Agent Readiness Audit
Paste the instructions, FAQ or policy text you gave your AI agent. You get back the five holes it is carrying into production, ranked by what will hurt first — and the exact wording to close each one.
Your text is sent to the model to be reviewed and is not stored by us. Don't paste customer names, card numbers or anything confidential — replace them with placeholders first.
What it checks
Five classes of failure. These are the ones that actually reach a paying customer, in the order they usually hurt.
Is there a quotable file?
The audit looks for a plain-text source the agent can lift sentences from verbatim. If the facts live inside prose instructions, the agent will paraphrase — and paraphrasing numbers is inventing them.
"Quote only what's listed"
Missing this one sentence is the single most common cause of invented prices and policies. The audit tells you exactly where to put it.
Does it stop before answering?
Most setups escalate after the agent has already answered. Trigger lists only work if they tell the agent this class of question isn't theirs to answer.
Does the human start blind?
Checks for the five fields a human needs — contact, summary, what was tried, what's open, suggested reply — so the customer never repeats themselves.
Has anything tried to break it?
Friendly tests pass on every prompt. The audit gives you the adversarial message set to run before launch.
A fix list, not a score
You get the ranked gaps, the reason each one matters, and copy-paste replacement text for your instructions. No percentage score to feel good about.
Want the fill-in versions?
The audit gives you the method for free. If you'd rather edit than invent, the facts-file templates, the 20-message hostile test suite and the escalation rules are available as ready-to-use packs.
See the packs →