All posts Governance & Control

Keeping Humans in the Loop: Human Oversight for Enterprise AI Agents

Human oversight is the set of controls that keep a person accountable for what an AI agent does: approval gates on consequential actions, escalation paths, the right to override or reverse an output, and a stop control that halts the agent mid-task. For a regulated European company it is not a compliance afterthought bolted on at the end. It is the design decision that determines whether an agent reaches production at all, and for the high-risk use cases listed in Annex III it becomes a legal requirement under Article 14 of the EU AI Act from 2 December 2027.

Last updated: October 28, 2026

The failure mode is well documented. Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027, naming rising costs, unclear business value and inadequate risk controls. The last of those three is a design problem, and oversight is how you close it.

Key takeaways

  • Oversight is a production gate, not a compliance tax. The agents that scale in regulated environments are the ones a business leader can approve, audit, pause, and reverse. Design the controls before you design the autonomy.
  • Article 14 turns that into law for high-risk use cases. The EU AI Act requires oversight measures commensurate with risk and level of autonomy, the power to override or reverse an output, and a “stop” procedure. After the AI Omnibus it applies from 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
  • Six major agent platforms document some human oversight, and none oversees the whole decision. Agentforce, Copilot Studio, ServiceNow, Joule, Gemini Enterprise and Bedrock each offer some mix of approval, escalation, logging and stop features. Whatever a platform approves, logs or stops, it covers only its own agents.
  • Match oversight to consequence, not to the whole workflow. Blanket human review kills the ROI; zero review kills the deployment. The durable pattern is graduated autonomy tied to the reversibility and risk of each action.
  • In-the-loop and on-the-loop are both needed. Approve-before-act for the few irreversible actions, monitor-and-intervene for the reversible volume.
  • Oversight is an engineering artifact, not a policy PDF. Approval logic, escalation and the stop control belong in the execution path, embedded by the people who build the system, not documented after the fact.

Why human oversight decides whether agents reach production

Enterprise leaders are not short on agent ambition. Gartner expects that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. The constraint is not appetite or availability. It is trust that survives contact with a risk committee.

Consider what an agent actually does in a regulated back office: it moves money, adjudicates a claim, changes a customer record, generates code that ships, or files something with a downstream authority. Every one of those actions has a consequence that someone in the organization is personally accountable for. An autonomous system with no oversight does not remove that accountability. It just detaches it from any control the accountable person can exercise. That is precisely the “inadequate risk controls” Gartner names as a leading cancellation driver.

This is why oversight is a production gate. A model can be impressive in a demo and still be un-shippable, because shipping requires a chief risk officer, a compliance lead, or a line-of-business owner to sign off on what happens when the agent is wrong. If the answer is “we would find out later, from an incident,” the project stalls in pilot regardless of accuracy. Human oversight is the mechanism that turns “the model is good” into “the business will let it run.” For a fuller treatment of why pilots stall, see our breakdown of why enterprise AI pilots fail.

What “human in the loop” actually means beyond the slogan

“Human in the loop” is used loosely enough to be meaningless. In practice, enterprise-grade oversight is four distinct capabilities, and most stalled projects are missing at least one.

Approval gates. Before an agent takes a consequential, hard-to-reverse action, it pauses and requests explicit human authorization. The design question is not whether to have gates but where to place them, because a gate on every action makes the agent slower than the manual process it replaced.

Escalation paths. When an agent hits low confidence, an edge case, or a policy boundary, it routes to a named human rather than guessing. The absence of a clean escalation path is how agents produce confident, wrong, unsupervised actions at scale.

Override and reversal. The person reviewing an output can disregard it, change it, or undo an action already taken. An approval gate without a real option to say no is a rubber stamp.

Stop controls. Operators can pause a running agent or halt a class of behavior across the fleet, and the agent stops in a state that leaves nothing half-written. This is the control that lets a leader approve deployment, because the downside is bounded.

Each of these has to leave a record, or nobody can later show it was used. What that record must contain is the subject of our piece on audit trails for AI agents; this post covers who decides.

Notice that none of these is a model capability. They are system and process design. That is the core reason oversight is an engineering problem, not a prompt-engineering one.

What enterprise AI platforms offer human oversight?

