AI Agents and Automated Actions Policy
How Everdeck agents act, the approvals and human oversight around them, and who is responsible for what.
Draft for review. This policy is a draft pending review by a Singapore-qualified lawyer. It is not final and may change before Everdeck launches.
This policy explains how agents in Everdeck take actions, the safeguards around them, and the responsibilities of FLOW FN PTE. LTD. (UEN 202617303Z) and of our customers. It is part of the Terms of Service.
1. How agents work
An agent is an AI-driven worker you set up in Everdeck. Each one has a charter you write: its role and duties, who it reports to, which tools it may use, its work hours and its limits. Agents work in short sessions on tasks from their inbox, and every session is recorded.
Agents act only through tools. A tool is one action in an app you connected, an API or MCP server you added, or Everdeck itself, such as "create a task" or "send an email". An agent can use only the tools its charter grants.
2. Risk tiers
Every tool has a risk tier, which sets how carefully its use is checked:
| Tier | Meaning | Examples |
|---|---|---|
| 0. Read | Looks things up, with no side effects. | Search tasks, read a document |
| 1. Internal | Writes to Everdeck's own records or notifies your people. | Create a task, leave a comment |
| 2. Reversible | Changes something outside Everdeck that one step can undo. | Move a ticket, add a label |
| 3. External | Anything a customer or the public sees. | Send an email, publish a post |
| 4. Money or irreversible | Moves money, deletes data or changes permissions. | Refund a payment, delete a file, share access |
3. Approvals and human oversight
Before every tool call, Everdeck checks your approval policies. A policy can allow the action, refuse it, or require approval from people, roles or reviewer agents, in steps.
Some safeguards apply whatever your policies say:
- A person for tier 4. Moving money, deleting data and changing permissions always need at least one person to approve.
- No self-approval. The agent or person that asks for an action can't approve it.
- Silence is a no. If an approval request isn't answered in time, it's refused, never approved.
- Outside content raises the bar. Once an agent has read content from outside your company, such as a customer's email or a web page, its tier 3 and 4 actions need a person to approve by default for the rest of that session. An Owner can allow a named reviewer agent to approve tier 3 actions instead, but never tier 4.
- Exact details. An approval covers the exact details shown to the approver. If they change, the action must be approved again.
- Tier 4 is decided in Everdeck. If your company sets up approvals in Slack, people who linked their Slack account can decide agents' requests there with buttons, up to the tier your company chooses and never above tier 3. Every decision made in Slack is recorded like one made in Everdeck.
4. Your responsibilities
You decide what your agents may do and you supervise them. In particular, you are responsible for:
- writing charters, granting tools and setting limits and approval policies that suit your business and the risks involved;
- making sure the people and agents you name as approvers are able to judge the actions they approve;
- reviewing agent output before relying on it, especially for legal, financial, medical, employment or other significant matters;
- following the laws that apply to what your agents do, such as consumer protection, anti-spam and data protection laws, including telling people they are dealing with AI where the law requires;
- following our Acceptable Use Policy.
Actions an agent takes within your configuration are taken on your behalf, as set out in the Terms of Service.
5. The limits of AI
AI models can be wrong. They can misunderstand instructions, state false information confidently, or be manipulated by content they read. Our safeguards reduce these risks but cannot remove them. We do not guarantee any particular outcome from an agent's work.
6. Your model providers
With Everdeck you bring your own AI model keys. You choose each agent's model provider, and your agreement with that provider, including its terms, usage policies and data handling, applies to what your agents send it. You are responsible for the provider's charges. Spending caps in Everdeck are estimates based on published prices; your provider's bill is final.
7. No training on your data
We do not use your content, or the actions and results of your agents, to train or fine-tune AI models. For the platform features we pay for, such as the embedding model behind knowledge search, we choose providers whose terms do not allow them to train on what we send.
8. Stopping and reviewing
You can pause one agent, all the agents of a company or every agent in your account at any time. An agent pauses automatically when it reaches its monthly budget, after 3 failed sessions in a row, when no AI key works for its model, or when your account's billing stops allowing work (no active plan, or a payment overdue for more than 7 days), and stays paused until someone resumes it from its page. Your account's monthly spending cap holds every agent once it is reached, until you raise the cap or the next month starts. Each session also stops at the limits its charter sets, such as its time and estimated cost. Every session, tool call and approval decision is recorded in the run timeline and the audit log, so you can review what happened and why.
9. Questions and problems
If an agent did something you didn't expect, or you think a safeguard didn't work, email [email protected]. We take reports of safeguard failures seriously and will investigate them.