TL;DR
Human in the loop AI agents pause before chosen actions so a person can approve, edit or reject them. Not every action needs that pause. Read-only and easily reversed steps can run alone, while payments, external messages and production changes wait for approval. Confidence scores and policy rules decide what else goes to a reviewer queue. Frameworks such as LangGraph pause the agent, save its state and resume after the decision. Track approval rate, override rate, time-to-decision and escalation precision to prove the human checkpoint works.
Key Takeaways Agents with a human in the loop stop at defined decision points, so people review high-impact actions instead of every step. Sort agent actions into risk tiers by impact, reversibility, uncertainty, external exposure and legal duty before you add any approval gate. Five patterns cover most needs, namely approval gates, review and edit, confidence escalation, exception queues and an emergency stop. High model confidence never replaces approval for an irreversible or regulated action, because confidence measures the model, not the consequence. Approver experience decides whether oversight is real, so each review needs evidence, a deadline and a safe default when nobody answers. A KPI scorecard and the EU AI Act Article 14 checklist show whether the human checkpoint changes outcomes or only adds clicks. Watch on YouTube
How to Build Effective AI Guardrails for Enterpises?
A Kanerika discussion on the guardrails that keep enterprise AI in bounds, including where human review belongs and which limits should be enforced in code.
The Approval Dilemma Every Agent Team Hits
Your support agent can issue refunds, update orders and also email customers. If you put a human approval on every one of those actions, the queue fills by lunch. By the afternoon, reviewers click approve without reading, so the control exists only on paper.
Now remove the approvals. At first, nothing goes wrong for weeks. Then one prompt-injected ticket convinces the agent to refund a large order twice, and the payment leaves before anyone looks.
So teams face a real dilemma. Too many checkpoints make the agent slow and turn reviewers into rubber stamps, while too few let one bad action reach customers. This guide shows where an agent should pause for a person, how to build the pause and how to measure whether it works.
What Are Human in the Loop AI Agents?
Human in the loop AI agents hand a decision to a person at set points before they continue. First, the agent plans the task, calls tools and prepares an action. Then a rule checks that action, and if the rule fires, the agent waits for a human to approve, edit or reject it.
That pause is also what separates an agent from a chatbot with a review step. An agentic AI system acts on real systems, so the cost of a mistake is a wrong action rather than a wrong sentence. The human checkpoint therefore sits in front of the action, not after it.
Agent Oversight vs Human Feedback in Machine Learning
The phrase “human in the loop” also describes people labeling training data or ranking model outputs. That older meaning improves the model over weeks or months, so its payoff is slow. Agent oversight, by contrast, happens in real time on a single task, and the person decides whether this specific action goes ahead.
Both are useful, and they solve different problems. Labeling raises average quality, while an approval gate stops one bad action today. This article covers the second meaning, which is the one that decides how safely an agent runs in production.
The Review Cycle in Four Steps
Every working design follows the same loop. First, the agent proposes an action with its inputs and reasons. Second, an oversight rule scores that action against risk tier, confidence and policy. If the rule asks for a human, the agent pauses and saves its state.
Third, a reviewer approves, edits or rejects the proposal. Finally, the agent resumes from the saved point, and the decision is logged for audit and tuning. The log matters as much as the pause, because it is the only record of why the action went ahead.
Because of this, each step needs an owner. For instance, the rule belongs to the risk team, the queue belongs to operations, and the resume logic belongs to engineering. When nobody owns a step, the cycle fails at that point and nobody notices.
In the Loop, On the Loop or Out of the Loop
Human oversight generally comes in three strengths. In the loop means a person approves before the action happens. On the loop means the agent acts while a person watches and can step in, and out of the loop means the agent runs without case-level review inside fixed limits.
Choosing between them is a risk call that the business owner makes. In practice, the same agent often uses all three, one per action type. For instance, it might classify tickets alone, update records under monitoring and send refunds only after approval.
Table 1: Three Levels of Human Oversight for AI Agents
Dimension In the loop On the loop Out of the loop
When the human acts Before the action runs While actions run, on alerts Only when limits or reviews change Human role Approve, edit or reject Monitor, intervene, stop Set limits and audit samples Fits actions that are High impact or irreversible Moderate impact, reversible Low impact, routine, reversible Main cost Review delay and reviewer load Alert quality and attention Errors found after the fact Example Issuing a large refund Updating internal ticket fields Tagging inbound emails
Even on the loop, people need a working stop control, and they also need alerts they trust. Without good AI agent observability , monitoring turns into watching a dashboard nobody reads. In practice, on-the-loop designs fail more often from poor alerting than from poor judgment.
