AI is answering from years of contradictory reports, and $210 of agents opened five accounts

Take the phrase active customer. In the sales dashboard it counts anyone who bought something in the past twelve months. In the finance report it counts anyone with an open contract. In a spreadsheet a departed analyst built in 2019, it counts both, and a few of them twice. Someone who knew the company knew which of the three to trust, and which had been abandoned years ago.

Most companies are now putting an AI answer layer on top of exactly that, according to a survey of senior data leaders: years of unused, contradictory reports, left as they are. An assistant asked a question in ordinary language does not rank its sources. It finds a table, reads what is in it, and writes one clean sentence in seconds.

The analytics firm Omni, writing up a 2026 industry survey, describes the most common form of this: the same measure defined in more than one system, with the analytics team, the dashboards and the AI copilot each keeping their own version, so a change to one definition has to be made three times.

Cloudera's report this year, drawn from 1,270 IT leaders at companies with more than 1,000 staff, has 96% saying AI is in their core business processes and 85% saying they have a clear data strategy, while about 80% say they still cannot reach all of their own data across the systems holding it.

A Strategy survey of 100 senior data and technology leaders at companies above $2 billion in revenue found that nearly 80% of their data teams spend more than half their time preparing data rather than answering questions with it.

Five hours of rented computing cost $210, and by the end of it five of Shrivu Shankar's own online accounts were open. Shankar rented about 100 unrestricted open-source AI programs on machines fenced off from his home network, and told them to research his name and get into anything belonging to him. They also produced his home address and phone number, starting from the name alone.

Three of the five accounts they opened through known weaknesses in software. They found no new flaw in anybody else's product, and never reached what Shankar calls his tier 0 accounts: his Gmail, his password manager, his bank. His own arithmetic came to $40 an account, and he expects a comparable result for under $5 an account within a year as the same models get cheaper to run.

Cyber officials from the United States, the United Kingdom, Canada and New Zealand told business leaders this week that basic security habits, not AI tools, still stop most attacks, even as state-backed hackers use AI deepfakes to get through job interviews and be hired into remote roles with company access.

That is the same guidance the FBI, CISA and NSA gave on deepfakes in 2024: confirm who a person is through a separate, agreed channel, rather than relying on anyone to notice the fake.

Google is testing a mode in the Gemini desktop app called Assign. A person hands it one folder on their own computer, and the model works through the apps on screen, opening files, filling things in, moving between programs. OpenAI's Codex and Anthropic's Claude agents do the same thing now.

The same design is arriving where the work is discussed. An assistant inside Slack or Microsoft Teams takes the request in the channel and completes it, instead of returning something for a person to carry into another system by hand. More than 2,000 businesses pay for one of them, sold by a Warsaw company called Viktor that has raised $75 million.

Each of these needs standing access to do the work. In Slack or Teams the app is installed once, and the channels, files and connected systems it may read are chosen at that moment; it keeps them until somebody takes them away. Google's Assign narrows the same decision to a single folder on one computer, chosen by the person handing it over.

A rough drawing is now an accepted instruction. OpenAI's ChatGPT Images 2.5 adds a Sketch tool that takes a hand-drawn layout, boxes, arrows, a scribble where the product goes, and returns a finished image, in about half the generation time of the version before it. OpenAI released it to keep pace with Google's rival image tool.

For a business with nobody in a design role, that covers the slide, the product concept and the social post. The sketch on the back of a printed agenda goes to the subscription somebody in marketing pays for.

Technology chiefs are increasingly blamed when a workplace AI tool produces a bad answer, according to new industry survey data. In most of those cases the same person did not choose the vendor, did not approve the rollout, and is never shown how the system arrived at the answer being complained about.

The rollouts they answer for tend to begin in a department that switches on an AI feature inside software the company already pays for, which needs no purchase order and no launch date.