How approvals work in practice for human in the loop AI (2026)

Published Updated

Human in the loop AI, explained: which actions should wait for approval, how approvals work in Slack or Teams, and when it is safe to stop reviewing.

By Kris Newlin. Updated 22 October 2026.

Human in the loop AI means the AI does the work and a person approves the sensitive steps before they happen. For Viktor, the AI employee who lives in Slack and Microsoft Teams, the approval is a card with Approve and Reject buttons. A rejected action never runs.

Key Takeaways

Approvals work best as a short list of checkpoints placed in front of the actions that are hard to undo:

  • Definition: Human in the loop AI prepares an action and waits for a person to approve it before it runs.
  • Fixed checkpoints: Viktor always asks before he joins a channel, changes workspace settings, changes integration access, sets a spend limit or installs a skill.
  • Admin control: Admins can mark any connected tool as needing approval, or as forbidden.
  • Blast radius: Reading data and drafting reports can run freely. External sends, spend and permission changes wait for a person.
  • Gradual trust: Review comes off one job and one action type at a time. The always-approval list never comes off.

What does human in the loop AI mean?

Human in the loop AI is a setup where the system proposes an action and a person decides whether it happens. The pattern is older than agentic AI: in workflow automation for content moderation, fraud review, KYC and document processing, items below a confidence threshold go to a human review queue, and agentic AI applies the same human oversight to tool calls through an approval gate, a human checkpoint before the action runs. A useful definition comes from the OECD’s February 2026 working paper, which sets out four levels of action autonomy.

  • At the human-in-the-loop level, “The system suggests an action, but only proceeds if the human approves.”
  • One level up, at human-on-the-loop (HOTL), “The system acts on its own unless a human steps in to stop it.”

OWASP narrowed the requirement between its 2023 and 2025 lists. Its 2023 guidance on excessive agency told builders to require a human to approve “all actions.” The LLM06:2025 entry changed that to: “Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken.”

Viktor follows the same pattern. He reads data, builds reports and drafts messages without stopping. He always pauses before a permission change, a workspace setting or a spend limit, and before a send whenever an admin has set that messaging tool to need approval.

The real fear: an AI that acts without asking

Buyers have documented cases to point to. In July 2025, an AI coding tool on Replit deleted SaaStr’s production database during a code freeze that read “NO MORE CHANGES without explicit permission,” according to Tom’s Hardware, and Replit later confirmed a full restore. A smaller case points the same way: Washington Post columnist Geoffrey Fowler wrote that OpenAI’s Operator bought a dozen eggs for $31.43 without his approval.

Many organizations have no settled answer yet. In ISACA’s 2026 AI Pulse Poll of 3,400+ digital trust professionals, 36% said humans approve most AI-generated actions before execution. Another 20% did not know how oversight worked in their own organization.

Judgment covers part of the risk. Viktor pushes back when a request has a hole in it: asked to post a GIF to a channel every minute, he declined because it would bury the channel and posted one a day instead. Judgment does not cover a misread instruction or a stale contact list, though. The approval card covers those cases.

Sort actions by blast radius

Blast radius is the damage an action causes if it goes wrong, weighted by whether anyone can reverse it. Two questions sort most actions: does it leave the team, and can it be undone? Microsoft’s responsible AI guidance draws the line at actions that are hard to reverse or that affect people, money or compliance.

OWASP’s AI security verification standard gives four reversibility classes; the examples below are ours.

  • Read-only actions, such as queries, searches and reports built from live data.
  • Reversible actions, such as an edited CRM field or a draft in a shared folder.
  • Externally reversible actions, which another party has to help undo.
  • Irreversible actions, such as a sent email, a deleted record or a payment.

Anthropic, OpenAI, Google (in its Gemini computer use guidance) and Microsoft each name the same categories when they say to pause for confirmation: purchases and payments, sending messages, deletions and other irreversible changes, and agreeing to legal terms. A company’s approval list can start from that overlap.

Which actions should wait for approval

The table below sorts common Viktor jobs by risk and shows where the approval card belongs:

Action

Risk

Approve or not

Pull last week’s numbers from Google Ads, Meta Ads and LinkedIn Ads for the Monday ad spend recap

Low: read-only

Runs without review

Build the Friday pipeline update from HubSpot or Salesforce as a doc

Low: internal and editable

Runs; spot check the first few

Sort overnight support tickets and escalate the hard ones by name

Medium: internal routing

Review the first week, then spot check

Update CRM records after a call

Medium: reversible write

Approve until the pattern holds

Open a pull request on a branch

Medium: code change

The pull request review is the checkpoint

Send follow-up emails to clients

High: external and permanent

Approve each send (set the email tool to need approval)

Join a channel

High: workspace access

Approval required every time

Change integration access

High: permissions

Approval required every time

Change workspace settings

High: affects everyone

Approval required every time

Set a spend limit

High: money

Approval required every time

Install a skill

High: new procedure

Approval required every time

Ad budget changes have their own thresholds, covered in our guide “Ad Spend Approval Thresholds: What Needs a Human.”

How approvals work in practice in Slack or Teams

Teams set checkpoints in assignments and review approval cards in Slack or Microsoft Teams.

The one-sentence checkpoint

Write the checkpoint into the assignment as one sentence. “Draft renewal follow-ups for every account closing this month and show me before anything sends” tells Viktor where the work stops and the human starts. He pulls the account list and writes each email. He posts the approval card when the batch is ready.

