Agents can now spend. Few teams can say on what, up to how much, or who approved it.
Agents buy software, pay invoices and renew subscriptions. When you run them for several clients, the risk isn't one bad answer: it's an action nobody agreed to, and no record to show what happened.
No spending limits per agent
An API key or a saved card gives an agent the same authority as the person who set it up.
Approvals live in chat threads
"Is this OK?" in a message isn't a control: it's easy to miss, and hard to prove later.
No record you can show a client
When a client asks why their agent bought something, you need an answer with evidence.
Four controls, one place.
Each card says what you can try in the sandbox today and what is still being built.
Spending policies
Per-agent limits per purchase and per month, and a list of approved vendors. Rules are written and testable. No AI model makes the allow/deny call, and anything not permitted is denied.
Vendor checks match by name today; verified payees are in development.
Human approvals
Purchases above a threshold wait for a signed-in approver using two-factor sign-in. An approval lets that exact request run once, and nothing else.
Agent controls
Each agent gets its own key and connects through an API or an MCP tool. Pause one agent, or all of them, at any time.
Hosted, sign-in-based connection for chat assistants is in development.
Audit trail
Every request, decision, approval and pause is recorded with the reason, and you can export it as a spreadsheet.
One purchase request, start to finish.
An illustrative example. Fisher makes the decision; making a payment is a separate step, and in the sandbox no payment is ever made.
-
1
The agent asks first
A purchasing agent for an example clinic asks Fisher before buying.
request_payment agent: front-desk-assistant vendor: "Team chat (Pro plan)" amount: 65.00 CAD / month reason: "Free plan's message history is full" -
2
Fisher checks the policy
- Pass Agent is active and its key is valid
- Pass Vendor is on the approved list
- Pass Within the $500 monthly limit
- Rule Over $40: needs the clinic manager
-
3
A person decides
Needs approval · expires in 23 hTeam chat (Pro plan), $65.00/month, requested by the front-desk assistant.
Approving lets this exact amount run once.
-
4
Decision and payment, kept apart
DecisionApprovedPaymentNot run · sandboxDecisionDeniedPaymentNothing to run -
5
Everything is on the record
- 20:39:02 Request received from front-desk-assistant
- 20:39:02 Rule matched: over $40 needs approval
- 20:51:47 Approved by clinic manager (two-factor sign-in)
- 20:51:47 Payment step: not run (sandbox)
Illustrative sandbox example. Names, amounts and times are made up.
Four steps.
Connect an agent
Give each agent its own key, through an API or an MCP tool.
Set its rules
Limits, approved vendors and when a person must approve.
Approve what matters
The right person sees what the agent wants and why, and decides.
Show the record
Every request and decision, ready to share with the client.
Built for teams that run agents for many clients.
This is what we're building next, with design partners. Today's sandbox manages one business at a time.
- A separate workspace, policy and record for each client In development
- Approvals routed to each client's own approver In development
- Policy templates you reuse across clients In development
- A per-client report you can hand to the client Planned
Where payments fit, stated carefully.
Fisher's job is the decision. Over time we want approved decisions to connect to real payment rails. None of the later stages below is available today, and we have no partnerships to announce.
Decisions in a sandbox
Policies, approvals and audit, tested with made-up requests and test-mode data only.
Supported payment providers
Let an approved decision run once, exactly as approved, through providers we support. Subject to each provider's eligibility and terms.
Cards and agent payment standards
Virtual cards through eligible infrastructure partners, and tokenized agent payments as industry standards mature.
Clear about what Fisher can and can't do.
Fisher only controls actions that go through it. If an agent has its own direct access to a card, account or system and doesn't ask Fisher first, Fisher can't stop that action. Part of working with us is deciding together which actions to route through Fisher.
Fisher is in a private pilot and isn't generally available. It does not move money or issue cards today. Everything you see on this page runs in a sandbox with made-up or test-mode data.
Apply to be a design partner.
For agencies running action-taking AI agents for client businesses. You get early access to the sandbox and a direct line to the founder; we ask for a 30-minute conversation about how you control your agents today.
Apply by emailThis opens your email app with a few short questions. Prefer to just say hello? Write to hello@fisherpay.ca.
- You email us with a few lines about your agency.
- We reply personally: no mailing list, no automated emails.
- We talk for 30 minutes and walk through the sandbox together.