AI Credits are easy to burn through quickly if you don't plan for it - a chatty Vippy user can blow a starter pool in a week. This guide covers the budget patterns we've seen work, the alerts that catch problems early, and the four ways tenants tend to over-spend.
The model: org pool + per-user allocations
- Org pool - the total monthly AI Credits available to your tenant (from your plan + any add-on packs)
- Per-user allocations - sub-pools assigned to individual users and apps so no single user drains the pool
- When a user runs an agent, the platform deducts from both the personal allocation and the org pool. Either running dry pauses new work until a top-up or a plan upgrade
Setting realistic allocations
- Heavy users (finance team running document processing all day) - allocate generously, expect their share to dominate
- Light users (sales using Vippy occasionally) - a smaller monthly allowance is usually fine
- Pilot users testing a new agent - allocate small until the agent is proven; raise after
- Reserve about 10% of the org pool unallocated, as a buffer for ad-hoc heavy use
Alerts that catch problems early
- Low balance alert - fires when the org pool drops below a threshold you set (default 10%), giving you time to top up before anything pauses
- Per-user allocation low - fires per user when their personal allowance drops below a configurable threshold
- Per-app spike alert - daily report on any app where AI Credit usage is more than 2x its 7-day moving average
- All alerts land in-app; admins can also route them to Slack, Teams, or email for immediate visibility
Four cost traps to avoid
- Trap 1 - Defaulting every agent to the most expensive model. Most tasks don't need the flagship tier; a mid-tier model handles ~80% of agent work at a fraction of the cost. Try a fast/cheap model for classification and routing
- Trap 2 - Unbounded loops in pipelines. A for-each over a long list with an agent step inside burns credits fast. Cap the max iterations on every loop and set a per-iteration timeout
- Trap 3 - Re-asking the same question. If a Vippy user asks the same thing three times because they didn't see the first answer, that's three billed calls. Improve the answer rendering instead of accepting the repeat cost
- Trap 4 - RAG on very large knowledge bases. Every retrieval query costs a small amount plus the AI Credits for the model response that follows. Chunk your content sensibly and use multi-stage retrieval to keep query cost predictable
Add-on AI Credits packs are forever - they don't expire and roll over month to month. Plan credits reset on every billing cycle. So if you're trending over your plan tier, two paths: buy an add-on pack (one-shot but permanent), or upgrade the plan tier (recurring). Add-on packs work well when the over-spend is one-time / seasonal; plan upgrade when it's the new normal.
Auto top-up
Admin Settings > Billing > Auto Top-Up. Configure a trigger ("org pool drops below 5%") and the auto-purchase pack size. The platform fires the purchase via Stripe and lands the credits within 60 seconds. Useful for tenants with critical 24/7 workflows that can't afford a hard stop.