Building Trust With Agentic AI From Pindrop: What the Approach Is, What the Products Do and What to Ask Before You Buy

At a glance

• The idea: Pindrop’s view is that trust in the agentic era can’t rest on identity alone. It frames the problem as three questions: is this the right entity, does it have the right intent, and is the right action being enabled?
• Two angles: Pindrop says it helps enterprises defend against third-party AI agents that impersonate people, and it applies agentic AI inside its own fraud-investigation workflows.
• Newest product: BotStopper, announced Sept. 16, 2026, detects AI voice agents and automated callers in real time. Pindrop describes it as a first step toward “know your agent.”
• Pindrop’s platform: Three engines cover synthetic detection (Pulse, BotStopper), risk (Protect, Fraud Assist) and authentication (Passport, VeriCall).
• Read the numbers carefully: Most figures come from Pindrop’s own releases and research. They’re useful context, not independent proof.
• The open problem: Trust breaks down when agents cross company boundaries, and no single vendor has solved that yet.

Why This Topic Matters Now

A few years ago, “trust” in a phone call or video meeting mostly meant checking that a person was who they claimed to be. Today the other party might be a person, a cloned voice, a script, or an AI agent acting for a real customer. Some of those are welcome. Others are not.

That shift is the subject of this guide. It looks at how Pindrop, a voice-security and deepfake-detection company, describes building trust with agentic AI, what its products do about it, and what readers should check before relying on any of it. Our review covered Pindrop’s official articles, announcements, and platform documentation, as well as trade media reports about a Pindrop webinar. This is an analysis of those public documents, not a hands-on evaluation, and we haven’t carried out any product assessment ourselves. Where a number comes from Pindrop, we say so.

What “Agentic AI” Changes About Trust

From answering to acting

A chatbot answers questions. An agent takes actions: it can check a balance, file a claim, schedule a meeting or negotiate on someone’s behalf. Pindrop’s June 2026 interview with its Chief Product Officer, Nicholas Holland, describes agents that may verify insurance benefits, resolve support requests and complete administrative tasks. He adds that some will create real value while others may be deceptive, fraudulent or malicious.

To understand how autonomous systems differ from conventional conversational tools, see our guide to AI agents vs. chatbots, including their architecture, capabilities, and security implications.

Pindrop’s September 2026 BotStopper release puts the same shift plainly: AI agents are no longer just calling humans, they’re calling enterprise systems directly. Most contact centers, it says, can’t tell a legitimate agent from a malicious bot.

Why voice is an exposed channel

Voice and live video feel trustworthy. During a June 2025 webinar, Pindrop’s Mo Merchant highlighted how widely available voice-generation tools have become. As reported by Biometric Update, he cited more than 2,400 text-to-speech engines discoverable through a simple search and described deepfakes capable of convincing real-time interactions. Amit Gupta, Pindrop’s VP of Deepfake Detection, said fraud rings that once needed hundreds of people can now automate much of the work, and that organizations should assume real-time communication is at risk from deepfakes.

That’s a vendor’s view, and vendors have an interest in stressing the threat. But it matches a broader pattern: cheap synthetic voices and automation make impersonation easier to scale.

Who Pindrop Is and Where It Fits

Pindrop describes itself as a cybersecurity company providing continuous identity verification and deepfake detection across voice, video and digital communications. Its platform page calls its offering an “identity trust platform” for real-time communications. The company reports that its Pindrop Intelligence Network (PIN) draws on more than 8 billion analyzed interactions, 2,300+ detection signals, 1,000+ modeled AI tools, 2.7 million known bad actors and 5,000+ known AI voices. It also says it’s trusted by 14 of the top 20 banks and five of the top seven life insurers, and that TIME named it one of 2026’s most influential software companies. All of those are company statements.

The three engines

Engine	Question it answers	Products
Synthetic engine	Is this identity synthetic?	Pulse for Contact Centers, Pulse for Meetings, BotStopper
Risk engine	Is this identity risky?	Protect, Fraud Assist
Authentication engine	Is this identity known?	Passport, VeriCall

Pindrop’s chief product officer said the company wants customers to view its offerings as a single, integrated platform rather than a collection of individual products. This is important for buyers because the underlying technologies are built to work together, even though some are available as standalone solutions.

Pindrop’s Framework: Entity, Intent and Action

The core of the company’s thinking on building trust with agentic AI is a three-part test, as described in the Holland interview.