Which Agent Actions Need Human Approval?
The answer depends on the action, not the agent. OWASP lists excessive autonomy as a root cause of its Excessive Agency risk for LLM applications , and its mitigation is plain. Use human-in-the-loop control so a person approves high-impact actions before they are taken.
To decide what counts as high impact, score each tool call the agent can make against five factors. Then place it in a tier, and attach the tier’s oversight pattern to that tool.
Impact. What happens to money, customers or systems if this action is wrong?Reversibility. Can it be undone fully, and how fast?Uncertainty. How strong is the evidence behind the action?External exposure. Does it reach customers, partners or production?Obligation. Does a law, contract or internal policy require a person to decide?Turning Scores Into Risk Tiers
Tiering works at the tool level, which keeps it enforceable. A rule that says “the refund tool always pauses above a threshold” can be tested. A rule that says “use judgment on risky refunds” cannot.
Table 2: Agent Action Risk Tiers and Their Oversight
Tier Typical actions Oversight Human involvement
Tier 0, read only Search, retrieve records, summarize Out of the loop Audit samples only Tier 1, reversible internal Draft replies, tag tickets, update internal fields On the loop Monitor, act on alerts Tier 2, uncertain or unusual Low-confidence cases, conflicting records, policy exceptions Confidence or exception escalation Review and correct Tier 3, irreversible or external Payments, customer emails, production changes, data deletion Mandatory approval gate Approve before execution Tier 4, regulated decisions Hiring, credit and other decisions about people Approval by a qualified person Named, trained approver decides
Tiers also need least-privilege access behind them. After all, an approval gate means little if the agent holds credentials that let it skip the gate. Pair tiering with AI access control so each tool runs with only the rights its tier allows.
Industry context also moves actions between tiers. A payment agent in banking will treat far more actions as Tier 3, as the guide to agentic AI in financial services shows. Likewise, screening candidates sits in Tier 4, which is why agentic AI in HR keeps a recruiter on every final decision.
Five Human in the Loop Design Patterns for AI Agents
Once actions sit in tiers, five patterns cover nearly every oversight need. Each one has a trigger, a human job and a fixed agent behavior, so teams can test each one. Mixing them by tier gives you oversight that is strict where it counts and light everywhere else.
Approval gate. The agent stops before a Tier 3 action and waits for an explicit yes or no.Review and edit. The reviewer can change the draft or the action parameters, and the agent then resumes with the edited version.Confidence escalation. When evidence is weak, the agent routes the case to a person instead of guessing.Exception queue. Policy violations, conflicting data and failed checks land in a prioritized queue for investigation.Intervention and stop. A supervisor can halt a running agent, cancel pending actions and leave systems in a safe state.Table 3: Human in the Loop Design Patterns Compared
Pattern Trigger Human job Agent behavior
Approval gate Tier 3 or Tier 4 action proposed Approve or reject Pauses before the action Review and edit Output needs a human eye Correct, then approve Resumes with edited input Confidence escalation Evidence below threshold Resolve the ambiguity Waits for the decision Exception queue Rule breach or unusual case Investigate and decide Parks the case, moves on Intervention and stop Unsafe or unexpected behavior Halt or redirect Stops in a safe state
Why Review and Edit Beats Approve or Reject
Review and edit deserves more use than it gets. Many teams only offer approve and reject, so reviewers reject drafts that need one small fix. Letting them edit cuts rework and gives the team corrected examples for later AI agent evaluation .
Multi-agent setups also need the same patterns at the handoff points. When one agent passes work to another, the gate belongs on the agent that finally touches the outside world. The guide to AI agent orchestration covers where those handoffs sit.
Setting Confidence Thresholds and Escalation Rules
At first, confidence escalation sounds simple. If the agent is unsure, ask a person. The hard part is that “confidence” means several different things, and some of them are not reliable.
