Human-in-the-Loop Approval Tiers for AI Agents, by Risk and Value

In 30 seconds

What it is Risk tier and amount set the approval, and some decisions are never automated.

You use it when You are deciding which AI agent actions need approval, by whom, and above which amounts.

Guide covers AI Governance

Two questions set the approval: how risky is the action, and how much money is involved. Pay, legal and people decisions are never automated.

Part 23 of 26 in SAP Business AI, Explained.

Not every agent action needs a human, and some always need one. On projects, the question I get most is where that line sits. I answer it with two questions: how risky is the action, and how much money is at stake? Here’s the grid that follows, and the habits that keep approvers from becoming rubber stamps.

If you only read one section

Draw a grid: risk tier down the side, value band across the top. Each cell says auto, one approver, two approvers or never automate.
Pay, legal outcomes and anything that changes someone’s employment never leave human hands, whatever the amount.
An approval only works if the approver looks. Show the evidence, re-check a random sample and watch for anyone who never says no.

The everyday picture

Finance already knows this as the four-eyes principle, or maker-checker. One person prepares a payment and another approves it. Above a certain amount two people approve, and some things, like a colleague’s salary change, never take a shortcut.

An artificial intelligence (AI) agent is just a very fast maker. The checker rules don’t change, but volume does. An agent can prepare in a minute what a team prepared in a week, so the checking has to be designed, not assumed.

Setting the tiers, step by step

Classify the action’s risk. Low is internal and easy to reverse, like assigning a payment to the matching invoice. Medium is when money leaves the company or a customer sees the result, such as a refund or reminder letter. High touches pay, legal position or someone’s job.
Set value bands. I start from the existing delegation of authority for people, and the agent never gets more authority than the human role it supports.
Read the tier off the grid.
Route and record the approval as a task with evidence, a named approver and a deadline. I covered this pattern in Build Work Zone and Process Automation.
Review the tiers monthly, moving cells up or down on override rates and incidents, not gut feel.

OWASP (Open Worldwide Application Security Project) names too much autonomy as a root cause of agent harm and lists human approval for high-impact actions as a mitigation. Your tiers are where you decide what “high-impact” means for your business.

Four payments through the NextSAP grid

At NextSAP Utilities, the fictional Dubai utility used across this site’s SAP content, the payment clarification agent (agent 4) works in FI-CA (Contract Accounts Receivable and Payable), proposing how to assign unclear incoming payments to open items, such as Marina Bay Café LLC paying several bills in one transfer. It sits at L2 Propose today. Before anyone talks about more autonomy, Layla, the chief financial officer (CFO), asks Omar, the finance controller, and his shared-services team to agree tiers.

NextSAP is invented, but this is the kind of grid I’d draw with a real finance team. They agree three value bands in UAE dirham (AED): under AED 5,000, AED 5,000 to 499,999, and AED 500,000 or more.

AED 4,800, exact match

A customer pays AED 4,800 quoting the invoice number, matching to the fils. Low risk, under AED 5,000 and reversible, so it’s the lightest tier. For this agent that means a one-click approval by a person: a clerk confirms the match and posts it. Even when the agent later reaches L3, these items get one-click approval and never post automatically. One in 20 is re-checked by a collector the next morning.

AED 500,000, partial and split

Al Noor Property Management pays AED 500,000 against 37 invoices for its premises, totalling AED 512,400, with no remittance advice. The agent proposes which invoices to clear and flags a short payment of AED 12,400. That’s the top band, so two approvers, Elena, the collections lead, and a finance manager, both see the evidence the agent used.

A refund of AED 7,500

When Marina Bay Café LLC moves out of one of its premises, the final meter read is billed and its security deposit is set against that bill, leaving a credit balance of AED 7,500. The agent proposes refunding it. Money leaves the company, so it’s medium risk with two approvers, despite the far smaller amount.

Salary changes

Mariam’s salary-change agent (agent 3) helps managers prepare requests. Salary changes are high risk at any amount, so never automated. The agent drafts and checks, and people approve in the normal human resources workflow.

Keeping approvers awake

Rubber-stamping is the real enemy of human-in-the-loop. The European Union (EU) AI Act’s human oversight article names it: overseers should stay aware of “automation bias”, the habit of over-trusting a system’s output. Even if the Act doesn’t apply to you, copy the controls.

