Will an AI agent understand your email?
Paste an email and what you want the recipient to do. An AI agent reads it blind — three independent times — and we grade what survived against what you meant. A context-free simulation of machine readability, not a prediction of any specific inbox assistant.
Core ask survived, but the specific deliverable was lost.
Snag: "touch base on the thing" → Name the deliverable: 'approve the Q3 budget draft'.
| ||
The agent logged no concrete action — 'circle back' read as optional.
Snag: "maybe circle back when you can" → State the action directly: 'Please reply with approval by Thu'.
| ||
You needed this urgently; the agent filed it as low.
Snag: "no rush at all" → Drop the softener if it's actually time-sensitive.
| ||
'Next week' left no concrete date for the agent to schedule.
Snag: "sometime next week" → Use an explicit date: 'by Thursday, Aug 14'.
|
A context-free simulation: one agent, no thread history, no recipient data. Not a prediction of what a specific inbox assistant will do.
We email a confirmation link first (double opt-in) and run the score only after you confirm. Fair-use limit: two scores per email address per day. See Privacy.
Scoring one email by hand is the free version. Amino can check every outbound message — intent, action, deadline and injection risk — before it leaves. See how →
Agent Comprehension Score FAQ
What is an Agent Comprehension Score?
More inboxes now use AI to summarize, prioritize and suggest actions before — or alongside — a human reader: Gmail, Outlook Copilot and Apple Mail all ship features like this. The ACS is a context-free simulation that tests whether your email's summary, action, urgency and deadline survive machine extraction. It grades the message, not the mailbox.
Does this predict what Gmail, Copilot or Apple Intelligence will do with my email?
No, and we won't claim it does. Those assistants use different models, carry the prior thread, know the recipient, and apply their own policies — none of which we have. We run one context-free reader with no thread history and no recipient data. What that reader loses, a real assistant may well also lose; what it keeps is no guarantee. Treat the score as a machine-readability check, not a forecast.
How does it work?
You paste your email plus what you want the recipient to do. An AI agent reads the email BLIND — without your stated intent — and reports what it understood and would do. We repeat that blind read three independent times, grade each one against your intent in a separate call, and email you the median score plus how much the runs disagreed, with the exact phrase behind each miss and a one-line fix.
Why a band and a stability rating instead of just a number?
Because a single model run isn't a measurement. The same email can score differently across runs, model versions and prompt wording, so a bare '62/100' implies precision we can't support. We report the median of three runs, a band (Clear / Usable with ambiguity / Likely misread / Action unclear), and a stability rating. If three runs disagreed about your deadline, that disagreement is the finding — the ambiguity is in the email.
What are the dimensions?
Four make up your Message actionability score: Summary fidelity (did your core point survive), Action clarity (did it get the single ask), Priority match (did it triage urgency as you meant), and Deadline & context sufficiency (does the message alone pin down when). Confidence calibration is reported separately as an agent-behavior diagnostic and deliberately does NOT feed the score — it describes how sure the simulated agent was, not how well you wrote, and averaging it in used to reward vague emails for making the agent appropriately unsure.
Which model grades my email, and can I reproduce a score?
Every scorecard names the exact model that produced it (currently @cf/meta/llama-4-scout-17b-16e-instruct on Cloudflare Workers AI, with a smaller Llama fallback if it times out), the rubric version, the number of runs and the date. When we change the rubric we bump its version, so an old score stays interpretable.
Do I have to say what I want the recipient to do?
It's optional but it unlocks the graded score. Without it you still get the blind read plus machine-legibility and agent-instruction flags — useful on its own. With it you get the four-dimension comparison and your single highest-leverage fix.
Do you store my email?
Your pasted subject, body and intent are transmitted to us and held only until you click the confirmation link, and are deleted as soon as we've scored them — within 24 hours at the latest, even if you never confirm. Scoring runs on Cloudflare Workers AI, which does not use your content to train models. We then keep only your email address, briefly, to deliver the result and honor unsubscribes. Please don't paste secrets, credentials, regulated data or confidential threads. See our Privacy page.
Why a confirmation link instead of scoring instantly?
Two reasons: it stops the tool being used to score forged addresses, and it means we only run the model after you opt in. Fair-use limit is two scores per email address per day.
Isn't this just a spam or deliverability check?
No. Our /audit and /headers tools check whether your email arrives and authenticates. The ACS checks something different — whether an automated reader between you and your recipient extracts the right summary, action, urgency and deadline. Machine-actionability is a newer axis, and it's the one no deliverability tool measures.