Question	What it asks	Why identity alone isn’t enough
Is this the right entity?	Is the caller, participant or agent who or what it claims to be?	Synthetic voices and identities can pass static checks
Does it have the right intent?	Is it behaving in a way consistent with a legitimate purpose?	An authentic account can still be used for fraud or reconnaissance
Are we enabling the right action?	Should this entity be allowed to do this particular thing now?	Permission depends on context, not just a credential

Holland’s summary of the era is that the last decade connected people and systems, while the next is about deciding what and who you can trust.

Two complementary angles

Pindrop says it approaches agentic security in two ways.

1. Defending against external agents. Third-party AI agents can use voice, video and other channels to impersonate people, manipulate interactions and commit fraud. Pindrop’s detection products target that.
2. Using agents internally. The company applies agentic AI in its own fraud-detection and investigation workflows. Fraud Assist already supports analysts by reviewing calls, summarizing activity, and helping investigations move faster. Pindrop says future security agents will go further by linking signals across cases over extended periods.

The cross-company gap

Holland’s most useful point is about scope. Organizations can govern agents inside their own walls, he says, but trust breaks down when agents interact across enterprises. That’s the gap Pindrop says it wants to fill. It’s also the hardest part, since it depends on shared standards and registries as much as on any single product.

The Threat Picture, as Pindrop Reports It

These figures come from Pindrop’s own releases, research and webinars. We haven’t been able to verify the underlying datasets, so treat them as company-reported.

Claim	Source	How to read it
AI-powered attacks up nearly 1,680% since late 2024, across more than 700 million customer interactions	Pindrop BotStopper release, Sept. 2026	Company research; the definition of “AI-powered attack” isn’t spelled out in the release
IDC projects 1.3 billion AI agents by 2028	Cited in the same release	A third-party projection quoted by Pindrop
Pindrop expected a 162% rise in deepfake fraud in 2025; 1 in 599 calls was fraud	Webinar coverage by Biometric Update, June 2025	Company forecast and statistic from 2025
1 in 6 engineering candidates showed signs of fraud	Pindrop platform page	Sample size and method aren’t stated on the page
At one healthcare organization, bot calls peaked at 6,668 in a month; bot activity fell 94.3% within four months	BotStopper release	A single customer example from the vendor

The pattern across all of them is consistent: rapid growth in automated and synthetic interaction. The precise percentages are less reliable than the direction.

What the Specifications Tell Us: Pindrop’s Products Through an Agent-Trust Lens

Below, each product is described using Pindrop’s own pages. Our discussion of “limits and questions” represents independent analysis, not the company’s official position.

Pindrop Pulse for Contact Centers

What it does: Identifies synthetic speech in phone interactions. Pindrop’s release describes Pulse as the detection engine that answers whether a caller is a human or a machine in about two seconds.

Strengths as described: Real-time analysis, integration with contact-center platforms such as Genesys, Five9, Amazon Connect, Cisco Webex and NICE, and access to PIN’s shared intelligence.

Limits and questions: Accuracy figures are vendor-reported. Ask how performance is measured on your own call audio, how false positives are handled, and what happens to a call flagged as synthetic. Detection tells you something is synthetic. It doesn’t decide what that entity is allowed to do.

Best for: Contact centers worried about cloned-voice impersonation.

Pindrop Pulse for Meetings

What it does: Detects deepfakes in virtual meetings. Pindrop says it’s adding continuous video analysis during screen sharing and integrations with hiring platforms that flag proxy candidates and identity swaps, with results written back to the candidate record. It works with meeting platforms including Zoom, Webex and Microsoft Teams.

Strengths as described: An audit trail of what was detected, and coverage of remote hiring and workforce interactions.

Limits and questions: Pindrop’s own FAQ says a detection signal should inform further verification and human review, not replace it. Consider candidate privacy, notice and consent, and how you’ll treat a false alarm.

For a broader discussion of privacy, human oversight, and monitoring in recruitment workflows, see our guide to AI tools for HR.

Best for: Organizations that hire remotely or run high-stakes meetings where impersonation is a risk.

Pindrop BotStopper

What it does: This standalone solution spots AI agents and automated callers as they connect to a contact center. It combines Pulse technology with the AI Voice Consortium, a registry managed by Pindrop that includes more than 5,000 AI voices. The registry helps the system identify known agents instead of only flagging synthetic voices.

