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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
We are part-way through a run on what actually changed in business AI during 2026 — the regulation that took effect in August, the first agent-driven breach at a company that should have been hard to breach, why the returns did not show up, and where the line on autonomy belongs. A new piece goes up every two days, 12 still to come, next on 21 September 2026.
Each one 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.