The Task Approval Agent: Human-in-the-Loop AI Automation for Email-Driven Work
By Anas Shaikh · June 10, 2026 · 4 min read
Full autonomy is the wrong default for most business automation. When an AI agent can create tasks, send replies or move money, the right design isn't "let it run" — it's "let it propose, and let a human approve." That single design choice is what makes AI workflow automation safe to deploy in a real company.
The Task Approval Agent we built does exactly this: it turns a noisy inbox into a clean queue of suggested actions, each one a tap away from approval. Here's how it works and why the human-in-the-loop pattern is worth copying.
The problem it solves
Important work hides inside email. A client asks for a change, a vendor confirms a delivery, a teammate requests a follow-up — and each one should become a tracked task. Doing that by hand is exactly the kind of repetitive work that quietly eats a team's week.
But fully automating it is risky. An agent that creates tasks unsupervised will eventually create the wrong one, and trust evaporates fast. So we kept the human in the loop — without keeping them in the grind.
The flow
Incoming email
│
▼
[ Classification agent ] → Is this actionable? What kind of task?
│ (actionable)
▼
[ Teams Adaptive Card ] → Suggested task, posted for a human to review
│
Approve / Reject (one tap)
│ approve
▼
[ Task created + tracked ]
Each stage is deliberately simple, which is what makes it reliable.
1. Classification
The agent reads each incoming message and decides two things: is there an action here, and if so, what is it? Non-actionable mail (newsletters, FYIs, noise) never reaches a human. This is where the time savings come from — the queue only contains things that might actually need doing.
2. The Adaptive Card
Instead of silently creating a task, the agent posts a Teams Adaptive Card with its proposed task: title, summary, source email, suggested owner. The reviewer sees the agent's reasoning and approves or rejects with a single tap — no context-switching into another tool.
3. The approval queue
Approved items move into a tracked queue and become real tasks. Rejected items are logged so the classifier can be tuned over time. The two-stage design — a pending queue feeding a tracked store — means nothing is created without sign-off, and nothing approved is ever lost.
The engineering that makes it dependable
State that survives restarts. Approval state and message dedup are persisted to disk (JSON), so a service restart never double-posts a card or loses a pending approval.
Provider-agnostic email intake. Different mailboxes need different access paths — the Microsoft Graph API for Microsoft 365 mailboxes, IMAP for Gmail and Yahoo, and COM-based automation for client tenants where we don't have app-registration rights. The agent routes per account, so onboarding a new mailbox is configuration, not a rewrite.
UID-based deduplication. Every message carries a stable identifier. We check it before processing so the same email never spawns two cards, even if it's fetched twice.
Why human-in-the-loop wins
The instinct with AI is to automate end to end. But the most adopted automations leave a human at the decision point and automate everything around it: the reading, the classifying, the drafting, the routing. The person does the one thing humans are good at — judgement — and the agent does the ten things they shouldn't have to.
This pattern generalises far beyond email. Invoice approvals, content publishing, customer responses, expense categorisation — anywhere a wrong action is costly, "propose then approve" is the safe, fast path to AI automation people actually trust.
Want this for your team?
If your inbox is where tasks go to get forgotten, this is a workflow worth automating — with the guardrails to make it safe.