Use casesAI products, agent platforms and copilots that act for users
Did a person approve what the agent does?
Before an agent's consequential action goes through, HEIP checks that a person in your web app approved it.
Request
Agent proposes a transfer
The user clicks Approve in your app
Agents act faster than oversight can respond.
Agents now click, submit and pay on users' behalf. Most actions are fine. The risky ones need a human checkpoint, and a checkpoint only works if the Approve click came from a person, not from the agent or a script.
Why now
- +1,700%
- growth in daily AI-agent requests in one year, on Cloudflare's network
- Source: Cloudflare, “The Internet has a second audience”, Sep 2026
- 49.8%
- of surveyed security professionals reported or suspected an AI agent took an unintended or unauthorised action
- Source: Salt Security, State of Agentic AI & API Security 2H 2026 (vendor survey, n > 380)
What you get
- A human checkpoint the agent can't click through itself
- Because HEIP checkswhether a person was in control of the session when Approve was pressed
- Approvals tied to one specific action
- Because HEIP checksthat the approval token was issued for this action and used once
- A record of what a person approved
- Because HEIP checksand keeps a decision record with reasons
- Agents that stay fast for low-risk work
- Because HEIP checksonly on the actions you choose to protect
Actions to protect
| Action | Suggested policy | On step up, you usually |
|---|---|---|
| Approve a payment | Critical | Passkey or re-authenticate |
| Approve deleting data or changing permissions | Critical | Re-authenticate |
| Approve a purchase | Fast | Ask the user to confirm again |
| Approve sending an email or message for the user | Standard | Show the message and ask again |
Policies are named presets you set per action. Critical asks for the most evidence. Fast suits frequent actions that still matter. Standard fits everyday actions. Observe never blocks.
Where it fits
HEIP sits in the approval step of your web app. It checks approvals made in a browser. It does not identify agents or authorise calls an agent makes directly to an API.
Works alongside
- Your app's approval screens
- Your existing sign-in and MFA
- Your agent's action log
- Your permission model
Roll it out on agent approvals
Baseline
Pick agent approvals. Note how they're protected today.
Observe
HEIP decides on real traffic but blocks no one.
Compare
Review its decisions against what you already know.
Enforce
Turn it on for agent approvals only if it earns it.
Most teams go from observe to enforce in 6 to 8 weeks. Observe mode blocks no one while you compare. How onboarding works
Questions
Does HEIP identify or authenticate agents?
No. It checks whether a person approved an action in your web app. It doesn't tell agents apart or vouch for them.
Does it work when an agent calls our API directly?
Not today. HEIP protects actions taken in a browser, so it fits the moment a person approves what the agent proposed.
Will it slow our agents down?
Only on the actions you protect. Everything else runs as it does now.
Can an approval be reused for a different action?
No. Each approval gets a token that works once, for that action in that session, and expires in seconds.
Trust has to move with the action.
When digital actions can be automated, trust must move with them.