Salesforce Agentforce, Microsoft Copilot Studio, ServiceNow AI Agents, SAP Joule, Google Gemini Enterprise and Amazon Bedrock all document some human oversight for agents, each in a different form. These are the six major platforms we compared, not the whole market. Copilot Studio, ServiceNow, Gemini Enterprise and Bedrock document a step that pauses an agent until a person approves or confirms. Salesforce defines a per-action confirmation flag in Agent Script, and SAP describes human review inside named Joule agents. Whatever a platform approves, logs or stops, it covers only its own agents. It cannot see another vendor’s agent, or the decision as a whole.

The table maps three of the four capabilities above (approval, escalation and stop) and the record each must leave, as each vendor documents them on 28 October 2026. It uses only the vendors’ own pages and repositories, linked in the cells. Where a cell reads “not documented”, the feature may exist; we did not find it on a page the vendor publishes openly. The caveats below the table carry the detail.

PlatformApproval before an actionEscalation to a personAudit recordStop or isolate
Salesforce AgentforceA per-action user confirmation flag in Agent ScriptIn Agentforce Voice, the Escalation topic fires an Omni-Channel flowSession tracing in Data Cloud captures all agent activityCommand Center alerts; a stop control is not documented
Microsoft Copilot StudioPer-tool confirmation, off by default; request for information; multistage approvals in previewEscalate topic shows a message by default; live-agent transfer added by handMicrosoft Purview logs admin, maker and user activityAdmins can quarantine an agent through the Power Platform API
ServiceNow AI AgentsSupervised mode per tool; in flows, a named Support User approvesNot documentedThe kill switch audits trigger activations; action-level records not documentedThe kill switch stops runaway triggers; its default only warns
SAP JouleJoule Agents engage users to review and approve significant actionsNot documentedAssistants work while “maintaining audit trails”, inside SAPNot documented
Google Gemini EnterpriseA Workflow Builder approval step pauses the run as Needs ReviewA request info step; queue hand-off not documentedCloud Audit Logs; prompts and responses only when enabledNot documented
Amazon Bedrock Agents and AgentCorePer-action user confirmation; AgentCore temporal policies can require approvalNot documented; return of control hands the step to your applicationCloudTrail once configured; AgentCore Policy logs decisions to CloudWatchStopRuntimeSession ends a session; forbid policies block tool calls

Four caveats sit in the documentation rather than on the product pages. The first is status and defaults. Copilot Studio’s multistage approvals add manual approval stages for several stakeholders, and Microsoft’s page says “All of these capabilities are in preview.” Gemini Enterprise’s approval step works only after an administrator switches the feature on. Copilot Studio’s per-tool confirmation is set to No by default. It asks the person in the chat, as Bedrock’s user confirmation asks the end user, and that person is not necessarily the one accountable for the action. Salesforce’s Agent Script schema describes its flag as “Whether to require user confirmation before executing.” Salesforce’s own sf-skills repository lists an open issue: in Agent Script, the flag compiles but produces no confirmation at runtime. Google’s Agent Development Kit labels its tool confirmation feature experimental. A preview or unproven feature is fine to pilot and weak evidence for a risk committee.

The second is that the audit column depends on configuration. Bedrock records InvokeAgent calls in CloudTrail only after someone configures advanced event selectors for data events. Gemini Enterprise logs the fields that can hold prompts and responses only when “Enable logging of prompt inputs and response outputs” is turned on (Google Cloud). Microsoft notes that some Copilot Studio channels are excluded from logging. Where each platform logs is a different question from what the log must contain, which our guide to audit trails for AI agents covers.

The third is that the stop column is the thinnest. ServiceNow’s kill switch watches AI agent triggers that fire again and again on the same records, to stop wasted Now Assist consumption. Its shipped default, warn_only, sends warning emails and never deactivates a trigger. That is a cost guard, not a person deciding to halt an agent. Microsoft’s quarantine isolates a noncompliant agent while security or compliance issues are investigated, and cuts it off from every channel except maker testing. AgentCore’s StopRuntimeSession ends one running session. Gemini Enterprise limits agents to the connectors and tools an administrator provisions, which bounds their scope but is not a stop. None of these reaches work the agent has already handed to another system.

The fourth is that some oversight lives in agent behavior rather than in a setting. In SAP’s Q3 2026 release highlights, the Material Reservation Processing Agent comes “with built-in human review”, and another agent “executes corrections only after human approval”. The SAP pages we read document no escalation or stop control. Salesforce’s escalation has a limit of its own. In the default escalation for Agentforce Voice, Salesforce’s blog notes that the classifier bypasses the topic’s instructions, so the agent cannot collect data or run tools before the hand-off. Dynamic escalation needs the canEscalate flag set on the topic. Salesforce’s architecture guidance keeps the human in the loop “for decisions, approvals, and strategic guidance”. Joule’s record stops where SAP stops, a gap we cover in SAP Joule’s limitations outside SAP.

