it is the modal outcome, and the 2025-2026 research is specific about why.

22 September 2026 10 min read Read the field notes
Regulation

Colorado's AI Law Keeps Changing Dates. Here's What Actually Applies

Delayed by a special session, paused by a federal court, then repealed and rewritten into a narrower law. The Colorado AI Act's fourth effective date is 1 January 2027 — who counts as a “deployer,” what's actually required, and what to do regardless of whether the date moves again.

21 September 2026 10 min read Read the timeline
Security

The Resume, Invoice, or Support Ticket Could Be Attacking Your AI Tool

You did not choose to be vulnerable here. A customer sends you a completely ordinary document, and hidden inside it — invisible to you — can be instructions aimed at the AI tool that reads it on your behalf. Real documented cases, and how to cut the blast radius.

20 September 2026 10 min read Read it
Cost control

Why Your AI Bill Came In Higher Than the Quote

The quote was built on a demo: one prompt, one answer. The bill was built on real use — and the multi-step, agentic workflows AI is best at can burn far more of the metered unit per task than the chat window you tested with. What drives the multiplier, and the five questions that catch it before you sign.

19 September 2026 10 min read Read the breakdown
Cost control

Your Insurer Quietly Changed Its Mind About AI

If an AI tool you use causes a loss, is it covered? Until recently the answer was usually yes — but only because policies never mentioned AI. Insurers spent 2026 ending that in both directions, and the change arrives at renewal without a conversation. Two form numbers to search for, and a three-minute check.

16 September 2026 10 min read Read the check
Buyer's guide

Six Clauses to Check Before You Sign With an AI Vendor

Nobody reads AI terms of service, and you do not need to read all of it. Six clauses account for nearly everything that goes wrong, each takes two minutes to check, and two of them decide whether adopting the tool is reversible at all. Plus the one-line search that reveals which answers have a shelf life.

15 September 2026 10 min read Read the checklist
How it actually works

Should You Tell Customers It's AI?

The nervous question owners actually ask. The 2026 data answers it clearly and then points somewhere else entirely: disclosure is not what costs you customers, having no way to reach a person is. Roughly three quarters say they would leave a brand offering only AI — and the disclosure statistics contradict each other.

14 September 2026 10 min read Read the data
Cost control

Your Software Bill Stopped Being Predictable

Per-seat pricing is ending, and not because vendors got greedy — an agent never logs in, so there is no seat to charge for. Usage-based went from about 30% of SaaS companies to roughly 85%, hybrid is now the norm, and the risk moved to you. The six models and the three terms worth negotiating.

13 September 2026 10 min read Read the models
How it actually works

Your Developer Is Using AI. Here's What to Ask Them

Between a quarter and nearly half of AI-generated code carries a security flaw before review, and output rose three to fourfold while security findings rose roughly tenfold. The more troubling finding is the false confidence. Why asking whether your developer uses AI is the wrong question, and the six that are not.

12 September 2026 11 min read Read the questions
Buyer's guide

You Probably Don't Own the AI Content on Your Website

You paid for the tool, so the output is yours to use. Whether it is yours to stop anyone else using is a different question, and the answer is frequently no — US copyright needs human authorship and the Copyright Office says prompting is not enough. Where that bites, and why trade mark rather than copyright is the fix.

11 September 2026 10 min read Read it
Regulation

AI Is in Your Hiring Whether You Chose It or Not

Ask most owners whether they use AI in hiring and they say no. Then you look at the software: the tracking system ranks, the job board shortlists, the parser decides what counts. New York has required an independent bias audit since 2023, compliance ran at about 5%, and enforcement is going proactive.

10 September 2026 11 min read Read the guide
Field notes

Your Customers Are Asking an AI Instead of Googling

Search traffic is down and nothing obviously broke. The part nobody expected is what happened to the visits that still arrive — a channel that converted at half the rate a year ago is now the strongest one measured. What the numbers say, who they are actually about, and why a figure we went to verify was not on the page it is attributed to.

9 September 2026 11 min read Read the field notes
Security

The FBI Started Counting AI Fraud. The Number Is Nothing Like What You've Been Sold

For the first time there is an authoritative figure rather than a vendor estimate. AI-involved business email compromise came to just over $30 million against a $3.05 billion category — far less than the marketing implies, and the widely quoted alternatives cannot be traced to a source. It is also an undercount by the FBI's own admission, and the five controls that actually stop it never mention AI.

9 September 2026 11 min read Read the numbers
Regulation

The State AI Law Everyone Told You to Prepare For Was Repealed

For two years the Colorado AI Act was the reference point for American AI regulation. It never applied to anyone — repealed in May 2026, before its own start date. Meanwhile Texas, California and Illinois laws quietly took effect on 1 January. Which of them actually reach a business that uses AI rather than builds it, and why compliance advice keeps going stale.

