Mercury launched Mercury Spend on August 11, 2026. The product does what most spend management platforms do: employee cards, intelligent budgets, policy enforcement at the point of sale. But one feature sits in a different category entirely. Mercury now lets businesses issue corporate cards to AI agents.
Not to the person who built the agent. Not to the team that manages the agent. To the agent itself.
A human creates the card. A human sets the budget ceiling, the merchant restrictions, the category limits. Then the agent spends autonomously. No approval queue per transaction. No human in the loop at checkout. The card gets declined if it exceeds the budget. The card freezes if a required receipt goes missing. Every transaction logs automatically.
The practical version of this looks like a procurement agent that buys SaaS licenses when a team request comes in, checks it against budget, pays, and files the receipt. Or a marketing agent that purchases ad inventory within pre-approved limits. Or an operations agent that orders supplies when inventory drops below threshold. All of it without a person clicking “approve” on each line item.
That is not a software feature. That is a coworker with a spending limit.
The Mental Model Problem
Most organizations still treat AI agents the way they treated spreadsheets in 2005. Useful. Passive. Something you open when you need it and close when you are done.
The spreadsheet mental model made sense when AI was a text box you typed questions into. You asked, it answered, you decided what to do with the answer. The human stayed in the center of every action.
Agents broke that model. An agent holds context across tasks. It remembers what happened yesterday. It connects information from different systems. It acts on its own within boundaries you set. It does not wait for a prompt. It works.
When Mercury issued an agent its own card, it made the mental model shift concrete. You do not give software a credit card. You give a team member a credit card. That distinction matters because it changes how you structure the work around them.
The organizations still running agents as tools are asking the wrong questions. They ask “what can this tool do?” instead of “what should this team member own?”
What the Controls Actually Look Like
The interesting part of Mercury Spend is not the autonomy. It is the governance.
Every Agent Card operates under constraints set by a human. Budget ceilings enforced at the point of sale, not after the fact. Merchant category restrictions that block spending outside approved vendors. Automatic freezing when documentation requirements are not met. Full audit trails on every transaction.
This is how you manage a junior employee on their first company card. You give them authority within clear limits. You trust them to operate, but you build the guardrails before they start, not after something goes wrong.
Mercury is not the only company thinking about this. The broader pattern is showing up across enterprise software. Asana launched AI Teammates in April 2026 that live inside projects alongside human team members, visible to everyone, updating shared roadmaps when priorities shift. These agents do not sit in a private chat with one user. They work in the open, in the same workspace, on the same tasks.
The common thread is transparency and constraint. The agent acts. The human sets the boundaries. The system logs everything. That is collaboration, not automation.
What this means if you run agents
If your organization runs AI agents today, or plans to within the next year, Mercury Spend is worth studying not for the product but for the operating principle underneath it.
The question is not whether your agents will need spending authority, tool access, or system permissions. They will. The question is whether you have a framework for granting that authority before the need arrives.
Most teams do not. A June 2026 Metrigy study of over 300 companies found that 65 percent already have a security and compliance strategy for AI agents. That sounds high until you realize it means 35 percent are deploying agents with no governance framework at all. And “having a strategy” does not mean the strategy works in practice. It means someone wrote a document.
The practical framework looks like this: define what the agent owns, set the boundaries before it starts, make its actions visible to the team, and review its work like you would a new hire’s. That is the whole trick.
Mercury built that framework into a credit card. The principle applies to every system your agents touch.
The Line That Already Moved
The shift from tool to team member is not theoretical. It already happened in the organizations paying attention.
Google shipped the Pixel 11 last week with an on-device agent that coordinates across forty apps without being asked. Asana built a project management layer where agents and humans share the same workspace. Mercury gave agents purchasing authority. Each of these moves treats the agent as a participant in the work, not an accessory to it.
The gap that mattered in 2025 was adoption. The gap forming now is structural: companies that reorganized around agents versus companies still treating them as fancy search bars.
One side builds governance, sets boundaries, and gives agents real work. The other side asks the chatbot a question, reads the answer, and does the work themselves.
The card is just the part you can see.