What happens when an agent triggers a human-in-the-loop approval?

The agent pauses, and where it waits depends on the platform. ServiceNow holds a supervised tool until a person approves it; in a flow, it waits for the named Support User in the Now Assist panel. In a background or unattended run, ServiceNow says the tool “stalls rather than proceeds.” Copilot Studio approvals reach approvers in the Microsoft Teams approvals app, Outlook or the Power Automate portal. Its request-for-information step goes by Outlook only, inside your tenant, and the flow uses the first reply. Gemini Enterprise shows the run as Needs Review in the Runs tab. Bedrock returns the proposed action to your application, and if the user answers DENY, the action is not run.

All four behave as the in-the-loop column of the human-in-the-loop vs human-on-the-loop table predicts: if nobody responds, nothing happens. That is safe and slow, and a queue nobody watches becomes a backlog. AWS’s Well-Architected guidance names the failure: approval workflows “without timeout policies or escalation paths, so agent execution stalls indefinitely when reviewers are unavailable.” It also asks for approvals logged with timestamps and reviewer identities. The same page warns against the opposite error: “Routing every agent action through human review produces rubber-stamp approvals.” The platforms supply the pause. Which actions get one is the graduated-autonomy question further down.

Can one platform oversee an agent decision that spans several vendors?

No. Whatever a platform approves, logs or stops, it covers only its own agents. It cannot see another vendor’s agent, or the decision as a whole.

The scoping is in the vendors’ own documentation. AgentCore Policy evaluates agent traffic that passes through AgentCore Gateways, and its temporal rules look at the history of an agent’s actions within one session. Copilot Studio’s audit records sit in Microsoft Purview. Gemini Enterprise’s Cloud Audit Logs track admin and data-access operations on Gemini Enterprise resources. ServiceNow’s kill switch watches ServiceNow triggers, and SAP’s audit trail covers the steps Joule runs in SAP. A platform’s own agent can still gate its calls into other systems, through a per-tool confirmation on a Copilot Studio connector or Bedrock’s return of control. What it cannot see is the agent another vendor runs on the same case. None of this is a defect. Each vendor governs what it can see.

Take the billing dispute that Nobody owns the handoffs uses as its cross-vendor case, a generic example and not a client one: raised in Salesforce, parked in ServiceNow and settled as a credit note in SAP, with each vendor’s agent handling its own step. Put the three oversight questions to it. Who approves: a person steps in at three points (an Omni-Channel hand-off, the Now Assist panel, an SAP approval), and nobody approves the credit as a whole. Where the record lives: in three vendors’ logs, each holding its own step, with no record of the decision as a whole. What stops the work: quarantining or stopping one agent leaves the other two running, and neither of them knows the first was stopped. An invoice dispute is not an Annex III use case, so Article 14 does not bind this process. Its test is still the right design standard: a stop that lets the work “come to a halt in a safe state”. Here the work spans three vendors, and no single stop reaches all of it.

SAP says AI Agent Hub helps organizations govern every agent “across SAP and non-SAP environments”, and Nobody owns the handoffs compares it with four other vendors’ registries. A registry keeps an inventory of agents, but it does not approve a decision or halt a process that runs across several vendors’ agents, so it does not close the gap. Monitoring ends at the same edge: when Salesforce launched Command Center in June 2025, Constellation Research analyst Martin Schneider questioned how well it would measure multi-agent flows that use agents from other platforms.

Cross-vendor oversight belongs in a layer the company owns. That layer answers the three questions once for the whole decision: the approval policy for each action tier, the decision record across systems, and a stop that reaches every system an agent touches. It is part of the context engine the company runs on, and it is model-agnostic, so it outlasts a change of vendor or model. The price is real. Someone has to operate the layer and keep its approval policies current as processes change, and it needs an owner in the business as well as in IT.

Human-in-the-loop vs human-on-the-loop

The two terms describe where the person sits relative to the action. In the loop, the agent cannot proceed until a human acts. On the loop, the agent proceeds and a human watches with the power to step in. Neither is better in general; each fits a different kind of action.

