What happened
A 404 Media investigation published August 7, and highlighted the same day by researcher Simon Willison, documented companies scrambling to rein in spending on AI tokens, the per-use billing unit behind most AI services. The report drew on leaked recordings of internal Accenture meetings, and the findings upend a common assumption: most token consumption at the consulting giant was driven not by engineers but by non-technical staff. One of the largest single drivers turned out to be mundane document handling, converting PDFs into images and then into markdown text so AI systems can read them, a process that quietly burns through tokens at scale. In the recording, one executive asks, turning PDFs into markdown, is that right, and a colleague confirms the data. The report lands the same week as coverage of Microsoft imposing internal token budgets on its own teams, routing work to cheaper models and treating tokens as a scarce resource, in part because a single user request can fan out into multiple model calls, searches, and retries.
Why it matters for your business
If a company with Accenture's resources is surprised by where its AI money goes, a small business should assume it does not know either. Flat-rate, per-seat AI subscriptions shield you from this, but the moment your tools run on API pricing, and many chatbots, automations, and agency-built integrations do, costs scale with usage that nobody is watching. The PDF detail is worth taking literally: the format of your documents changes what AI processing costs. Feeding systems clean text instead of scanned PDFs is cheaper as well as more accurate.
What to do about it
- Ask which of your AI tools bill by usage rather than per seat, and get the usage dashboard in front of someone monthly.
- Set spending alerts or hard caps on any AI API account your business pays for.
- Where documents feed AI workflows, prefer text formats over scanned PDFs, and ask vendors what drives their token consumption.