Strengths as described: It’s available to any enterprise without another Pindrop product, and Pindrop says existing Pulse for Contact Centers customers receive it and access to the consortium at no added cost.

Limits and questions: Pindrop calls it step one toward agentic identity. It answers “is it a machine?” Pindrop describes these follow-up questions—what agent is involved, who it acts for, and what it can do—as areas it plans to address. The registry recognizes known voices, so ask how new or unregistered agents are treated.

Best for: Contact centers in healthcare, financial services, insurance and retail that are seeing automated traffic.

Pindrop Protect

What it does: Assesses fraud risk during live interactions, using behavioral, acoustic, network and device signals through a call, according to Pindrop’s FAQ.

Strengths as described: Context beyond a caller’s answers to security questions, supported by 2,300+ signals and a database of known bad actors.

Limits and questions: Risk scores need decision rules and workflows. Ask who reviews flagged calls and how thresholds are set and tuned.

Best for: Fraud and risk teams that need continuous scoring, not a single check at the start of a call.

Fraud Assist and the New Agents

What it does: An optional add-on for Protect that adds AI-assisted fraud investigation workflows. At its September 2026 showcase, Pindrop introduced a Fraud Investigation Agent to help analysts make sense of evidence and uncover relationships, as well as an Enrollment Agent that uses Pindrop intelligence and enterprise rules to promote consistency in enrollment decisions.

Strengths as described: Pindrop says enterprise policies and human judgment remain in control.

Limits and questions: The new agents were presented as capabilities being showcased, so ask what’s generally available, how outputs are explained and how analysts can override them.

Best for: Fraud teams with heavy investigation volume.

Pindrop Passport

What it does: Verifies recognized callers by comparing them with their previous enrollment data, reducing the need for passwords and security questions.

Strengths as described: Less friction. Pindrop’s platform page says Passport can reduce average member authentication time by up to 50% and lift NPS by up to 10 points, citing a 2026 customer story.

Limits and questions: That’s a single customer’s reported result, so ask for results from comparable environments. Enrollment is required, and biometric rules apply.

Best for: High-volume contact centers where authentication adds noticeable time and frustration.

VeriCall Technology

What it does: An optional add-on for Passport, also available standalone, that adds ANI validation and spoof detection. Pindrop’s public pages give little detail beyond that, so ask for technical documentation.

Best for: Teams that need caller-number validation alongside authentication.

Feature Comparison: How These Products Compare

Product	Trust question	Channel	Role with AI agents	Maturity note
Pulse for Contact Centers	Is it synthetic?	Phone	Flags synthetic speech	Described as already in enterprise use
Pulse for Meetings	Is it synthetic?	Video and audio meetings	Flags deepfakes and proxy candidates	Expanding with hiring integrations
BotStopper	Human or machine?	Phone	Detects automated callers and known AI voices	Launched Sept. 2026; described as a first step
Protect	Is it risky?	Contact center	Scores behavior and context	Established product
Fraud Assist and agents	What happened and why?	Investigations	Pindrop’s own AI agents assist analysts	Some capabilities newly showcased
Passport	Is it a known person?	Phone	Authenticates humans	Established product
VeriCall	Is the number spoofed?	Phone	ANI validation	Limited public detail

A Closer Look at BotStopper

BotStopper is the product most directly tied to the agentic shift, so it deserves a closer reading.

What’s new

Pindrop lists three new elements: it’s standalone, it’s powered by the AI Voice Consortium, and it’s positioned as step one toward agentic identity. The CEO, Vijay Balasubramaniyan, framed the philosophy this way: the goal is not to block AI agents, but to know which agent is calling, who it represents and what it should be allowed to do.

That framing is useful. A blanket ban on automated callers would also stop legitimate agents acting for real customers. The harder and more valuable work is sorting them.

The reported results

In Pindrop’s healthcare example, the organization identified more than 30,000 bot calls in under a year. Bot calls peaked at 6,668 in one month, when roughly $15.1 million in HSA balances was exposed to risk, and activity fell 94.3% within four months of real-time bot detection. Pindrop says an Optum Financial team also reported benefits for its advocates and members. These are company-published customer results, and they describe a single organization.