Human-in-the-loopHuman-on-the-loop
When the human engagesBefore the action executesWhile or after the action runs
What the human doesApproves, edits, or rejects a prepared actionMonitors, investigates alerts, intervenes or halts
Default if nobody respondsNothing happensThe action goes ahead
FitsIrreversible or high-consequence steps: payments, customer commitments, regulatory filingsReversible, lower-consequence, high-volume steps: classification, drafting, internal routing
Main riskRubber-stamping out of fatigue, the over-reliance Article 14 calls automation biasErrors that run until someone notices them
Engineering it needsA gate in the control flow and a queue with a named ownerMonitoring, alert thresholds, rollback and a stop control

The row that decides most designs is the default. If a missed approval means nothing happens, in-the-loop is safe but slow. If a missed alert means the action goes ahead, on-the-loop is fast but only acceptable when the action can be reversed. That is why the choice belongs to each action, not to the whole workflow.

Graduated autonomy: matching oversight to consequence

The instinct to require human review of everything is understandable and fatal. It reintroduces the labor cost the agent was supposed to remove and guarantees a negative ROI. The opposite instinct, full autonomy, is what Gartner’s cancellation forecast warns against. The workable answer is graduated autonomy: the level of human oversight scales with the risk and reversibility of the specific action, not the workflow as a whole.

A practical way to think about the tiers:

  • Autonomous. Low-consequence, easily reversible, high-confidence actions run without review, logged for later audit. Drafting an internal summary or categorizing a ticket belongs here.
  • Notify-and-proceed (on the loop). The agent acts but surfaces the action to a human who can intervene within a window. Suitable for medium-consequence, reversible actions.
  • Approve-before-act (in the loop). The agent prepares the action and waits for explicit sign-off. This is where money movement, customer-facing commitments, and regulatory filings live.
  • Human-owned, agent-assisted. The human makes the decision; the agent gathers evidence and drafts options. Reserved for the highest-stakes, least-reversible calls.

The engineering work is deciding which tier each action falls into, and building the routing so the agent moves through tiers correctly. Done well, most volume runs autonomously while human attention concentrates on the small fraction of actions where judgment actually pays. This is also the design that lets you demonstrate ROI honestly rather than through demo theater; we cover the measurement side in our 2026 benchmarks for enterprise AI agent ROI.

EU AI Act Article 14: what human oversight means in law

For a European mid-cap, graduated autonomy is no longer only good practice. Article 14 of the AI Act requires high-risk AI systems to be designed so that natural persons can effectively oversee them while they are in use, and it says the oversight measures must be commensurate with the system’s risks, level of autonomy and context of use. That last clause is the tiering above, written as a legal duty.

The application dates changed in July 2026. Under Article 113 as first adopted, the rules for the high-risk use cases listed in Annex III were due from 2 August 2026, and the rules for high-risk AI embedded in products covered by Annex I from 2 August 2027. Regulation (EU) 2026/1744, the AI Omnibus, entered into force on 27 July 2026 and moved both back. According to the European Commission, the Annex III rules now apply from 2 December 2027 and the Annex I rules from 2 August 2028. Annex III includes CV-sorting software for recruitment, a workflow any employer may automate. Whether a given agent is high-risk depends on its use case, not on the fact that it is an agent.

Article 14 is written for the provider, who must build oversight into the system. Article 26 puts a second duty on the deployer, who must assign oversight to natural persons with the necessary competence, training and authority, and support. A company that builds a system and puts it into service under its own name meets the Act’s definition of a provider, so a mid-cap building its own agent for its own workflow can hold both sets of duties at once.

Platform features are parts of that design, not compliance with it. Article 14(3) says oversight is ensured through measures “identified and built, when technically feasible, into the high-risk AI system by the provider”, or through measures the provider identifies for the deployer to implement. A supervised-mode setting or an approval step is a component the provider can use. Which actions get one, who approves, and how the system halts are still the provider’s decisions. A mid-cap that builds an agent on Copilot Studio or Bedrock and puts it into service under its own name is usually that provider, as the paragraph above sets out.

