Phishing and scam emails your agent should never answer
You built an agent that reads the inbox and answers it. It drafts replies, books meetings, sends files, maybe even updates a record or approves an invoice. That is the whole point. The trouble is that the inbox is the one place where strangers are allowed to write to your agent, and some of those strangers are writing specifically for it.
A human assistant who gets a strange message usually pauses. An agent built to be helpful tends to do the opposite. It reads the email, finds the request, and tries to complete it. When the email is a phishing or scam attempt, a helpful agent becomes the attacker's best employee.
Here is what goes wrong when an email agent answers the wrong message, why homemade filtering is hard to keep up, and where a dedicated check fits.
What a reply to the wrong email costs
Most teams worry about the obvious failure, an agent clicking a bad link. The quieter failures are more common and often more expensive.
A reply confirms you exist. Even a polite "Thanks, we'll take a look" tells the sender that the address is live, that someone (or something) reads it, and that it answers. That address now gets more mail, and better targeted mail.
A reply can carry data out the door. Agents are often given context so they can be useful: a calendar, a customer list, order history, a shared drive. A scam email that asks for "the latest invoice" or "the contact sheet for the event" is asking your agent to share what you gave it. Picture a fictional outdoor brand, Copperleaf Outfitters, whose support agent cheerfully sends an order export to someone posing as its shipping partner. Nobody hacked anything. The agent simply did its job for the wrong person.
A reply can start a money conversation you never meant to have. Fake invoices, changed bank details, "pay the deposit to secure the booking." An agent that negotiates, confirms or forwards these messages moves a scam one step closer to a payment, and it does so with your name on the thread.
A reply can be steered. Some emails are not written for a person at all. They contain instructions aimed at the AI that reads them, hoping the agent will treat the email as a command. An agent that obeys text from a stranger is no longer working for you.
A reply lends you to the scam. If your agent answers on behalf of a fictional studio like Tallgrass Media, the scammer can quote that answer to the next victim: "Their team already confirmed it."
None of these need a dramatic breach. They only need an agent that answers first and thinks later.
Why "just tell the model to be careful" is not enough
The first fix most developers reach for is a line in the system prompt asking the model to ignore suspicious mail. It helps a little. It does not hold up for long, for a few reasons.
A general model is trained to be helpful and to take the email at face value. Scam mail is written to look normal, polite and urgent in exactly the way that earns a helpful answer. The same model that writes your replies is now being asked to be the skeptic, the judge and the assistant in one pass, and the assistant usually wins.
It is also inconsistent. The same email can be read as fine on one run and suspicious on the next. That is tolerable for drafting a friendly reply. It is not tolerable for a decision about whether to send data or money.
It also gets expensive at scale. Running a large model with long instructions over every inbound email adds up fast.
Why homemade filters fall behind
The next fix is a homemade filter: a list of things to block, a few rules, maybe a second prompt. This is where many teams spend far more time than they planned.
Scams change faster than your rules. Each new pattern needs someone to notice it, understand it and update the filter. The people writing scam mail change their wording as soon as it stops landing.
False positives cost real business. Block too much and the agent ignores a genuine customer, a real partner or a legitimate offer. Those misses rarely show up in a dashboard. They show up as a lost deal weeks later.
Edge cases pile up. Real senders sometimes look odd, and careful scammers sometimes look perfect. Every exception you add makes the filter harder to reason about, and every change risks breaking something that used to work.
Nobody owns it. A filter written during launch week tends to sit untouched while the threats move on. Keeping it current is ongoing work that competes with everything else you are building.
The honest summary: screening email well is its own specialty. If it is not the product you are selling, it is a cost you carry.
A check before your agent acts
This is the gap Ready Check is built for. Your agent sends one inbound email's text to Ready Check before it decides what to do, and gets back structured findings: the kind of email, a verdict, scores, the named checks that fired, and what to watch for. Phishing and scam signs are flagged, and so are emails that try to give instructions to an AI assistant.
A few things it deliberately does not do:
- It does not write a reply. Your own agent, or a person, decides what happens next and writes in your own style.
- It does not connect to your inbox. It only sees the email text you send it, plus attachment file names if you include them. Files themselves are not downloaded or opened.
- It is not a guarantee. The checks are a sorting aid. A careful scammer can pass every check, and a real sender can trip one, so read the result before you act, especially on money, sign-ins and attachments.
What it gives you is a separate, structured opinion from a step built only for screening, not from the same model that is trying to be helpful. Your agent stays in charge of the conversation. Ready Check answers one question first: what is this email, really?
What it costs
A standard check is $0.0025. Longer emails may count as two checks. You buy a prepaid credit pack, reuse the same key, and the amount charged is returned with every result, so the cost per email is never a mystery. Part of each check is a model step run through OpenAI, and that model cost is already covered by the flat price.
For most agents, that is a small line item next to the cost of one bad reply.
Where to find it
- The facts page, with exactly what is checked, what is returned and the limits: Ready Check
- The MCP server for agents and coding assistants: ready-check-mcp on npm and ready-check-mcp on GitHub
- If you are wiring it in by hand: How an AI agent can call Ready Check to triage inbound email
Your agent was built to answer email. Give it a way to know which emails should not get an answer. Are you ready?
Copperleaf Outfitters and Tallgrass Media are fictional examples.
ReadyUpAgents is independent and not affiliated with or endorsed by OpenAI, Stripe or Cloudflare. Product and company names belong to their owners. Company names in the examples are fictional.
