Gartner® names Reality Defender a Market Shaper in deepfake detection, June 2026

Get the report

\

Insight

\

Agentic AI bots aren’t the enemy, and blocking them hangs up on real customers

Tim Male

Product Marketing Manager

Published November 3, 2026

Agentic AI bots are flooding your IVR. Containing them, not deflecting them, is what protects your queue and the customer experience

Agentic AI bots are flooding contact center IVRs, and the goal is to contain them, not deflect them. Most are legitimate. Customers and partners send them to do routine work that belongs in an automated lane, not the human queue. A few are hostile. Detection tells you which calls carry AI-generated speech so your workflow can route, hold, or pass them. The false-positive rate determines which actions you can safely automate.

Agentic AI bots now check eligibility, chase claim status, and ask for balances at volumes no staffing plan anticipated. Every one that reaches a live agent impacts handling time, cost-to-serve, and employee sentiment.

The reflex is to deflect them. That’s the wrong goal. Deflecting bots acting on legitimate customer requests jeopardizes your customers’ experience. Containment keeps them in an automated lane where it gets resolved, and successful containment tells you your IVR is working.

As a detection company, it’s awkward to admit that most AI callers are legitimate. A customer’s assistant rebooking a flight is still a customer transacting. A provider’s bot running eligibility checks against a payer line saves billing staff from waiting on hold.

Still, a few are hostile. Applying one policy to every AI bot ignores nuance and gets it wrong in all directions. You turn away traffic you invited, and the adversary adapts by next week.

Agentic AI bots are worth routing, not refusing

  • They cost less to serve. A call handled in an automated lane doesn’t take time away from an agent. Every bot routed off the human queue is capacity returned to live customers who need to connect with a person.
  • They wait without complaining. Agentic AI bots don’t abandon, don’t get irritated, don’t need hold music. Queue them, batch them, push them to off-peak.
  • They ask for a narrow set of things. Eligibility, claim status, balance, and order status. Low variance, high volume, the calls you want off the human line.
  • They carry a real customer behind them. Behind the bot sits a customer, a patient, or an account holder. Refuse the bot and your customer does the same work again, only slower and more frustrated.
  • They show you what to build. Bots converging on the same three questions are describing a workflow or an API you haven't built.

Each of those is an AI call routing decision. To succeed, routing decisions need input. You can't send a bot to the automated lane if you don't know it's a bot, and you can't hold the hostile one if it looks like the rest.

Run the math before you block anything

Across our deployments, agentic AI bots account for 3–5% of inbound volume. Most contact centers can't yet say what their own split looks like, which is why this next part matters.

Screening for something rare is a studied problem with a known trap. Stefan Axelsson set it out in The Base-Rate Fallacy and the Difficulty of Intrusion Detection (ACM Transactions on Information and System Security, 2000). When the target is rare, even an accurate classifier will trigger more false alarms than real catches.

Work it through with round numbers. Say your contact center takes a million calls a month, and 2% of them are actually agentic AI bots. A classifier catches 95% of those bots, so it flags 19,000 of them. It also wrongly flags 3% of human callers, accounting for 29,400 real people.

That’s 48,400 calls flagged as AI bots, but only 40% are actually bots.

6 out of 10 calls you act on are customers

Tighten the false-positive rate to 1% and precision reaches 66%, or 9,800 people. Reach 0.1%, and it's 95%. You’d need to reach that level of precision before automated refusal is defensible, and it's 30 times tighter than the 3% this example started with. The tighter the false-positive rate, the more of your policy you can safely automate. Until you get there, route rather than refuse.

No detector arrives at 0.1% out of the box. Your traffic, your goals, and the threshold you set during deployment decide whether you reach it. A benchmark figure is where that tuning starts.

The arithmetic applies to every detection vendor, Reality Defender included, so take it as a design rule rather than an argument against detection. How often a detector is right decides what you safely do with that information, not whether it’s useful at all. A signal that's right four times in 10 can still send a call down a cheaper path, because being wrong only costs a real person a few seconds of routing. Hanging up on someone needs far better odds.

If a vendor tells you to block AI bots, ask them to run that calculation on your call mix in front of you and tell you how many customers you’d hang up on.

Ask what a list of known voices can't cover

One acceptable approach matches the incoming voice against a catalog of known synthetic voices. A catalog of known bots and generators is useful as one input. When a caller matches something already seen, you get a fast, confident resolution.

The failure is in what it can't do. It can't carry the primary decision, and it can't tell you which system you're hearing once the generator sits outside the reference set.

Identifying which system generated a piece of audio is called source tracing. At Interspeech 2025, systems reached roughly 95% accuracy when the generating system sat inside the reference set and about 45% when it didn't (Unveiling Audio Deepfake Origins, arXiv:2506.02085). A companion open-world study across 699,000 utterances and 13 synthesizers reported error rates of 35% to 50%, near chance (STOPA, arXiv:2505.19644).