Article 14(4) lists what the people assigned to oversight must be able to do. Each item maps to something an engineer has to build:

Article 14(4) powerWhat it means in practiceThe control that delivers it
(a) Understand the system’s capacities and limits, and monitor itThe reviewer knows what the agent is for, what it is bad at, and sees anomalies as they happenDocumented scope per action, live monitoring, alerts on drift and unexpected behavior
(b) Stay aware of automation biasThe reviewer is not trained by the interface to click “approve”Show the evidence behind a recommendation, not just the recommendation; sample approved items for second review
(c) Correctly interpret the outputThe reviewer can tell why the agent proposed what it didReasons, sources and confidence shown next to every proposed action
(d) Decide not to use the system, or disregard, override or reverse its outputSaying no is a real option, and undoing is possibleApproval gates with reject and edit paths, rollback for executed actions
(e) Intervene, or interrupt through a “stop” button or similar, halting in a safe stateSomeone can stop the agent now, without corrupting the work in flightA stop control per agent and per action class, with transactions that leave no half-finished state

The obligations do not start until December 2027 at the earliest, but none of these controls can be added in the month before. An approval gate changes the control flow; a safe stop changes how actions are committed. The extra time is the window to build them into the system. For the full high-risk picture, including how to classify a use case, the Commission’s regulatory framework page and the Omnibus text in the Official Journal are the primary sources.

Oversight is an engineering artifact, not a policy document

A recurring reason oversight fails is that it lives in the wrong place. Governance frameworks, responsible-AI policies, and risk memos are necessary, but a PDF does not stop an agent from taking a wrong action. Policy enforcement has to be part of execution: the approval gate is code, the escalation rule is a branch in the agent’s control flow, the stop control is an operational lever someone can actually pull, and every one of them writes to the log.

This is where the delivery model matters. When oversight is designed after the agent is built, it becomes a layer of friction retrofitted onto a system that assumed autonomy. When it is designed alongside the agent, by engineers embedded in the actual workflow, the controls fit the way the business really operates. That is the reason we work embedded; see what a forward-deployed engineer does in enterprise AI. The engineer who understands the claims process, the settlement rules, or the code-review standard is the one who knows where the approval gate belongs and which actions must never run unattended.

There is a durable business benefit here too. When oversight is instrumented, the record of approvals and overrides shows how the work is actually done: which decisions require human judgment, which policies apply, where the edge cases hide. That captured knowledge outlasts the individuals who currently hold it in their heads, and it stays with the company rather than with whoever built the system.

What decision-makers should require before approving an agent

If you are the executive being asked to approve an agent for production, oversight gives you a concrete checklist that cuts through vendor “agent washing” (Gartner’s term for products rebranded as agents without real capability):

  • Where are the approval gates, and are they on the actions my risk and compliance teams care about?
  • Which actions run on the loop, and can each of them be reversed?
  • What happens on low confidence: does it escalate to a named human, or does it guess?
  • Can I stop the agent, how fast, and in what state does it leave the work?
  • Who is accountable for each tier of action, do they have the competence and authority Article 26 asks for, and does the tiering match our actual risk tolerance?
  • Is every approval, override and stop recorded, so we can prove the oversight happened?
  • Is the oversight embedded in the system, or described in a document that the running agent has no awareness of?
  • If the workflow crosses vendors, where does the approval for the whole decision live, and can one stop reach every system it touches?

An agent that answers these cleanly is one a regulated company can actually run. An agent that cannot is a pilot, no matter how good the demo looked.

Nucleo works with European mid-caps to rebuild the workflows where the value is stuck around AI. Our forward-deployed engineers build the approval gates, the cross-system decision record and the stop controls into a context engine the client owns, with the data kept in the EU and the system staying with the client when we leave. If oversight is what your risk committee needs to see before a workflow changes, see how we work or write to us.

Frequently asked questions

What is human oversight for enterprise AI agents?

Human oversight is the combination of controls that keep an accountable person in command of an AI agent’s actions: approval gates on consequential steps, escalation to humans on uncertainty, the ability to override or reverse an output, and a stop control that halts the agent in a safe state. It ensures that when an agent acts, a named person decided whether it could, and can step in when it should not, which is what a risk committee needs before it lets an agent into production.

What does EU AI Act Article 14 require?

Article 14 requires high-risk AI systems to be designed so natural persons can effectively oversee them while they are in use. The measures must be commensurate with the system’s risks, level of autonomy and context of use. The people assigned to oversight must be able to understand the system’s capacities and limitations and monitor it, stay aware of automation bias, interpret its output, decide not to use it or disregard, override or reverse its output, and intervene or interrupt it through a stop button or similar procedure that brings it to a halt in a safe state. Article 26 adds that deployers must assign oversight to people with the necessary competence, training and authority.

When does Article 14 human oversight apply?

Under the AI Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, Article 14 applies from 2 December 2027 for the high-risk use cases listed in Annex III, such as CV-sorting for recruitment, and from 2 August 2028 for high-risk AI embedded in products covered by Annex I, according to the European Commission and the amended Article 113. Whether a given agent is high-risk depends on its use case, not on the fact that it is an agent.

