Managing AI Costs: The dkd AI Usage and Cost Dashboard

How can AI costs be transparently tracked, analyzed, and managed within a company? With the dkd AI Usage and Cost Dashboard, we provide visibility into the use of generative AI via LiteLLM—from individual budgets and AI models used to project-related costs. This article shows how centralized AI cost management helps plan budgets, allocate costs based on usage, and optimize AI usage within the company using data-driven insights.

Now that we’ve covered the basics in the blog post “Securing AI: The Reliable, Privacy-Compliant LiteLLM Platform,” this post will focus on specific examples of AI cost management.

Making AI Costs Transparent in the Company

The question eventually came up on its own. Many people at dkd now work with AI—via chat.dkd.de, in Claude Code, and in OpenCode. All queries are routed centrally through LiteLLM, which meant that all the answers were also stored there: in the database behind the LiteLLM backend. This was inaccessible to practically anyone, and anyone who wanted to know how much a project had cost in terms of AI or how close someone was to their budget had to ask someone with access—who would then query the API manually.

That’s not a basis on which to discuss budgets, prices, or model selection. That’s exactly why we now have the dkd AI Usage and Cost Dashboard: so everyone can see the same numbers without having to ask someone every time.

Development of an AI Cost Dashboard Using LiteLLM

The first version could do exactly one thing: display the cost per LiteLLM day. That was enough to roughly assign expenses to the right projects, and it was built in a single afternoon. That was all it was meant to be at first—the goal was to figure out whether it would be of any value to anyone at all.

Further questions then arose naturally: Which models are used, and how intensively? How much do AI costs amount to for a specific task (e.g., organized via a ticket)? How do I use this information to plan a budget? Subsequent versions expanded on this starting point accordingly.

From a technical standpoint, the dashboard is intentionally kept simple: it is an R Shiny application that runs in Docker containers on our ShinyProxy infrastructure and is accessible to all dkd employees via the intranet.

Features of the dkd AI Usage and Cost Dashboard

My Usage – Your Own AI Usage

The section that everyone at dkd sees. How much have I used in the last few weeks, which models am I actually using, and where will I end up at the end of the month if this continues?

• Usage broken down by area of use: API key and chat.dkd.de
• Trends over the past 30 days and the past six months
• Budget status per key, including a projection to the end of the budget period—with a warning if the projection exceeds the budget
• Top models and the proportion of cached input that significantly reduces costs
• “vs. others”: a percentile ranking of your own usage within the team. Completely anonymous; no individual values or other users' names are displayed.

Analyze AI Costs by Project, Ticket, and Tag

Every LiteLLM call can include tags, and anyone who tags their work can see afterward what AI costs were incurred by a ticket or a project. The Tags section organizes this by tag type: Issues, Projects, Credentials, and Everything Else.

• Tags containing a ticket URL are resolved in Redmine: instead of a bare URL, they display the ticket title, the associated project, and a direct link
• A filter can be set to display only your own tags, if desired
• The rollup view summarizes the data collected for each Redmine project

This is the foundation for fairly allocating AI costs among projects and customers in the future. However, it’s only as good as the tags that are sent. But that’s a story in itself—one for which we’ve developed our own plugins.

Analyzing AI Costs and Model Usage in the Enterprise

• Expenses for this month, last month, and the difference between them
• Daily Expenses by Provider
• Top models, sorted by cost or by tokens
• Top 10 Clients: Which Tools and Systems Actually Generate the Load
• Monthly trend over six months

AI Budgets and Cost Control for Management

Management gets its own tab: all defined budgets with their current status, broken down by personal and shared keys, along with each person's cost history and the tags used.

Privacy-Compliant AI Cost Management with LiteLLM

The dashboard does not store anything. It is stateless, has no database of its own, and requires no backup—all figures are pulled in real time from the LiteLLM API, whose data is backed up separately. The dashboard analyzes only the information that LiteLLM collects anyway: username, model, tokens, and costs. No prompts, no responses, no content.

Making AI Costs Transparent Within Your Own Company

Would you like to know how your company could implement an AI usage and cost dashboard like this? Feel free to contact us.

Request a dashboard for your business

Sources & Related Links

[1] AI Transformation? We're Your AI Agency
[2] The Path of dkd AI Transformation: On Adventurers, Stabilizers, and Balancers
[3] Prompt, please!
[4] Starfruit AI
[5] Securing AI: The Self-Sovereign, Privacy-Compliant LiteLLM Platform