How much of your AI is defensible?

Total monthly spend on AI services across all vendors. Include token billing, cloud AI services, and model hosting or inference fees.

Exclude general cloud infrastructure, internal salaries, and one-time training costs if you’re measuring production usage.

An estimate is fine — round to the nearest $10K.

Discrete AI outputs delivered this month: user-facing responses for chat, answered queries for RAG, completed tasks for agentic workflows.

Count finished outputs delivered to users or downstream systems, not every internal model call within a task.

Don’t know? Your vendor dashboards usually surface this. A directional estimate is fine.

The dominant type of AI workflow in your production environment.

Chat: simple conversational AI — user sends a message, model replies. One model call per turn, no document retrieval, no tool use.

RAG: Retrieval-Augmented Generation. The model is paired with a document store or search index; when a query comes in, relevant content is retrieved first, then the model answers using it.

Agentic: multi-step workflows where the model plans, calls tools, retrieves data, and chains reasoning to complete a task autonomously.

Mixed: most enterprises. Pick this if you have more than one of the above in production.

Whether your team can identify the documents, training data, or other inputs that influenced each AI output.

We can trace every output to source: full provenance is in place. For any output, your team can show which specific inputs shaped it and click through to inspect them.

Some outputs: source attribution exists for some workflows — often high-stakes or regulated use cases — but not across the board.

No: outputs are not currently traceable to the sources that influenced them. The model produces an answer; the inputs that shaped it are not exposed.

Whether each AI output ships with a confidence value your team can act on.

Yes: every output carries a score (or equivalent signal) indicating how reliable the model judged its own answer. Your team can use the score to accept, route to review, or reject the output.

No: outputs ship without confidence values. The answer comes through; whether to trust it is left to whoever reads it.