five signals.
one week.
nobody counted
anything._

the census bureau's business trends and outlook survey puts production use at roughly 17 to 20 percent of firms, under 20 percent at the smallest ones, against 37 percent at companies with 250 or more employees.

the 89 percent figure travels because it is cheap to collect. it counts anyone who opened a chatbot once in the last year, which is a survey answer rather than an operating change. the condition it creates is a room full of owners who believe they are two years behind and start buying at that speed, which is the exact state a vendor prices against. the cost is real money spent catching up to a race that four out of five of your competitors are also not running, and the spend lands on tools rather than on the process that is actually leaking hours.

the fake fix is an ai readiness assessment, which ends in a document and a score and no change to how anything runs on monday. take one process that costs you hours every week, write down what it costs now and what it should cost, and move that one into production before you buy anything else. that puts you in the 20 percent on something that pays, which is the only version of the number worth having.

the condition is a channel that kept being priced as though it were still scarce. one person can now send in a week what used to take a quarter, so every inbox your buyer owns received the same three paragraphs in the same month, and the filters learned the shape of it faster than anyone could rewrite the copy.

the cost is not the seat fee. it is the sending domain you burned and the long stretch where your company name reads as noise to the exact accounts you most wanted. that damage does not appear in the tool's dashboard, because the dashboard counts what left rather than what landed.

the fake fix is more of it with better disguises: a second domain, a warmup service, a tool that promises to sound human. sounding human was never the constraint. having a reason to write was. cut send volume down to accounts where something actually changed, a funding round, a new operator, a system they just bought, and write to that one fact. campaigns run that way report 5 to 18 percent replies against 1 to 3 percent for the generic version, which is not a copywriting result. it is what happens when the list is short enough to be true.

the condition is that ai was insured by accident. it sat quietly inside cyber and tech e&o policies that never mentioned it, the market called that silent ai, and in january the insurance services office published a standard endorsement removing bodily injury, property damage and advertising injury attributable to generative ai from commercial general liability. absolute ai exclusions started appearing in d&o, employment practices and fiduciary lines at the same time.

the cost falls hardest on companies that do not think of themselves as ai companies, because the exposure is an employee pasting a contract into a chatbot, a vendor tool screening applicants, or marketing copy that borrowed something. each of those gets pleaded as a different kind of claim, so four policies can each push it out through a different clause. most of the erosion is not a line on the declarations page either. it arrives as a revised base form or a slightly narrower definition, and it surfaces at the claim, which is the worst possible moment to learn what your policy means.

the fake fix is asking your broker whether you are covered for ai, which produces a reassuring yes, because coverage turns on how a plaintiff characterises the claim rather than on whether the word appears anywhere. do the version that holds. before the next renewal, put it to the broker in writing: which of our policies now contain ai exclusions or endorsements, what are the form numbers, and what is carved back in. then send one page listing where ai actually touches the operation, hiring, customer messages, contracts, code, marketing, so the answer comes back about your exposure instead of the category.

the condition is a purchase made on the strength of a demo, where the case for the tool is that it is obviously going to help, and obviously is carrying the whole argument.

the cost shows up two quarters later as four subscriptions, a shared login, and no way to say whether any of it worked. that is the same reason 61 percent of these projects were approved on a return nobody ever went back and measured. nothing failed loudly. it just never got checked, and an unchecked thing renews on schedule.

the fake fix is a heavier process: a scoring matrix, a pilot committee, a vendor bake off that runs six weeks and produces a slide, all of it expensive and none of it the variable in the data. what moves the number is one line written before money moves, naming which figure this is supposed to change, what that figure reads today, and the date you will look at it again. hours in the approval queue, cost per ticket, days to invoice, pick the one your business actually feels. a tool with a number attached either earns its renewal or gets cancelled in a quarter, and both of those are worth more than what you have now, which is a feeling and a card on file.

the condition is that ai never entered the business as a decision. one person expensed a writing tool, ops added a meeting notetaker, marketing bought a second one that does the same job under a different name, and the average company now carries 4.5 paid ai tools inside 27 ai touched apps, roughly a fifth of the entire software portfolio.

the cost is bigger than the line items. two thirds of it leaders were billed more than they expected this year off consumption pricing, average monthly ai spend is running about $85,500 and climbing 36 percent, and no single person owns the number that produced it. the consolidation everyone claims to be doing has also quietly reversed: apps per company are back up 11 percent, and the rate at which companies actually cut tools fell from 14 percent to 5.

the fake fix is a software rationalisation initiative, which means a spreadsheet, a steering group, and a decision that lands next quarter after the renewals have already run. do the ten minute version this afternoon. open the card statement, filter for anything with ai in the name, and write beside each one what it does in four words. any line that shares a job with another line gets cancelled this week, and anything that survives gets a hard usage cap before the card goes back on it.