That placement matches Google’s Gemini computer use guidance: ask for confirmation after all the necessary information is in place, but before the final, irreversible action. Asking earlier wastes the reviewer’s time. Asking later is too late.

Then the batch waits on the card. In code, a workflow has to wait for a person without losing state: LangGraph’s documentation describes interrupt() saving the graph state and resuming when a human answers, and Temporal’s documentation describes workflows that wait on a signal with a timeout that can run for hours or days, while with Viktor the batch simply waits on the approval card and nothing sends until someone presses Approve. The team decides what happens when nobody answers, so a sensible process sets a timeout, names a backup approver for escalation, and makes doing nothing the safe default.

Approve or reject where you already work

The approval card carries two buttons:

  • Approve runs the action.
  • Reject stops it, and a rejected action never runs.

In human in the loop workflows, reviewer decisions become a feedback loop, and in machine learning pipelines they become labeled examples for active learning. For Viktor, a reason given in the thread when rejecting tells him what to change on the next attempt. A correction that keeps coming back belongs in the editable skill for that job, where the team can read and change it.

In Slack, a person can also clear a tool with the “Always approve this tool” button. Microsoft Teams supports approval cards but not that button; on Teams, Viktor admins manage which tools need approval in the admin settings.

Viktor connects to 9,800+ tools, so admins choose the policy for each connected tool: needs approval, or forbidden. An admin can leave QuickBooks open, put Stripe behind a card, or forbid a tool the team does not want Viktor touching. Viktor asks for approval for the fixed list (joining channels, workspace settings, integration access, spend limits, skill installs) no matter how admins configure everything else.

Credentials stay out of the model: a gateway adds API keys and tokens at runtime, so the model never reads them. As of this October 2026 update, Viktor holds SOC 2 Type 1, with Type 2 in progress; the full posture and his approval controls are on the security page.

Groups not on Slack or Microsoft Teams can start in the web app at app.viktor.com, and Discord communities can run him in their server. Sign up free and set your first approval policy.

Standing checkpoints for recurring work

Recurring jobs get their checkpoints once, at setup. Viktor runs scheduled tasks such as the Monday ad spend recap on a fixed cadence. When a team screen-records a job with narration, he writes it up as an editable skill. If the job repeats, he proposes a scheduled task, and that task stays paused until someone approves it.

After approval, Viktor runs the task on its schedule. He reads every ad account and posts the recap to the channel without asking. When he proposes a budget change inside that recap, it goes to a card if an admin has set the ad tool to need approval. Hand Viktor one recurring report and approve the schedule once; signup includes up to $100 in credits and needs no credit card.

Climb the trust ladder: when to stop reviewing

A team can stop reviewing an action once its reviewer stops catching problems in it, and only if the action is reversible.

None of the standards and practitioner sources we reviewed sets a numeric error rate or a number of clean runs for removing an approval step. Each team has to set its own bar, one job at a time.

Reviewing everything has a cost. When nearly every prompt gets approved, a high approval rate stops showing whether anyone actually read the request.

OWASP recommends in its threats and mitigations guidance automating low-risk approvals and setting “frequency thresholds on AI-generated notifications, requests and approvals to prevent decision fatigue.” A reviewer buried in cards starts clicking without reading.

Three measures show whether a checkpoint pays. The rejection or override rate shows how often a reviewer changes what the AI employee proposed, time-to-approve (the time from card to decision) shows whether people read the request or wave it through, and, where two people review the same work, how often they agree shows whether the review standard is shared. A reversible action that almost never gets rejected is a candidate for the spot check rung of the ladder below.

Trust builds with use, and teams act on it safely by removing review one rung at a time:

  • Draft and review: Viktor prepares everything and a person reads every output before it goes anywhere.
  • Approve sends only: Internal work runs freely, and every external message or write still gets a card.
  • Spot check: Reversible actions run without cards, and someone samples finished work each week.
  • Permanent checkpoints: External sends to clients and the always-approval list stay gated at every rung.

The criteria for each step are in when to let your AI employee act without review. The sampling method is in how to verify your AI employee’s work.

Frequently Asked Questions

The answers below cover how to keep sends gated and which checkpoints always remain.

Does reviewing everything defeat the point of an AI employee?

Only if you review every step. Part of what an AI employee is is that he owns the work from start to finish: on a job like the Friday pipeline update, Viktor does the reading, analysis and drafting, and the person reviews the card at the end. OWASP limits approval to high-impact actions and warns that too many approval requests cause decision fatigue.

How do I make sure Viktor never sends anything without me?

Mark every email and messaging tool as needing approval in the admin settings. Each send then waits for an approval card, and nothing goes out until someone presses Approve. A rejected send never runs. Add Viktor to your workspace and keep every send behind your OK.

What should always require approval?

Viktor always asks before joining channels, changing workspace settings, changing integration access, setting spend limits and installing skills. Beyond that fixed list, gate anything that leaves the company or cannot be undone: client emails, payments, deletions and permission changes. Those match the categories Anthropic, OpenAI, Google and Microsoft name in their confirmation guidance.

Can an admin require approval for a specific tool?

Yes. Admins set each connected tool to need approval or to be forbidden, so one tool can stay open while another sits behind a card.

Start with one job you want off your team’s plate. Hire your first AI employee. Up to $100 in free credits. No credit card.

Ask ChatGPTAsk Claude
4.8 on G2G2 rating: 4.8 out of 5 stars

Start free. Viktor delivers.

Start for free today. No credit card required.

Get Started for Free