What an Agent Confidence Score Really Measures
A language model’s self-reported confidence is the weakest signal, because models can sound certain while being wrong. Stronger signals, by comparison, come from outside the model. Examples include retrieval scores, agreement between two independent checks, validation against a schema and whether the cited source actually contains the claim.
So build the escalation rule from several signals, not one number. For example, a refund request with a matching order, a valid policy clause and a consistent customer history is well supported. If any of those checks fails, the case escalates whatever the model says about itself, which also guards against AI hallucinations slipping through.
Why High Confidence Does Not Mean Auto-Approve
Confidence and impact are separate axes. A well-supported Tier 3 action is still irreversible, so it still needs approval. Confidence decides how much of the lower tiers you can automate. For high-impact actions, it only changes who reviews the case and what evidence they see.
Table 4: Confidence and Impact Decision Matrix
Evidence strength Low-impact action (Tier 0 to 1) High-impact action (Tier 3 to 4)
Strong Proceed automatically, sample for audit Proceed to the approval gate with evidence attached Moderate Proceed on the loop, flag for review Approval gate with a senior reviewer Weak or conflicting Escalate to the exception queue Stop and escalate, no action taken
Tuning Thresholds With Review Data
Start strict and loosen with evidence. In the first weeks, route more cases to people than you think you need. Then compare what reviewers did with what the agent proposed.
If reviewers approve a category of cases unchanged for a long stretch, that category can move down a tier. If they keep overriding another category, tighten its threshold or fix the prompt and tools behind it. This tuning loop is a core part of AI agent lifecycle management , because thresholds drift as data and models change.
Test the escalation paths before launch, not after. For instance, feed the agent known edge cases and confirm each one lands in the right queue. A suite built for agentic AI testing should include cases that must escalate, and the build should fail when they do not.
Checklist
Enterprise Agentic AI Checklist
A practical checklist for taking agentic AI from pilot to production, covering readiness, governance, human oversight controls and rollout.
Get the Checklist → How to Build Human in the Loop AI Agents With LangGraph
Implementation needs three things from a framework, since an approval can take hours. It must pause mid-task, save the full state and resume later from the same point. LangGraph handles this with its interrupt and resume API , which is one reason it shows up in many AI agent frameworks comparisons.
When a node calls interrupt(), the graph pauses and saves state through a checkpointer. Then the payload you pass appears under __interrupt__ in the result. You resume by invoking the graph again with Command(resume=...) and the same thread_id, and that value becomes the return value of interrupt().
A Refund Approval Gate in Code
The sample below gates refunds only when they exceed a set amount. Smaller refunds run straight through, which is the risk tiering from earlier expressed in code.
from typing import Literal, TypedDict
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.graph import END, START, StateGraph
from langgraph.types import Command, interrupt
class RefundState(TypedDict):
order_id: str
amount: float
status: str
def approval_gate(state: RefundState) -> Command[Literal["issue_refund", "cancel"]]:
if state["amount"] < 100: # low-risk tier, no human needed
return Command(goto="issue_refund")
decision = interrupt({
"action": "issue_refund",
"order_id": state["order_id"],
"amount": state["amount"],
"question": "Approve this refund?",
})
return Command(goto="issue_refund" if decision["approved"] else "cancel")
def issue_refund(state: RefundState):
# call the payments API here with an idempotency key
return {"status": "refunded"}
def cancel(state: RefundState):
return {"status": "rejected"}
builder = StateGraph(RefundState)
builder.add_node("approval_gate", approval_gate)
builder.add_node("issue_refund", issue_refund)
builder.add_node("cancel", cancel)
builder.add_edge(START, "approval_gate")
builder.add_edge("issue_refund", END)
builder.add_edge("cancel", END)
# InMemorySaver is for demos; use a durable checkpointer in production
graph = builder.compile(checkpointer=InMemorySaver())
config = {"configurable": {"thread_id": "refund-A4821"}}
result = graph.invoke({"order_id": "A4821", "amount": 640.0, "status": "pending"}, config)
print(result["__interrupt__"]) # payload for the review queue
# later, after the reviewer decides
final = graph.invoke(Command(resume={"approved": True}), config)
print(final["status"]) # refundedThree Rules That Prevent Lost Approvals
Three rules from the LangGraph docs prevent most production bugs. First, use a durable checkpointer backed by a database, since the in-memory one loses paused tasks on restart. Second, keep the same thread_id for the resume call, because it tells the checkpointer which state to load.
