The AI allowance memo

When limits arrive

In July, 404 Media reported that companies were throttling employees’ AI use as costs rose. In at least one case, monthly AI spending had tripled to more than $15 million. [1]

Since then, two reported answers have been a number per person. In October, 404 Media reported that a global consultancy told many employees it will centrally fund only $29 a month per person of one AI assistant, with use above that charged to a project budget code. On October 6, Channel Insider, relaying Yahoo Finance's summary of The Information's reporting, said monthly AI spending caps for individual employees in Microsoft's Cloud and AI division had in most cases fallen from $100,000 to roughly $10,000. The report carries no confirmation from Microsoft. [2] [3]

That is a $29 funding level at one company and a roughly $10,000 cap at another. Each is a reasonable reaction to a bill that grew faster than anyone planned. A flat allowance controls the size of the bill. It cannot tell you which of the spending was worth it, and that is what a board asks next.

The reflex is easy to understand. In Deloitte's second-quarter survey of 200 North American CFOs at companies with $1 billion or more in revenue, 46% named cost uncertainty or lack of transparency as their biggest internal concern about AI. In KPMG's US survey of 204 leaders at companies with $1 billion or more in revenue, 26% reported full, real-time visibility into what their AI systems cost to operate, while 66% had monitoring dashboards. When the invoice jumps and the dashboard cannot explain it, a number per person is a lever finance can pull this quarter. [5] [6]

Where acceptance is recorded

Now take two people at the same firm, both past $29 by the tenth of the month. These figures are illustrative, not reported. One is an engineer whose coding agent works through a ticket backlog. Last month it cost $600 and produced 40 changes that reviewers merged and that stayed merged, so $15 a change. The other is an analyst who uses the assistant to turn reports into slides. That cost $45, and no system records whether the slides were any good.

For the engineer, $29 is wrong by a wide margin, and the evidence for a better number is already sitting in source control. For the analyst there is no evidence either way. A flat allowance treats them the same. They need different answers, and only one of the answers can be computed today.

Where work leaves a record of acceptance, the measure is cost per accepted result. A merged change that stayed merged, or a ticket closed that stayed closed. The acceptance has to come from a reviewer or a system the person does not control, and the check has to look for reversals as well as approvals.

Any ratio that becomes a target gets gamed. ITPro, citing a February Financial Times report, said in June that a global consultancy's monitoring of staff AI use prompted “tokenmaxxing”, with employees showing off their uptake by using more. Cost per merged change can be gamed too, by splitting work into smaller pieces. So the ratio should set the allowance for a role, reviewed by people who know the work. It should never become a ranking of individuals. [4]

Where evidence is missing

The analyst is the harder case, and probably the bigger one. The same ITPro report, citing internal company data, described non-engineers driving excessive use on tasks such as converting PDFs to slides. For slides, memos and analysis, most companies have no reliable record of accepted work. A reviewer's sign-off or a deliverable's status in a project tool might serve. Until one does, the honest allowance for those roles comes from seat use and cost, and gets reviewed every quarter. [4]

One part of the reported memo is sound. Cost above the allowance goes to the project it served. The next step is to attribute it from the record of each call, to the project and to whoever approved the overage, so the charge does not depend on a code someone remembers to enter.

An allowance memo written this way runs longer than a paragraph. Each role gets a number, the evidence behind it, and the name of whoever approves the exception. Engineering roles could have that next month. The rest will take longer, and the memo should say so.