In the second session of FinOps Summer Camp, Lukáš Pokorný (Engineering Manager) and Martin Cangár (Senior Cloud Operations Engineer) from Make.com shared how they built AI cost observability across their engineering organization using Vantage. Joining them was Vantage CEO Ben Schaechter, who added perspective on the patterns emerging across the broader industry.

For Make.com, the challenge arrived quickly. As the company's internal AI usage accelerated, AI spend grew 8.7x in the first half of the year. Spreadsheets and provider-specific dashboards were no longer enough. The team needed a consistent way to understand where AI costs were coming from, who was driving them, and how to respond before costs got out of hand.

Here are four lessons from their implementation.

1. Start by consolidating every AI bill

The first step wasn't optimization. It was visibility.

Make.com connected Anthropic, OpenAI, Cursor, GitHub Copilot, and AWS Bedrock into Vantage to create a single view of AI spending. Native integrations eliminated the need to build and maintain custom pipelines for each provider.

AWS Bedrock proved especially important. Because Bedrock usage is embedded inside an AWS invoice, it's easy for AI costs to disappear alongside infrastructure spend. Breaking it into its own cost view and budget made one of the company's fastest-growing expenses visible.

You can't manage AI costs until you can see them all in one place.

2. Standardize attribution across providers

Every AI provider exposes usage data differently.

Rather than reporting against five different schemas, Make.com created a single virtual tag, AI User, that maps each provider's user field into one consistent dimension.

That gives engineers a simple way to understand their own AI usage while allowing leadership to roll costs up by team, department, or organization.

As Ben noted during the session, consistent attribution is what makes budgeting, governance, and chargeback possible as AI adoption grows.

3. Focus on behavior, not just spend

One of Make.com's most useful metrics isn't cost at all.

Cursor exposes aborted and errored requests: instances where an engineer cancels a generation because the output isn't useful. The team tracks these as a signal of AI adoption quality. A low rate suggests engineers are using AI effectively, while a higher rate points to opportunities for training and better prompting.

The team also emphasized the importance of timely data.

Near real-time cost updates allowed Make.com to catch a Bedrock workload that increased four to five times overnight after a feature launch. Instead of discovering the spike weeks later during invoice review, they were able to investigate immediately.

4. Put cost data where engineers already work

Dashboards only help if people use them.

Rather than asking developers to check another reporting tool, Make.com surfaces individual AI budgets directly inside Backstage, the internal developer portal engineers already use every day.

The company complements that visibility with internal AI education and plans to manage budgets by individual AI tool instead of a single organization-wide budget, recognizing that different models serve different workloads.

The goal isn't to slow AI adoption. It's to help engineers make informed decisions about how they use it.

Three takeaways for engineering teams

If you're building AI cost observability today, Make.com's advice is straightforward:

• Centralize AI cost data before trying to optimize it.
• Standardize attribution early, even if providers expose different metadata.
• Enable alerts immediately so unexpected spend is caught before month-end.

AI cost observability isn't just about tracking spend. It's about giving engineers and leadership the information they need to make better decisions as AI usage scales.

What's next

The next FinOps Summer Camp session shifts from observability to control.

Join Rem Baumann, Director of Product at Vantage, and Ben Schaechter for Managing AI Budgets in the Age of Tokenmaxxing, where they'll explore practical strategies for setting AI budgets that encourage responsible usage without slowing development. Thursday, August 6 at 1pm ET. Register for the rest of the series here.