Oversight

How do I keep humans in control?

A finance lead reviews something on her laptop while a small AI agent beside her holds up an envelope, waiting for her decision

Agents do their best work on their own. But some decisions should always stay with a person. You decide which ones, and you can change your mind as trust grows.

The challenge

An agent that has to ask about everything saves no time. An agent that never asks is a risk. Most teams want something in between: let the agent do the routine work, and keep people in charge of what matters. The hard part is drawing that line, and moving it as you learn.

A short story

Maria leads accounts payable. Her team is setting up an agent that checks incoming invoices and emails suppliers when something is missing, such as an order number.

Before it goes live, she decides what needs her yes.

  • The agent may only email suppliers on an allowed list. For now, it must ask for her approval even for them. Anyone outside the list is blocked.
  • Her instructions say: if you’re unsure, ask me instead of guessing.
  • The checklist the agent follows is marked Protected. Nobody can get the agent to change it unless they have Edit access and are in the chat.

Monday morning, the agent runs on schedule and drafts three reminders. Maria gets a notification. She opens the agent’s Approvals tab, edits one email to soften the tone, and approves all three.

Wednesday, an invoice doesn’t match its order and the agent can’t tell which is right. It sends Maria an owner notice. She clicks Discuss with agent and tells it what to do. Then she asks it to update its instructions, so next time it knows how to handle similar cases. The agent suggests the new wording and Maria approves it. The change is saved, and she can undo it at any time.

A month later, Maria reads the agent’s Diary to see why it flagged what it flagged. The reminders have been right every time. She changes one rule: suppliers on the list now get their emails straight away. Everyone else is still blocked.

The agent earned more freedom. Maria still decides how much.

How much should an agent decide on its own?

It depends on what’s at stake. Think of it as a scale, and place each type of decision on it.

  1. Agent decidesLow stakes, easy to undo.For example: Setting a ticket’s priority
  2. Agent decides, tells youMedium stakes.For example: Internal status updates
  3. Agent suggests, you approveHigher stakes.For example: A customer refund
  4. You decideCritical, unusual or unclear.For example: Escalating a complaint to leadership

How does Abundly keep people in charge?

Grouped by what you need, not by feature.

  • 1

    Decide what needs a person’s yes

    The agent pauses until someone approves or rejects.

    • For email, SMS and phone calls: an allowed list of recipients. Inside the list, send or ask. Outside it, block or ask.
    • Other actions, such as requests to outside services, can require approval too.
    • Requests appear in the chat, or on the agent’s Approvals tab when it works on its own. Owners get a notification.
    • You can edit an email before it goes out.
  • 2

    Let the agent ask instead of guess

    When nobody is watching, the agent can still reach a person.

    • Every agent can send an owner notice when it’s blocked or needs a decision.
    • Owners read them under Activity → Owner notices and can start a chat with the agent about them.
  • 3

    Keep the rules in human hands

    The agent can’t quietly rewrite its own job.

    • By default, an agent can only change its instructions in an Abundly chat with someone who has Edit access. Not from Slack, Teams or automated runs.
    • Protected documents hold checklists and scripts to the same standard.
    • Every instruction change is saved. Compare versions, see whether a person or the agent made each change, and undo it.
  • 4

    See what the agent did, and why

    Oversight without reading every message.

    • The Diary explains the agent’s reasoning in plain language.
    • The Activity log shows each automated run step by step.
    • Messages keeps every email and SMS the agent sent and received.

Tip: start further towards human control than you think you need. Loosen the rules once the agent has shown it gets things right.

Who is responsible for an agent?

A person, always. Every agent should have a named owner who sets its rules, answers its questions and follows up when something goes wrong. An agent can do the work. It can’t take the responsibility.

Where do I start?

  1. 1

    Name an owner for every agent

    One person, or a pair, who answers for what it does.

  2. 2

    Require approval for anything that leaves the company

    Emails, messages and calls to customers or suppliers are a good place to start.

  3. 3

    Loosen step by step

    Read the Diary, check the approvals, and give the agent more room as it proves itself.

For admins: you can also decide what may change on agents that are live in production. See How do I manage the lifecycle of agents?

See it in your own workspace

We will walk you through the platform with your own agents and rules in mind.

Request a demoMore questions answered