What the session covers
The session opens with a familiar scene: the AI bill doubled last quarter, finance wants to know why, and engineering cannot say which team, model or prompt caused it. Deeps explains that the cost unit of AI is the token, which is invisible in the way a server or a licence is not, and that agents multiply token use because one request becomes many tool calls, retries and reasoning steps. He then walks through the mechanics: input and output tokens priced per million, the hidden cost of system prompts, history, retrieved documents and tool output, and the way a growing context window makes each turn of a conversation more expensive than the last. The second half sets out five levers (smart model routing, prompt caching, per-user and per-team budgets, context discipline and agentic circuit breakers), argues that they only work when applied centrally at a gateway, and shows how Konsole implements them with a base-URL swap, bring-your-own-keys held in hardware vaults, PII masking and audit logs. A Q&A covers the biggest blind spot (observability), governance before cost cutting, enterprise and custom model integration, Konsole’s own security posture, and SaaS versus on-premise deployment.
What you’ll learn
- Why token-based pricing makes AI cost hard to see, attribute and control compared with cloud compute.
- What actually drives a bill: system prompts, conversation history, RAG chunks, tool outputs and long examples, not the user’s message.
- How context-window bloat makes turn ten of a conversation cost far more than turn one, and what to do about it.
- Why agents change the risk profile: loops, tool calls without end states, and a four-figure bill before the next stand-up.
- The five levers of AI cost governance and how much each saves, including routing to models that are 20 to 60 times cheaper for simple tasks and caching that cuts repeated-token cost by up to 90%.
- Why budgets, alerts at 50/75/90% and hard limits should live in a gateway that every app passes through, not in individual apps.
- How Konsole applies all five levers with a one-line base-URL change, plus PII masking, BYOK and audit logging.
In their words
“If you cannot answer who is spending what, on which model, for which task, you are not managing the cost. You are just receiving the bill afterwards.”
“Tokens are the new cloud compute, except they are much harder to see.”
“A runaway agent does not wait for the next monthly billing cycle. It will create a four-figure bill before the next stand-up.”
“Every production agent should have a budget, a boundary and a kill switch, and the kill switch should be with a human.”
“You cannot control AI spend only after the bill arrives.”
Key takeaways
- Measure first. Attribute every call to a user, team, app, model and model version before trying to cut cost.
- Route by task, not by habit. Classification, extraction and yes/no decisions belong on small models.
- Cache repeated context at the gateway so every app benefits without each team configuring it.
- Give every agent a token budget, a tool-call ceiling, loop detection and a human-held kill switch.
- Establish governance before cost reduction. Cost reduction follows from visibility.
More sessions

Clip · 2 min
Runaway agents and why you need circuit breakers
A two-minute excerpt on how agentic AI changes the cost risk profile: loops, tool calls with no end state, and why autonomy needs boundaries.

Conference talk
Emotional AI and voice forensics: DMWF Global 2025
Sheoli Biswas on the Emotional AI for Marketing panel at DMWF Global 2025 in London: what sentiment and voice analysis can read, where consumer trust breaks, and the controls that keep emotional data from becoming a liability.