What is the difference between human-in-the-loop and human-on-the-loop?

Human-in-the-loop means a person approves or performs a step before the agent proceeds, which suits high-consequence, hard-to-reverse actions like money movement or regulatory filings. Human-on-the-loop means the agent acts on its own while a person monitors and can intervene or halt it, which suits lower-consequence, reversible actions. Mature deployments use both and assign each action to a mode based on its risk, which is also what Article 14 asks for when it says oversight must be commensurate with risk and level of autonomy.

Does keeping humans in the loop cancel out the ROI of AI agents?

No, provided oversight is graduated rather than blanket. Requiring a human to review every action does erase the efficiency gain. The workable model ties the level of review to the risk and reversibility of each specific action, so most low-consequence work runs autonomously with logging, while human attention concentrates on the small fraction of high-stakes, hard-to-reverse decisions where judgment actually adds value.

How do audit trails support human oversight?

Oversight decides who may approve which action. The audit trail is the evidence that the decision was actually made: every approval, override and stop is recorded with who decided, when and on what information. Without that record, an organization can describe its oversight but cannot prove it to an examiner. What the trail must capture, field by field, is a separate design question with its own requirements under Article 12 of the AI Act, covered in our guide to audit trails for AI agents.

Should human oversight be a policy document or built into the system?

It should be built into the system. Policies and governance frameworks are necessary context, but a document cannot stop an agent from taking a wrong action. Effective oversight lives in the execution path: approval gates as code, escalation as control-flow logic, and stop controls as operational levers. Oversight designed alongside the agent, by engineers embedded in the workflow, fits the business far better than controls retrofitted after the fact.

Which AI agent platforms have built-in human oversight?

Six major enterprise agent platforms document some human oversight: Salesforce Agentforce, Microsoft Copilot Studio, ServiceNow AI Agents, SAP Joule, Google Gemini Enterprise and Amazon Bedrock. Copilot Studio has a per-tool confirmation, a request-for-information step and multistage approvals, the last still in preview. ServiceNow runs a tool in supervised mode until a person approves it. Gemini Enterprise has an approval step in Workflow Builder. Bedrock Agents ask the user to confirm or deny an action, and AgentCore policies can require an approval first. Salesforce defines a per-action user confirmation flag in Agent Script and documents escalation to a person. SAP describes human review inside named Joule agents. Whatever a platform approves, logs or stops, it covers only its own agents. It cannot see another vendor’s agent, or the decision as a whole. The platform table above links each vendor’s documentation.

What happens when an AI agent triggers a human-in-the-loop approval?

The agent pauses, and where it waits depends on the platform. ServiceNow waits for a person to approve the tool. In a flow, that is the named Support User in the Now Assist panel. In an unattended run, the tool stalls. Copilot Studio approvals reach approvers in the Microsoft Teams approvals app, Outlook or the Power Automate portal; its request-for-information step goes by Outlook only, inside your tenant, and the flow uses the first reply. Gemini Enterprise shows the run as Needs Review in the Runs tab. Amazon Bedrock returns the proposed action to your application, which answers CONFIRM or DENY; on DENY the action does not run. In each case, if nobody responds, nothing happens, so every approval queue needs a named owner and an escalation path.

Can one platform oversee AI agents across SAP, Salesforce and ServiceNow?

No. Whatever a platform approves, logs or stops, it covers only its own agents. It cannot see another vendor’s agent, or the decision as a whole. AgentCore Policy sees traffic through AgentCore Gateways, ServiceNow’s kill switch watches ServiceNow triggers, and Joule’s audit trail covers the SAP steps. Cross-vendor registries such as SAP AI Agent Hub keep an inventory of agents, but they do not approve a decision or halt a process that runs across several vendors’ agents. Oversight of a decision that spans vendors needs a layer the company owns, holding the approval policy, the decision record across systems and a stop that reaches every system involved. Nobody owns the handoffs covers the registries and the decision record in detail.

Does building an agent on one of these platforms make it compliant with Article 14?

No. For a high-risk use case, Article 14(3) puts the oversight measures on the provider, either built into the system or identified for the deployer to implement. A company that builds an agent on Copilot Studio, Bedrock or another platform and puts it into service under its own name is usually the provider. The platform supplies parts, such as an approval step or a stop API. Which actions need them, who approves, and how the system halts in a safe state remain the company’s design.

The future belongs to those who see it before it is obvious, and build it while everyone else is still out there scouting for the best AI tools.