Problem solved

AI agents make spoken commitments that are not recorded, leading to broken promises discovered only when a client forwards a transcript. UnfakeIT provides deterministic monitoring outside the model to catch moved ETAs, invented follow-ups, dropped details, and non-answers, re-voices only the offending turn, and logs all commitments in a ledger on your own data, so you have proof before the client does.

Use cases

• Catching when a voice agent changes an ETA mid-call (e.g., 30 min to 90 min)
• Detecting unauthorized follow-up promises like 'the owner will call you back'
• Auditing that delivery windows from agent answers survive re-voicing
• Identifying high-stakes questions answered with generic non-committal reassurances
• Maintaining a per-call commitment ledger for compliance or client review
• Adding human-sounding ownership of misses without rewriting every turn

Audience

• Developers
• ML Engineers
• Business Users

Pros

• Lean mode adds no latency, model calls, or token cost on clean turns
• Deterministic checks run outside the model, independent of prompt reliability
• Logs on your own data behind your own auth
• One-time license fee of $499
• Surgical re-voicing preserves the agent's original wording when fixing a turn

Cons

• Requires technical integration with an existing agent stack
• Focused narrowly on commitment tracking, not general agent observability
• Full re-voicing mode is opt-in and can add latency/token cost
• Only detected commitment categories are shown; custom types may need extension
• No public documentation or API details provided on homepage