9 September 2026 12 min read Read the map
Security

Your Staff Are Already Using AI You Didn't Approve

Somebody has pasted customer information into a free AI tool this month. They were not being reckless — they were behind, and nobody had given them a sanctioned way to do it. Shadow AI went from 20% to 43% of AI-related breaches in a year while the share of firms requiring approval fell, and two thirds of workers used AI believing it broke policy. Why a firmer ban is the option that measurably fails.

9 September 2026 11 min read Read it
Regulation

The EU AI Act's August 2026 Rules: What a US Small Business Has to Do

The transparency duties went live on 2 August and the expensive high-risk regime was quietly pushed to December 2027, which most of the compliance email being sent to you has not caught up with. A four-question scope test, which of the four duties are yours rather than your vendor's, and the point where we would tell you to stop reading and call a lawyer.

8 September 2026 11 min read Read the guide
Security

An AI Agent Breached a Real Company. What That Means If You Run Five

In July an autonomous agent spent two and a half days inside Hugging Face's production systems. It got in through credentials sitting on the public web, which is the most ordinary failure there is. What generalises to a business running three automations, the five permissions worth auditing this week, and the one question to put to your AI vendor.

8 September 2026 12 min read Read it
Field notes

Half of Companies Bought AI and Changed Nothing Else

Around 48% introduced AI without redesigning the workflow around it, and only 12% redesigned at scale. The tool was probably fine — it made one fast step faster inside a process whose duration is mostly waiting. How to timestamp a flow honestly, and the four-step fix for a business with no transformation budget.

8 September 2026 10 min read Read the diagnosis
Cost control

The Median AI-Using Small Business Now Runs Five AI Tools

Nobody designed that stack, it accumulated one defensible decision at a time. The cost is rarely the count — it is the same capability arriving three times through tier upgrades nobody compared, and two systems giving customers different answers. A capability grid and a ninety-minute audit that needs only a card statement.

8 September 2026 11 min read Read the audit
How it actually works

Which Decisions You Should Never Hand to an AI Agent

By 2026 the answer to whether an agent can do the job is nearly always yes, which makes it the wrong question. Four tests for what is safe to delegate, a three-column decision table, the five things that stay human regardless of how good the model gets, and the four ways autonomy arrives without anyone deciding.

8 September 2026 10 min read Read the framework
Buyer's guide

Your Software Is Growing Its Own AI Agents. Do You Still Need to Buy Any?

Gartner expects 40% of business applications to carry task-specific agents this year, up from under 5%. Much of what you are about to buy is arriving free in software you already own — and one category never will, because every vendor's AI stops at its own boundary. A wait-or-buy table, and why “included” means a tier upgrade.

8 September 2026 10 min read Read the comparison
How it actually works

AI That Remembers Your Business: What Changes, and What Breaks

Memory is why these tools finally stuck — it removed the re-explaining tax that quietly killed most 2024 pilots. It also means something holds a durable record of how you work, which drifts out of date, can surface the wrong client's context, and may be hard to erase on request. The three failure modes and an afternoon's setup.

8 September 2026 11 min read Read it
Buyer's guide

Custom AI Software for Trades: When It's Worth It, and What It Costs

Most trades who ask us for custom software do not need it. There are three layers, at roughly $29 a month, $20 a month, and $5,000 upwards, and the middle one gets skipped because nobody earns commission selling it. The four tests for when a build is genuinely justified, and the five ways custom fails after delivery.

5 September 2026 9 min read Read the guide
How it actually works

What an AI Receptionist Does When It Doesn't Know the Answer

Every demo shows the call that goes right. This is the one that goes wrong: the five ways a real call derails, the four escalation behaviours ranked, why warm transfer is often a paid upgrade two tiers above the price you were quoted, and seven test calls that take twenty minutes and tell you more than any demo.

5 September 2026 11 min read Read it
Buyer's guide

Best AI Website Builders for Small Business Owners (2026)

All seven build you a decent site in ten minutes, so that stopped being the question. This compares what the headline price actually commits you to, what it renews at, and whether you can leave with your site. Ours is in there, and it is the only one that does not publish what publishing costs.

5 September 2026 12 min read Read the comparison
Buyer's guide

Best AI Automation Agencies for Small Business (2026)

Eight vendors compared on what they actually sell and what they publish about price, including the platforms that cost $12 a month and the reasons you might not need an agency at all. We put ourselves in the same table, with the situations where you should hire somebody else.

5 September 2026 14 min read Read the comparison

What goes here next

The run on what changed in business AI through 2026 is complete: the regulation, the first real agent breach, the missing returns, the five-tool stack, where the line on autonomy belongs, what is arriving free in software you already own, and what it means that the assistant now remembers you.

Each piece answers a question a buyer actually asks us, with every figure dated and attributed, including the ones that argue against buying anything. If there is a question you want answered honestly — including one where the answer might be that you should not buy from us — send it over.

Suggest a topic