The warranty

Pindrop says new customers who buy BotStopper as part of the full suite on a three-year subscription may qualify for up to $1 million in reimbursement under its Pulse Warranty, with terms and conditions. That’s a significant commitment, but the details are important. Review the warranty terms before counting on it.

Published Benchmark Comparison: What to Make of the Accuracy Claims

Pindrop’s release says that in a June 2026 benchmark of audio deepfake detection systems by Podonos, Pulse was one of only four systems to exceed 95% accuracy, and was identified as the fastest accurate detector evaluated. We couldn’t open the Podonos report, so we can’t confirm those details. Treat it as a vendor-cited third-party result.

When you read any detection benchmark, a few questions help:

Question	Why it matters
Which audio and attack types were included?	A test set can leave out the latest synthetic voices
How was accuracy defined?	Overall accuracy can hide high false-positive or false-negative rates
Was the test run on phone-quality audio?	Compression and noise change results
Who funded or ran the evaluation?	Independence affects how much weight to give it
How often is the detector updated?	New voice tools appear continuously
What’s the false-positive cost for you?	Blocking a real customer is a business harm

A strong benchmark is a reason to shortlist a product, not a reason to skip an evaluation on your own traffic.

The Human Approval Problem

One of Pindrop’s more distinctive arguments appears in an April 2026 article by Elie Khoury, its VP of Research, responding to NIST’s National Cybersecurity Center of Excellence concept paper on software and AI agent identity and authorization.

The core argument

Pindrop agrees that agent identity matters. Its point is that identity for the agent isn’t enough, because an agent acting on delegated authority depends on a human approval event at the start. If that approval is faked, every downstream action can look valid while being untrustworthy. In phone calls and video meetings, synthetic media can convincingly simulate human presence.

Four implications Pindrop outlines

Implication	What it means in practice
Treat human approvals as identity events	Record who approved a delegation, how, at what assurance level and under what conditions
Build delegation provenance chains	Link the human authorization, the delegation policy and the agent’s later actions for audit
Include synthetic impersonation in assurance models	A strong machine identity doesn’t help if a fake human can authorize a sensitive action
Make trust adaptive	Authorization should change when an agent’s tools, behavior, context or accessed data changes

Pindrop also argues that the future of AI trust isn’t just bot detection, since enterprises depend on automation daily. The real question is whether an agent is authorized, governed and trustworthy for this action, in this moment, under these conditions.

This is Pindrop’s perspective, and the article states it isn’t security advice tailored to any organization. It’s still a sensible way to think about agent security, and it overlaps with independent frameworks that stress least-privilege access and accountability.

How Pindrop’s View Compares With Other Voices in the Same Conversation

The June 2025 webinar coverage placed Pindrop alongside two other companies working on identity, and the comparison helps show where each focuses.

Company	Focus, as described in the coverage	Core idea
Pindrop	Real-time deepfake detection and risk in voice interactions	Assume real-time communication is at risk and verify it
Anonybit	Biometric-bound identity	Link an AI agent’s actions to the human who authorized it through a time-bound biometric signature, which its CEO calls a “Circle of Identity”
Validsoft	Real-time biometric authentication for voice	Without identity assurance, malicious instructions to AI systems become more likely

These are separate companies with separate products. Several SEO-style pages online describe a combined “agentic AI Pindrop Anonybit” system with specific accuracy numbers and architecture. We found no Pindrop or Anonybit document describing a joint product, so treat those pages as unverified. All three companies emphasize the importance of verifying the person behind an AI agent and ensuring the agent’s actions can be traced back to that individual.

Our Analysis: What a Practical Agent-Trust Model Looks Like

Drawing on Pindrop’s materials and the NIST discussion, here’s how we’d structure agent trust in practice. This is our interpretation, not a Pindrop product description.

Layer	Question	Example controls	Typical owner
Entity	Is this a human, a synthetic voice or an agent?	Synthetic-speech detection, bot detection, AI-voice registries	Security / contact center
Identity	Is it who it claims to be?	Passive authentication, enrollment, device and network signals	Security / fraud
Delegation	Who authorized the agent, and how?	Verified human approval, logged delegation chain	Identity / governance
Authorization	What may it do right now?	Least-privilege scopes, step-up checks for sensitive actions	Application owners
Monitoring	Is behavior still consistent?	Continuous risk scoring, anomaly alerts	Fraud / SOC
Audit	Can we reconstruct what happened?	Evidence logs, case records, human override	Compliance

Strengths of Pindrop’s position

• Depth in real-time voice. Its roots and its data are in voice and contact centers.
• A clear conceptual frame. Entity, intent and action is easy to explain and to turn into policy questions.
• Attention to human approval. Many agent-security discussions focus on machine credentials, so this is a useful addition.
• Policy-aware use of its own agents. Pindrop says human judgment and enterprise policy stay in control.