Third, remember that the node runs again from its first line when it resumes. Any side effect placed before interrupt() will repeat, so keep those calls idempotent or move them after the gate. The docs also warn against wrapping interrupt() in try and except blocks.
Choosing a Framework for Pause and Resume
Other runtimes follow the same shape, even if the APIs differ. Temporal’s human-in-the-loop cookbook pauses a workflow until an approval Signal arrives and cancels it on rejection or timeout. Whichever stack you pick, check six things before you commit, and the LangChain vs LangGraph comparison explains where each library fits.
Persistent state that survives restarts and deploys. A review payload with the action, its inputs and the evidence. Decision capture that records who decided, what and why. A resume path that accepts approve, edit and reject. Timeout handling with a safe default for each tier. A decision history you can query for audits and tuning. Designing the Approver Experience and Review Queue
Oversight is only as good as the screen the reviewer sees. If a reviewer gets “Approve refund? Yes or No” with no context, the human checkpoint becomes a formality. So the review card has to carry enough evidence for a real decision in under a minute.
What a Review Card Must Show
Every card needs the proposed action in plain words and the exact parameters the tool will receive. It also needs the evidence behind the proposal, the risk tier and the reason the rule fired. Put the agent’s reasoning last, so reviewers check the facts before they read the argument.
Give reviewers four buttons, not two. Approve, edit, reject and escalate cover the real choices, while each rejection should require a short reason. Those reasons become the best training signal the team has, and they also feed a defensible AI auditing framework .
Queues, Deadlines and Safe Defaults
First, route each case to someone with the authority to decide it. A refund over a set limit goes to a finance lead, while a customer email about an outage goes to the incident owner. Then sort the queue by risk tier and deadline, not by arrival time.
Every paused action also needs a deadline and a default. For Tier 3 and Tier 4, the safe default on timeout is reject or expire, never approve. For Tier 2, a timeout can escalate the case to a second reviewer instead of letting it sit.
Finally, long-running tasks need one more check. If the situation changes between approval and execution, such as a price update or a new customer message, the approval should lapse. The agent then asks again with the new facts, because approving yesterday’s proposal is not approving today’s action.
Approval Fatigue and Rubber-Stamping
Reviewers who see hundreds of routine approvals stop reading them. This is automation bias, the tendency to trust a system’s output without checking it. So watch for approvals that take only a few seconds, approval rates near one hundred percent and reviewers who never edit.
The fix is fewer, better reviews. Move consistently approved categories down a tier, and sample low-risk actions for audit instead of gating them. Occasionally insert a known-bad test case into the queue, and treat a missed one as a signal to retrain or rebalance the load.
Watch on YouTube
Enhancing Brand Compliance and Approval Workflows with Conversational AI
How Kanerika used conversational AI to speed up compliance approvals while keeping reviewers in charge of the final decision.
A KPI Scorecard for the Oversight Loop
Counting approvals tells you that reviews happened. But it does not tell you whether they changed anything. A scorecard for human in the loop AI agents should measure both the agent’s judgment and the reviewers’ attention.
Seven measures cover the loop. Track them per action type and per tier, since a blended number hides the problems. Then review them monthly with the agent owner and the risk lead together.
Table 5: Human Oversight KPI Scorecard
KPI How to measure it What it tells you Warning sign
Approval rate Approved unchanged / total reviews How often the agent’s proposal is accepted as is Near total for months, so the gate may be unnecessary Override rate Edited or rejected / total reviews How often humans correct the agent Rising after a model or prompt change Time-to-decision Request time to decision time Whether oversight slows the business Long tails that breach service deadlines Escalation rate Escalated cases / total cases How much human work the agent creates Climbing while volumes stay flat Escalation precision Escalations that needed a human / total escalations Whether the triggers are well tuned Low, which means reviewers waste time Missed escalations Incidents found later that should have escalated Gaps in the trigger rules Any repeat in Tier 3 or Tier 4 Reviewer load Reviews per reviewer per day, time per review Fatigue and staffing needs Very fast reviews at high volume
Reading the Scorecard
Read the measures in pairs. A high approval rate with very fast reviews suggests rubber-stamping, while a high override rate with slow reviews suggests the agent is not ready for that action. Low escalation precision means the triggers are too loose, and any missed escalation means they are too tight in the wrong place.