Problem	Control	NextSAP example
Approver can’t see why	Show the evidence on the task: source documents, the rule matched, what differs from usual	The AED 500,000 task shows the 37 invoices, the AED 12,400 gap and the customer’s last three payments
Approver approves everything	Track approval rate and time to decide per approver	A collector approving 100% in a median of 4 seconds triggers a coaching conversation
Light-touch items never checked	Re-check a random sample after the fact	1 in 20 one-click AED 4,800-type approvals reviewed next morning
Drift goes unseen	Track override rate per agent and per tier	Tier rule: if overrides in a tier exceed 10% in a week, that tier moves up one level. Agent 4 also has its own pause rule, a weekly override rate above 15%
Approvers can’t test themselves	Seed known wrong proposals into the queue now and then	Two deliberately wrong test proposals per month, logged and scored
Batch approval hides problems	Cap batch size for medium and high tiers	Maximum 20 items per batch approval, and two-approver items are never batched

Where tiers help, and where they stop

Good tiers match human checking to the real risk, spare people from approving trivial reversible actions, and record who approved what on which evidence. They can’t make a wrong proposal right, though. They only catch it if someone looks. They don’t replace technical limits on the agent either, and they drift out of date as volumes and error rates change.

Explain it to your CFO

Small, routine, reversible items get the lightest check, such as a one-click approval, plus spot checks afterwards.
Bigger amounts, or money leaving the company, need one or two people, as today.
Salary and legal decisions always stay with people, whatever the amount.

Decisions for your finance team

What are your value bands? Copy your existing delegation of authority, capped at the human role the agent supports.
Which actions are never automated? I’d list pay, employment changes, legal commitments and anything a customer can’t easily undo.
What sample rate checks automatic items? Start at 1 in 20 for the first 90 days, then adjust to the error rate you find.

Go deeper

Technical detail for architects and developers

Who sets the tier? An SAP Community post on agent management lists properties tooling doesn’t provide out of the box, including the autonomy level from which a human approves. Store tier rules as data your workflow reads, not inside an agent’s prompt.

Enforce the tier outside the model. OWASP’s excessive agency mitigations include checking authorisation in the downstream system, not trusting the model’s judgement. So the backend role behind the agent’s identity must refuse a posting above the agent’s band, even if the agent tries.

SAP’s Joule Studio post describes standard agent extension, including a planned after-processing hook that can trigger an n8n workflow to route an exception for manual review and log the outcome. That’s a natural spot for tier logic on a prebuilt agent. SAP sources differ on whether n8n is available in Joule Studio yet, so check the current status in the Joule capability watch.

SAP describes SAP Cloud ALM (Application Lifecycle Management) as tracking agent actions once agents are live. Feed approval outcomes, override rates and time-to-decide into that monitoring so tier changes rest on evidence.

For high-risk systems, Article 14 of the EU AI Act expects overseers to understand the system’s limits, interpret its output, decide not to use it or override it, and halt it safely. The NIST (National Institute of Standards and Technology) AI risk framework puts this monitoring under Measure and Manage. Both make good checklists for the approval screen itself.

Words you’ll hear

Human-in-the-loop: a person reviews or approves before an agent’s action takes effect.
Four-eyes principle (maker-checker): one person prepares and a second person approves.
Delegation of authority: the company’s rules for who may approve what, up to which amount.
Automation bias: the tendency to trust a system’s output without checking it.
Rubber stamp: an approval given without real review.
Override rate: the share of agent proposals that humans change or reject.
FI-CA (Contract Accounts Receivable and Payable): SAP’s high-volume receivables and payables component.

Keep going

Independent explainer, not affiliated with or endorsed by SAP SE. SAP, Joule and other SAP product names are trademarks of SAP SE.

Frequently asked questions

Does every AI agent action need a human approval?

No. Low-risk, reversible, low-value actions can run automatically if a sample is checked afterwards. Approvals should grow with risk and amount, and some decisions, such as pay changes, should always stay with a person.

Where do the AED thresholds come from?

Start from your existing delegation of authority for people. An agent should never be allowed more than the human role it supports. The amounts in this post are an illustration for a fictional company.

What is a rubber-stamp approval?

An approval given without really looking, usually because the approver sees hundreds of tasks and the agent is usually right. What helps is evidence on screen, random re-checks and watching who never overrides.

Does SAP set the approval tiers for me?

No. SAP tools route and record approvals, but an SAP Community post on agent management notes that the autonomy level from which a human approves is not set out of the box. That's your policy call.