Limits worth keeping in mind

• Company-reported evidence. Most figures come from Pindrop’s own releases.
• “Know your agent” is early. By Pindrop’s own account, BotStopper answers only the first question.
• Channel dependence. Value depends on how well detection integrates with your contact-center or meeting stack.
• Cross-enterprise trust needs standards. A vendor registry helps, but agents moving between organizations need shared rules.
• Detection can’t replace governance. Knowing a caller is a machine doesn’t set the policy on what it may do.

For a broader look at how enterprises inventory AI systems, assess risks, enforce policies, and oversee autonomous agents, explore our guide to AI governance tools and platforms.

An Implementation Roadmap for Enterprises

The steps below are a general plan, not a Pindrop-specific procedure.

Stage 1: Understand your exposure

1. Map which channels carry sensitive actions: phone, video, chat and API.
2. Identify which actions an agent or impersonator could abuse: balance checks, claims, password resets, hiring and payments.
3. Review current authentication, including knowledge-based questions that synthetic voices can pass.

Stage 2: Add visibility

4. Measure how much automated and synthetic traffic you already receive.
5. Decide which kinds of agents you want to welcome, such as those acting for real customers, and which you want to block.
6. Choose detection and risk signals, and set clear actions for each outcome: allow, challenge, route or block.

Stage 3: Strengthen authorization

7. Define least-privilege scopes for what an agent may do on a customer’s behalf.
8. Require stronger verification for sensitive actions, and record human approvals.
9. Keep an audit trail linking approvals, delegations and actions.

Teams designing these controls can also consult our agentic AI implementation roadmap, which covers guardrails, human approval checkpoints, evaluation, and monitoring.

Stage 4: Review and adapt

10. Review false positives and false negatives regularly.
11. Retune thresholds as attacker behavior and agent tools change.
12. Revisit policy as standards and regulations evolve.

Questions to Ask Any Vendor, Pindrop Included

1. What does your detection measure, and how was it evaluated on audio like ours?
2. How do you report false positives and false negatives, and what’s typical?
3. How often are models updated as new voice and agent tools appear?
4. What happens to a flagged interaction, and who decides?
5. How do you treat legitimate agents acting for real customers?
6. What data do you store, for how long, and where?
7. What biometric-privacy and consent obligations apply, and how do you support them?
8. Which integrations are production-ready for our platforms?
9. What do your warranties or guarantees actually cover, and under what conditions?
10. How do you support audit, explainability and human override?

For a practical example of agent-level controls in a development environment, see our guide to the Claude Agent SDK and its permission controls.

Privacy, Compliance and Ethics

Trust tools handle sensitive data, so they raise their own trust questions.

• Biometric rules vary. Voice authentication can involve biometric data. Pindrop’s platform page says a court-recognized exemption under Illinois’s biometric privacy law permits qualifying authentication using biometric data without written consent, and its newsroom lists an August 2026 item about a federal appeals court affirming that it qualifies as a financial institution under that law. That’s the company’s position, and applicability depends on your use and location, so get legal advice.
• False positives have a human cost. Blocking a real customer or candidate damages trust. Plan for review and appeal.
• Accessibility matters. Voice-based checks can affect people with speech differences, so offer alternatives.
• Be transparent. Inform customers and candidates when analysis tools are being used, wherever disclosure is required.
•  Keep human judgment in the decision-making process. Pindrop’s FAQ indicates that a detection signal needs follow-up verification and human evaluation before conclusions are drawn.

Common Mistakes (Building Trust With Agentic AI From Pindrop)

Mistake	Why it hurts
Treating all bots as hostile	Legitimate agents acting for customers get blocked
Relying on one check at the start	Trust can change during a session
Ignoring the human approval step	A faked approval can make later actions look valid
Taking vendor accuracy at face value	Real traffic differs from benchmark audio
Skipping policy design	Detection without clear actions creates noise
Forgetting privacy and consent	Biometric and monitoring rules can apply

Where Agentic Trust Is Heading

These are informed expectations from visible trends, not predictions.

