Agents get one key. Customers hold the other.
Before an AI agent does anything risky in a customer’s account, the real customer okays that exact action with a passkey. The agent gets a time-locked session for that one job, and you keep a signed record of it.
Now taking a small number of design partners.
Try it: an agent asks to add a payee
2207Open for one job
- Agent
- Budget Buddy, by Orbit Labs
- May do
- Add payee J. Okafor, send once
- Up to
- $2,400.00
Locked
Okay this agent action?
Budget Buddy wants to add J. Okafor as a payee and send them $2,400.00.
- Agent
- Budget Buddy, Orbit Labs
- Signature
- Verified
- Request ends
- in 5 minutes
It can add J. Okafor and send $2,400.00, once, in the next 20 minutes. Nothing else.
- Limit
- $2,400.00, once
- Time lock
- 20:00 left
Entry record
signed, kept by you| Time | What happened | Record |
|---|---|---|
| Nothing yet. The agent is waiting at box 2207. | ||
You’re the customer here. Larkspur Bank, Budget Buddy and J. Okafor are fictional.
To your website, an agent in an account looks exactly like your customer.
Browser agents ride the customer’s logged-in session or type in their password. Cloud agents sign in from data centres. Either way your systems see the customer, so the agent inherits everything the account can do. Nothing it does can be limited, revoked or proved afterwards.
Hand over the master key
Let agents through and absorb the fraud, the account takeovers and the “my agent did it” disputes.
Brick up the door
Block agents, and turn away the customers who chose to use one.
HumanOkay is the third option: two keys, so the customer decides, one action at a time.
What the public figures say, 2026
| Figure | What it measures | When |
|---|---|---|
| 4.4× | Growth in AI-agent orders on one fraud network1 | Early Aug to late Sep |
| +38.6% | Fraud attempts in North America, while e-commerce sales grew 8.35%2 | First half |
| 84% | Bank fraud leaders who call agents their most exploitable weakness3 | June survey |
| +1,700% | Daily AI-agent requests seen by one network, over a year4 | To September |
“Customers must remain in control, understand what the agent is doing.”
How HumanOkay works
It sits between your bot management, which can see agents, and your login, which assumes a person is typing. The same four steps run every time an agent tries something risky, and each one writes to the customer’s entry record.
-
1
Recognise the agent
HumanOkay reads the signals your bot vendor already produces, plus the cryptographic signatures that well-behaved agents attach to their requests. Agents that don’t identify themselves are caught by how they behave, and get the same treatment.
- Agent
- Budget Buddy
- Identity
- Signed, verified
-
2
Ask the customer
Before a risky action, such as a new payee, a checkout to a new address or a booking change, the real account holder gets a push, a passkey prompt or an emailed link showing exactly what the agent wants to do. The agent never sees or touches the approval.
- Action
- Add a new payee
- Amount
- $2,400.00
-
3
Give the agent a limited key
The okay opens a session for that one job: a spending cap, the actions allowed, a time lock and a kill switch. Routine actions the customer has already pre-approved go straight through, so people aren’t asked twice.
- Limit
- $2,400.00
- Time lock
- 20 min
-
4
Keep the signed record
Each okay becomes a signed record of who approved what, when, and for which agent, formatted as chargeback and dispute evidence. Customers see every agent they’ve connected on one page and can revoke any of them.
- Record
- HO-9F2C-41AB
- SHA-256
- 9f2c41ab…e07d
Built for the people who own agent risk
For banks and fintechs, retailers and marketplaces, and travel and loyalty programmes, where agents already reach logged-in accounts.
Fraud and risk
Separate the agent sessions your customers approved from everything else. Fewer account takeovers and “my agent did it” disputes, without blanket blocks.
Identity and login
Step-up that still works when the “user” is an agent: passkey approvals, limited sessions and instant revocation, alongside the login you already run.
Digital and e-commerce
Say yes to agent orders as they grow, instead of losing them to a CAPTCHA or a blocked checkout.
Compliance
A record of exactly what each customer authorised, and a page where they can see and revoke it. That’s the control and transparency banks’ new agentic-commerce principles ask for.
Other tools see the agent. HumanOkay asks the customer.
| Tool | Decides based on | Covers | What’s missing |
|---|---|---|---|
| Bot management | The type of agent, such as every agent from one vendor | Traffic and sessions | The individual customer’s consent, and evidence for each action |
| Fraud scoring | A risk score for the order | Transactions | Account actions that aren’t payments, and the customer’s say |
| Card-network agent programmes | The cardholder’s approval of a card payment | Card payments | Logins, account changes and other ways to pay |
| HumanOkay | This customer’s okay for this exact action | Any risky action in the account, payment or not | Nothing to rip out: it’s designed to work alongside the three above |
Fits on top of what you already run
HumanOkay is designed to read the agent signals from the bot management you already pay for and to plug into your existing login, whoever provides it. It’s built on open standards, so neither you nor the agents are locked in.
Start with an agent exposure audit
We’re taking a small number of design partners. Each one starts with a paid audit of two to four weeks, run on your own logs, before anything is deployed.
- Who it’s for
- Banks and fintechs, retailers and marketplaces, travel and loyalty programmes
- What you get
- A written report on agent activity in your logged-in flows, and a recommended agent policy
- As a partner
- You shape the approval flows and the evidence format, and get first access to HumanOkay in production
Agent exposure audit
2 to 4 weeks, on your own logs
- Share of logged-in sessions that are really agents
- Which agents, and who runs them
- What they do: logins, checkouts, payees, account changes
- Fraud and dispute rates for agents compared with people
- Where a human okay would have stopped a loss
KeepA written report and a recommended agent policy
Questions buyers ask first
Do the companies that make AI agents need to integrate with you?
No. The okay goes to your customer on their own device, not through the agent, so it’s designed to work with any agent, including ones that don’t identify themselves. Agents that do sign their requests get a smoother path.
Does this replace our bot management or fraud tools?
No. HumanOkay is designed to use the signals those tools already produce. They tell you an agent is there; HumanOkay gets the customer’s consent for the specific action and keeps the evidence.
How is this different from Visa’s and Mastercard’s agent programmes?
Those programmes cover card payments. HumanOkay covers the rest of the account as well: logins, new payees, address and email changes, bookings, loyalty points and payment methods other than cards.
Won’t customers get tired of approving things?
They’re only asked about risky actions. Customers can set standing rules, such as reorders under $100, and those go straight through.
What does the evidence look like?
A signed record of the agent, the customer, the exact action, the limits, the time and how it was approved, formatted for card chargebacks and Regulation E disputes.
Where are you today?
We’re working with a small number of design partners, starting with the agent exposure audit. If agents are already reaching your customers’ accounts, express your interest below.
Give every agent action two keys.
If your team is preparing for customers to act through AI agents, leave your work email. We’ll follow up about the design-partner programme and agent exposure audit.
- A short reply from the HumanOkay team
- Your details stay private
- Start with one bounded agent action
Express interest
Interest noted.
We’ll reply to about the design-partner programme.
Thank you. There’s nothing else to schedule or confirm.