Feed these numbers into the same monitoring stack as latency and cost, so one team sees both. LLM observability tooling can also log each interrupt, decision and resume as events on the agent trace. That way one view shows what the agent did and also what the humans did about it.
EU AI Act Article 14 and Human Oversight for AI Agents
For high-risk AI systems, the EU AI Act turns human oversight into a design duty. Article 14 of Regulation (EU) 2024/1689 says such systems must be designed so natural persons can effectively oversee them while they are in use. It also says oversight measures must be commensurate with the risks, the level of autonomy and the context of use.
Whether an agent falls under Article 14 depends on whether it counts as a high-risk system under the Act, so legal review comes first. Many agent teams still use the article as a design benchmark, because its list of oversight capabilities is concrete. In particular, paragraph 4 lists five things the people assigned to oversight must be enabled to do, as appropriate and proportionate.
Table 6: EU AI Act Article 14(4) Mapped to Agent Controls
Article 14(4) capability Agent control that supports it
(a) Understand capacities and limitations, and monitor operation Documented tool scope, risk tiers and a live agent activity view (b) Remain aware of automation bias Evidence-first review cards, rubber-stamp alerts and seeded test cases (c) Correctly interpret the output Plain-language action summary with sources and the reason the rule fired (d) Decide not to use, or disregard, override or reverse the output Reject, edit and rollback actions in the review queue (e) Intervene or interrupt through a stop button or similar procedure An emergency stop that halts the agent in a safe state
Deployer Duties and Frameworks Beyond the EU
The Act also puts duties on deployers. Article 26 requires them to assign human oversight to people with the necessary competence, training and authority, as well as the necessary support. An approve button given to an untrained intern does not meet that bar, however well the software is built.
Outside the EU, the NIST AI Risk Management Framework offers voluntary guidance that points the same way. Together they make a practical baseline for AI compliance programs, especially in regulated industries. The wider policy work belongs in agentic AI governance , while this article covers the controls inside the agent.
Kanerika Service
AI Governance Services
Kanerika sets up the oversight policies, risk tiers, owners and audit evidence that let AI agents meet internal and regulatory review.
Explore AI Governance When Human Oversight Becomes the Bottleneck
Human review costs time and money, and reviewers make mistakes too. People get tired, disagree with each other and occasionally approve things they do not fully understand. As a result, oversight that ignores these limits looks good in an audit deck and catches very little.
The same seven mistakes show up again and again in agent programs. Most of them come from treating oversight as a checkbox rather than a design choice. Each one has a direct fix, listed alongside it.
Gating every action. Tier the tools and gate only Tier 3 and Tier 4.Trusting self-reported confidence. Build triggers from external checks instead.Placing the gate after the action. A review after a payment is an audit, not a control.Sending cards without evidence. Show the inputs and sources, then the reasoning.No timeout or default. Set a deadline and a safe default per tier.Treating approval as proof of correctness. Keep sampling approved actions for quality.Tracking approvals but not outcomes. Run the full KPI scorecard.How Oversight Shrinks Over Time
Done well, oversight shrinks over time. As override data proves an action type safe, it moves down a tier, and reviewers spend their attention on the cases that need it. That is the path from a cautious pilot to an agent that handles most volume alone, with responsible AI practices intact.
Case Study
82% Less Manual Processing in Vendor Agreements
A real estate developer cut manual agreement processing time by 82% and sped up vendor selection by 90% with an LLM solution built by Kanerika.
Read the Case Study → Pair the gates with technical limits as well. Rate limits, spending caps and output filters from AI guardrails catch problems that no reviewer would see in time. Human judgment then handles the gray areas, while code handles the hard boundaries.
How Kanerika Builds Human Oversight Into AI Agents
Kanerika designs oversight into the agent from the first sprint. Our agentic AI services teams follow five stages, and each one produces an artifact the client’s risk team can review before the next begins. That keeps the oversight design visible to the people who will answer for it.