• Agent registries and identity verification. Pindrop’s AI Voice Consortium and related initiatives suggest a path toward identifying the agent behind a call and determining who it represents.
• Agent identity standards. The NIST concept paper and industry commentary suggest agent identity, authorization and delegation will become more formal.
• Continuous, adaptive trust. Instead of a single login check, trust may be reassessed as behavior and context change.
• More channels. Pindrop says it intends to extend beyond voice to more channels and industries, and the same pressure applies elsewhere.
• Agents defending against agents. Both attackers and defenders are applying AI, so detection and investigation will keep evolving.
• Human oversight stays central. Even Pindrop’s own agent features are framed as working alongside human analysts.

The same oversight question applies to software development, where our comparison of agentic AI coding tools explores how autonomous coding workflows differ in capability and human control.

Frequently Asked Questions (Building Trust With Agentic AI From Pindrop)

Q: What does “building trust with agentic AI from Pindrop” mean?

A: It describes Pindrop’s effort to confirm the identity of people or AI agents involved in live interactions. The approach combines deepfake detection, risk assessment and authentication with AI-powered tools designed to assist investigators.

Q: What is Pindrop?

A: A cybersecurity company that provides identity verification and deepfake detection across voice, video and digital channels. It says it serves many large banks, insurers and healthcare organizations.

Q: What is Pindrop’s entity, intent and action framework?

A: A three-question model described by its Chief Product Officer: is this the right entity, does it have the right intent, and are we enabling the right action? It’s meant to go beyond identity alone.

Q: What is Pindrop BotStopper?

A: A standalone product, announced Sept. 16, 2026, that detects AI voice agents and automated callers in real time and checks them against a registry of more than 5,000 known AI voices.

Q: Does BotStopper block AI agents?

A: Pindrop says its goal isn’t to block AI agents, since legitimate agents acting for real customers should be welcomed. The idea is to know which agent is calling, who it represents and what it should be allowed to do.

Q: What is the AI Voice Consortium?

A: A Pindrop-managed registry of known AI voices, launched with more than 5,000 entries, that feeds BotStopper and the Pindrop Intelligence Network.

Q: How do Pulse, Protect and Passport differ?

A: Pulse detects synthetic speech and deepfakes, Protect assesses fraud risk using multiple signals, and Passport authenticates known callers against prior enrollment.

Q: Can I rely on Pindrop’s accuracy numbers?

A: They’re useful starting points, but most come from Pindrop. Pindrop cites a June 2026 Podonos benchmark, which we couldn’t independently verify. Evaluate performance on your own traffic.

Q: Why does human approval matter for AI agents?

A: Because an agent acts on delegated authority. If the human approval behind it was faked, later actions can look valid while being untrustworthy, which is why Pindrop argues approvals need verification too.

Q: Is there a joint “agentic AI Pindrop Anonybit” product?

A: We found no company documentation describing one. Pindrop, Anonybit and Validsoft appeared together in a webinar and related coverage, but each is a separate company with its own approach.

Q: Are biometric voice checks legal?

A: It depends on the place and the use. Biometric-privacy laws can require notice or consent. Pindrop cites a court-recognized exemption for qualifying authentication, but you should get legal advice for your situation.

Q: Where can I find Pindrop’s own materials?

A: Its platform page, newsroom and resources hub carry the product descriptions, the BotStopper release, the NIST response article and the interview with its Chief Product Officer.

Conclusion: Final Thoughts

The most useful idea in Pindrop’s thinking is also the simplest: trust in an agentic world is a set of questions asked continuously, not a gate passed once. Is this the right entity? Does it have the right intent? Should this action be allowed now? Those questions apply whether the other party is a customer, an employee, a stranger with a cloned voice or an AI agent acting for a real person.

Pindrop’s products map to parts of that problem. Pulse and BotStopper address whether something is synthetic or automated. Protect scores risk. Passport authenticates known people. Its newer agents help analysts work faster. The company is candid that “which agent is this and what should it be allowed to do” is still ahead of it, and the cross-company problem of trust between enterprises is bigger than any single vendor.

Three habits will serve any team well. Treat vendor statistics as leads to verify on your own traffic. Verify the human approval behind an agent, not only the agent. And design what happens after a signal fires, because detection without clear policy is just noise.