Detection shows the same shape. ASVspoof has separated known from unknown attacks since 2015. The challenge's own authors warned that countermeasures tuned to known attacks tend to perform poorly against unknown ones (Wu and colleagues, IEEE Journal of Selected Topics in Signal Processing, 2017).

Channel conditions compound it. Most published detection figures come from studio audio, while contact centers run on narrowband telephony that discards the frequencies carrying much of the evidence. One group built both paths and measured the gap. An equal error rate under 5% in the lab became a 23.7% miss rate over real handsets on real networks (On Deepfake Voice Detection, It's All in the Presentation, arXiv:2509.26471).

A catalog records what someone already collected. Your next AI caller could come from a generator you’ve never seen before

Route the traffic, don't refuse it

This argument ran on the web a decade ago and finished in June. Forrester renamed the category from Bot Management to Bot and Agent Trust Management in The Forrester Wave, Bot and Agent Trust Management Software, Q2 2026, published June 15, 2026. The report finds the market has moved from blocking bots to letting trusted automated traffic through. Websites faced automated traffic years before contact centers, and they learned a wall shuts everyone out. Nobody there sells a wall anymore.

The web is also building the plumbing. Web Bot Auth is a chartered IETF working group drafting a way for bots to cryptographically sign their requests. The receiving site then knows who's calling and on whose behalf.

The voice channel has none of it. No attestation standard, no signed caller, and nothing in the IETF, W3C, or NIST AI agent identity work that addresses telephony. Creation ran ahead of the trust layer here, exactly as it did everywhere else.

Some bots don't speak in the IVR at all. They navigate by keypad, which gives detection nothing to analyze, so the verdict arrives after the transfer rather than before it. That changes what the agent can do, not what the queue can do. If a vendor tells you otherwise, ask how well it worked in a real IVR.

Where speech exists, it's the one thing a bot can't withhold.

RealCall analyzes inbound voice streams in real time inside contact center infrastructure. It returns a signal on whether the speech shows signs of AI generation, with no enrollment and no stored voiceprints. It gives you the signal and your team writes the policy, which, given the arithmetic, is the only defensible division of labor.

For a closer look at integration points and deployment tradeoffs, see How to Detect AI Agent Callers: API Options Compared.

Sorting isn't deprioritizing. A provider's bot chasing a prior authorization is doing urgent work for a waiting member, and a lane built for bots should be faster than the human queue. If your automated path is where requests go to wait, you've built a wall with extra steps.

Detection is a continuous discipline. Nobody keeps the machines out. The enterprises that do well will know what's calling and have somewhere useful to put it.

Schedule a RealCall demo to see real-time voice detection in your contact center.

Frequently Asked Questions

Questions worth asking any vendor, including us

The biometric standards community settled how to report this years ago, and the discipline carries over even when the standard itself doesn't. ISO/IEC 30107-3 requires the worst case per attack type rather than the mean. A detector can post an excellent pooled figure while sitting near chance against one attack family, because the average absorbs it. Attackers don't use the average. They find the family you're weak against and use it repeatedly because it works. Ask for the full table and read the worst row.

Equal error rate is the operating point where false accepts and false rejects are equal, which is almost never where you'd run a contact center. ISO/IEC 19795-1:2021 deprecates single-number metrics like it outright, and the presentation-attack standard reports false positives at a fixed detection rate instead. That second number tells you how many real customers get flagged when the system is tuned to catch what you want caught. Ask for it at the setting you'd deploy, not the setting that demos well. Then ask who tunes that setting to your traffic once you're live.

Narrowband telephony throws away the high frequencies where much of the synthesis evidence lives, so a clean-audio result doesn't transfer to your queue. Ask which codecs were in the test, at what sampling rate, and whether the audio passed through real handsets. Ask whether it included packet loss, because your network has some. A vendor who can't answer is quoting a studio number at a telephony problem.

Precision isn't a property of the detector. It's a property of the detector meeting your traffic. A system that looks excellent when 10% of calls come from bots looks reckless at 0.5%, because the human population it's wrong about is so much larger. Nobody can quote you a precision figure without first asking what your mix is. If a vendor gives you the number before they ask the question, they're quoting someone else's contact center.

Your policy decides. Route it to automation, hold it before a refund or an account change, flag it for the agent, or let it through. RealCall returns a signal on whether inbound speech shows signs of AI generation, and your workflow acts on it. That boundary is deliberate, not unfinished. The right policy depends on your risk tolerance, your queue economics, and your regulatory exposure, and a detection vendor can't see any of that. Anyone who wants to own that decision for you is guessing.