Map actions. List every tool the agent can call and score it on the five risk factors.Assign patterns. Attach an approval gate, escalation rule or monitoring level to each tier.Build the loop. Implement interrupt and resume with durable state, review cards and timeouts.Test escalations. Run seeded edge cases that must reach a human, before any production traffic.Govern and tune. Review the KPI scorecard monthly and move action types between tiers on evidence.The same thinking also runs through our own agents. Klara , Kanerika’s compliance agent, checks contracts against a playbook, flags clause risks and suggests redlines. It then scores its findings by confidence and routes critical issues to a human reviewer, who keeps the final say. For wider programs, our AI governance services set the policies, ownership and evidence that sit around each agent.
Kanerika Results and Lessons From the Field
Results come from pointing human effort at the right cases. For a regulatory compliance solutions provider serving financial institutions, Kanerika built an AI compliance platform that cut manual compliance work by 60%. Regulatory response became 40% faster, and audit traceability improved fivefold.
Case Study
60% Less Manual Compliance Work with AI
A compliance solutions provider for financial institutions gained 40% faster regulatory response and 5X better audit traceability with a Kanerika-built AI platform.
Read the Case Study → The pitfalls our teams watch for are practical ones. Approval queues without owners, gates that agents can bypass through a second tool and thresholds nobody revisits after launch cause most of the trouble. For leaders sorting out where agents fit at all, an AI strategy engagement settles tiers and ownership before the first build. Complex setups with several cooperating agents, covered in our guide to multi-agent AI systems , need these controls at every handoff.
Putting the Human Where the Decision Is
So which agent actions need a human? In short, irreversible, external and regulated actions get an approval gate. Uncertain cases go to an escalation queue, and routine, reversible work runs alone under monitoring.
Build the pause with durable state, give reviewers evidence and safe defaults, and measure the loop with approval, override, time-to-decision and escalation precision. In the end, put the human where the decision is high-stakes, and let the agent handle everything else.
Frequently Asked Questions
What are human in the loop AI agents?
Human in the loop AI agents are AI agents that pause at defined decision points and hand the decision to a person before they continue. The person can approve, edit or reject the proposed action. Routine, low-risk steps run without a pause, so people only review the actions where their judgment changes the outcome.
What is the difference between human in the loop and human on the loop?
Human in the loop means a person approves an action before it runs. Human on the loop means the agent acts on its own while a person monitors it and can intervene or stop it. Most production agents use both, with in-the-loop approval for high-impact actions and on-the-loop monitoring for routine, reversible work.
Which AI agent actions need human approval?
Actions that are irreversible, reach customers or production systems, move money, delete data or make decisions about people usually need approval. Score each tool on impact, reversibility, uncertainty, external exposure and legal obligation, then gate the highest tiers. Read-only and easily reversed actions can run without case-by-case review.
How do confidence thresholds trigger human review?
A confidence threshold routes a case to a person when the evidence behind an action is weak or conflicting. Reliable triggers combine external checks, such as retrieval scores, schema validation and record matches, rather than the model’s own stated confidence. High confidence still does not skip approval for irreversible or regulated actions.
How does LangGraph support human in the loop workflows?
LangGraph provides an interrupt function that pauses a graph inside a node and saves its state through a checkpointer. The application shows the interrupt payload to a reviewer, then resumes the same thread with a Command carrying the decision. The node re-runs from its start on resume, so earlier side effects should be idempotent.
How can teams prevent approval fatigue and rubber-stamping?
Reduce the number of reviews by gating only high-risk tiers and sampling routine actions for audit. Give reviewers evidence and an edit option, not just approve and reject. Watch for very fast approvals and near-total approval rates, and seed known-bad test cases to check that reviewers still catch problems.
What metrics show whether human oversight of AI agents works?
Track approval rate, override rate, time-to-decision, escalation rate, escalation precision, missed escalations and reviewer load. Measure them per action type and risk tier. Reading them in pairs shows problems, for example high approval with very fast reviews points to rubber-stamping, while missed escalations point to weak trigger rules.
Does the EU AI Act require human oversight for AI agents?
Article 14 of the EU AI Act requires human oversight for AI systems classed as high-risk, so it applies to an agent when that agent falls into a high-risk category. It lists capabilities such as understanding limits, awareness of automation bias, overriding outputs and stopping the system in a safe state.
What should happen if a reviewer does not respond in time?
Every paused action needs a deadline and a safe default set by its risk tier. For irreversible or regulated actions, the default on timeout should be reject or expire, never approve. For uncertain lower-tier cases, a timeout can escalate the case to a second reviewer instead of leaving it waiting.
