Your AI Budget Just Became a Line Item.

For a while the pitch was: use AI as much as you can, more is better, don't worry about the meter. That's over. AI usage is now a cost line that leaders have to watch, defend, and govern. Let me catch you up on what changed and what to do about it before your CFO does it for you.

The perk had a meter the whole time

The fun part happened fast. Companies pushed staff to use AI, built usage leaderboards, and in one case warned people they'd risk losing promotions if they didn't use AI. Then the bills landed. Accenture is now trying to stop staff from burning token reserves on basic work like turning PDFs into slides.

One line from an internal Accenture meeting sums up the shift: "We're hitting this inflection point where AI is becoming material to the cost structure." Their agentic AI lead, Justice Kwak, added that spend is unpredictable and CFOs and CIOs are still asking if they're getting value. Translation for you: the free-for-all is being replaced by rationing, and someone will decide the rules. Better it's you.

Cheaper model, worse experience, and no way to know

The obvious move to cut cost is swap in a cheaper model. Amplitude tested exactly that, weighing whether to replace Claude Sonnet with Google's Gemini Flash at roughly a fifth of the price. On paper it was a clean win.

In production it got messy. Session cost dropped from $4.88 to $2.33 per active chat user, and conversion held flat. But response time jumped from 63.8 seconds to 119.7, and users on the cheaper model sent 10% fewer messages. The answers were just as good. People just waited longer and engaged less. They held off on the swap.

The point for your team: price and quality are not the whole story. Latency and engagement are part of the cost of a cheaper model, and you can't see them in a vendor benchmark. You have to test on your own traffic, watch what users do next, and decide with real numbers, not a spec sheet.

Build the meter before you need it

When the bill matters, you need to see it in dollars per session, not guess. Amplitude's first FinOps engineer, Hac Phan, faced the classic build-vs-buy call and chose to build the tooling in-house using AI coding assistants. He wired up cost anomaly detection, reservation analysis, and a Slack bot that answers cost questions on demand. His estimate: it freed up 50% of his time.

You don't need to copy his stack. The lesson is that the cost work used to be too slow and expensive to do in-house, and that math changed. What took days now takes under an hour. Small teams can now instrument their own AI spend instead of waiting for a platform purchase.

So the question for your next planning cycle is simple. Can you say, right now, what a single agent session costs you? If not, that's the first thing to fix.

The reader you forgot to design for

Here's a cost driver hiding in plain sight. When Amplitude rebuilt its docs, LLM crawlers requested more pages than humans in the first 18 days after launch: 198,843 bot page views against 124,044 from humans. Agents are now a real audience for your content, and they cost you tokens and bandwidth whether you planned for it or not.

They built for it on purpose, shipping a read-only MCP server and raw Markdown endpoints so agents fetch clean source instead of scraping HTML. They also log every bot hit as an event. That's the move: treat agent traffic as something you can see and shape, not a surprise on the bill.

The deep cut

The trap is cutting cost blind. The temptation under pressure is to grab the cheaper model or slash usage across the board, then find out later what it did to conversion and retention. Amplitude's swap looked like free money and would have cost them engagement, and they only knew because they measured session cost and downstream behavior on the same user.

Before your next review, get one number in hand: cost per session tied to what the user did next. That single view lets you defend spend that pays off and cut spend that doesn't, instead of guessing. Rationing is coming either way. The teams that measure first get to make the call. The ones that don't get the call made for them by finance.

Three questions for your team

1. Can we state the dollar cost of one agent session today, and can we tie it to whether that user converted or came back? If not, who owns building that this quarter?
2. Before we swap to a cheaper model to save money, how will we test it on real traffic and watch latency and engagement, not just price and quality?
3. Do we know how much of our AI and content cost is agent traffic versus human, and are we shaping it on purpose or paying for it by accident?