AI builds budget models in minutes, and one Microsoft outage stopped email, Copilot and Defender

Pigment, a planning software company, offers a free 14-day trial in which you describe the budget you want to an AI agent. It asks a few questions and builds a working model within minutes, and you can load your own figures and compare scenarios.

The slow part of a budget has been its structure: which lines exist, how headcount drives payroll, how volume drives fuel. When finance needs weeks to build that, a department head gets to ask two or three what-if questions a year. When it takes minutes, nobody needs to ration the questions.

The weak point moves into the assumptions. To build that fast, the agent fills every gap with something reasonable, such as a pay rise in January or customers paying in 30 days, and a model with a number on every line does not show which ones were guessed.

So test it on a year that has already happened. For example, a logistics manager could describe last year as it looked in January, with the volumes, drivers and fuel price she expected then. If the model spreads overtime evenly and her real December cost twice a normal month, she has found the assumption that would have broken next year's budget.

Budget rounds will soon bring thirty versions instead of three, and the argument will move from the totals to the assumptions underneath. The depot manager who knows why December is different becomes the most useful person in the room.

Copilot works by reading the mail, calendar and files in the same suite, so the outage cut it off from them as well. Defender was hit too, and its check on links that staff click in Outlook and Teams could fail, in the hours when an email promising a fix is most tempting.

Every job that a team moves into Copilot, such as drafting customer replies, stops at the same moment as email. So a continuity plan needs a line for each of those jobs: who can still do it by hand, and how they reach customers outside Microsoft.

Yet the cost of support per booking fell just over 17% in a year. Part of the gap is which cases the AI takes. A question about check-in time is cheap to answer, so the AI takes those first. A dispute between a guest and a host over damage is slow and expensive, and it stays with people.

So the human queue keeps the hard cases, and the average case in it takes longer than before, even while the team works well. A manager who still judges agents on time per case will punish them for the AI's success.

Fireworks AI, the supplier, runs AI models for Notion, and attackers broke into it. The chats came from fewer than 1% of workspaces, and most could not be linked to a person or a company.

A promise to keep nothing fits a simple request: the model answers and forgets. Insights, a Notion dashboard that shows administrators what their teams use AI for, needs the old conversations, so Notion sent stored chats to Fireworks to be sorted. Fireworks had promised to keep none of them, and copies were still there when the attackers arrived.

Any feature that looks back at past chats, such as a usage dashboard, needs a stored copy, and that copy goes wherever the feature's work is done. Before switching one on, ask where the copy is kept, which suppliers handle it and when it is deleted.

A Team Bot is one Grok Bot that a whole team shares in Slack or Microsoft Teams. When a teammate teaches it something for the team, the lesson waits under Review learnings until the owner or a Manager approves it.

That approval decides how the team's work gets done. If a new hire teaches the bot that credit notes under €500 need no sign-off, and the reviewer clicks Approve without checking, the bot applies that rule for everyone. The reviewer should be the person who knows the work, not whoever set the bot up.

The guide is voluntary and open for public comment. It suggests passing on the gains through higher pay or shorter hours, and spending some of the time saved on teaching and mentoring.

Without a decision, freed hours fill with more work, and when AI takes the routine tasks, the work that fills them is the hard kind. Mentoring deserves the most thought, because the routine tasks AI takes first are the ones that juniors used to learn on.