A new kind of caller is showing up in contact center queues. It is not a customer and it is not a robocall. It is an AI voice agent working on behalf of a real business, calling to check a claim, verify coverage, or get a case status. In healthcare and insurance operations in particular, provider offices are adopting these agents to replace long manual phone verifications, and the receiving contact center feels it as extra volume that nobody forecast.
The hard part for workforce management is that these calls are legitimate, they often come from the same phone numbers real people use, and they count toward service level. You cannot simply block them. You also usually cannot afford to buy new speech or bot-detection technology. This post covers what the calls do to your numbers, and the practical levers an operation running an Avaya telephony and IVR stack can try with what it already owns.
Why AI callers are different from normal volume
Industry commentary on these agents describes systems that call payers directly, navigate the IVR on their own, wait on hold, and write results back into the sender's own system. Manual phone verification is often cited at roughly 24 minutes and around 14 dollars of staff time per transaction, which is exactly why senders automate it. Portals handle most verifications, and voice agents pick up much of the remainder, so the phone channel is where the leftover, hardest-to-automate volume lands on you.
From a queue perspective, an AI caller behaves differently from a person in four ways:
- It passes the IVR. It answers prompts correctly, including "press 9 if you are a human" style challenges, so a simple DTMF gate does not separate it from a person.
- It waits patiently. A person abandons after a few minutes. An AI caller sits in queue as long as it takes, so it never relieves pressure by hanging up.
- It redials. If a call fails or is incomplete, a second attempt often follows, so one status question can become several calls.
- It shares a number with humans. When the caller ID belongs to a busy office, the same number carries both real staff and the automated agent, so ANI alone cannot decide who is who.
What it does to service level
To size the damage, take a queue that handles 180 human calls per 30-minute interval at 300 seconds of handle time. Staffed with Erlang C at an 80/20 target, it needs 36 agents and delivers about 86 percent service level at roughly 83 percent occupancy. Now add AI calls on top, assuming each one is short, about 120 seconds, and the roster does not change.
Notice that the AI calls are short, so average handle time actually falls, from 300 seconds to about 258 at a 30 percent load. Service level still collapses, because the number of calls rises faster than the average gets lighter, and the fixed roster runs from 83 percent to 93 percent occupancy. If those callers redial, with each call returning 1.5 times on average, a 30 percent load behaves like 45 percent and service level drops to about 16 percent.
This is why "they are short calls, they cannot matter" is a trap. Short calls still occupy an agent, still wait in the same queue, and still count in the denominator.
Do not block them, shape them
Before the levers, a ground rule. These are real calls from real organizations, and some of them support patient care or time-sensitive decisions. A hard block can damage a relationship or a regulatory commitment and often triggers redial storms. The goal is to reduce how much of this volume reaches a live agent at the busiest times, not to refuse it.
Free levers on an existing Avaya IVR
The levers below use call vectoring and IVR flow design that most Avaya Call Center Elite environments already include. Command availability and licensing depend on your release, so confirm each one with your telephony administrator and test in a non-production vector first. Nothing here is official vendor guidance.
1. Measure the volume first. Pull call records by caller ID, hour, IVR option chosen, and handle time for the suspected numbers. Automated callers leave a fingerprint: repeated calls from one number within an hour, near-identical handle times, the same IVR path every time, and heavy activity at a fixed time of day. You need the real share of volume before you spend anything on a fix, and you need the number for the next conversation with the business owner.
2. Finish the job inside the IVR. If the IVR already verifies an identifier and reads back a status, make that answer complete. The calls that cost you are the ones that fall through to an agent for a detail the IVR could have supplied. In the model above, completing half of the AI calls in the IVR lifts service level at a 30 percent load from 51 percent to about 73 percent.
3. Give known numbers their own lane. Call vectoring can branch on caller ID, and vector routing tables are commonly used to hold lists of numbers. Send the suspected numbers to a dedicated vector with its own treatment instead of the main queue. Because humans share the number, the lane still has to give a live person a clear path. A short question that needs a spoken comprehension step, or a randomized digit instruction, can be applied only to this lane, so the extra friction never reaches ordinary callers.
4. Move the volume to quiet hours. An AI caller is patient and will try again. Time-of-day conditions in a vector can play an announcement at peak intervals that tells this lane when status requests are fastest to complete. In the model, moving 60 percent of the AI volume out of the peak interval lifts service level from 51 percent to about 76 percent. The catch is that the destination interval needs real slack, so check the valley before you move anything into it.
5. Cap the lane, with care. A vector condition that counts simultaneous calls in a VDN can limit how many calls from the flagged lane are active at once. Excess calls hear a message instead of joining the queue. Use this only with a clear announcement, because a bare busy signal or disconnect invites immediate redialing.
6. Protect live humans with priority. Queue-to-skill priority levels let genuine callers move ahead of the flagged lane. This improves the experience for people, but it does not change the blended service level number if the business measures every call equally. It is a protection tool, not a reporting fix.
| Lever | Cost | Risk | Effect in the 30% load model |
|---|---|---|---|
| Measure by caller ID and hour | None | None | Enables everything else |
| Complete the status inside the IVR | Low | Low | Service level 51% to about 73% |
| Dedicated lane for known numbers | Low | Medium | Prerequisite for the levers below |
| Time-of-day demand shaping | Low | Medium | Service level 51% to about 76% |
| Cap simultaneous lane calls | Low | High (redials) | Protects the main queue |
| Priority for live callers | Low | Low | Better human wait, same blended number |
Combined, completing half of the AI calls in the IVR and moving about half of the rest out of the peak brings service level back to about 80 percent, without a single new license or agent.
The business conversation matters as much as the vectors
Two non-technical moves make the technical ones last.
Report two service level numbers. Keep the blended figure the business measures, and add a second figure for human-originated calls only. This is not an attempt to exclude anything. It is the evidence that shows how much of the miss comes from automated volume, and it keeps the discussion about facts rather than blame.
Offer the automated senders a better channel. The durable fix is a self-service status page or file-based status exchange that suits machine callers better than your IVR. Senders prefer portals, and most already use them for the majority of their work. A pointer in your IVR announcement to the web status option costs nothing and is often enough to move volume.
A sensible rollout order
- Measure for two weeks and quantify the share and the hours.
- Make the IVR status answer complete, then re-measure.
- Create the dedicated lane for the flagged numbers.
- Add the peak-hour message and watch the redial pattern.
- Add a cap or priority only if the first four steps are not enough.
- Share the two-number service level report with the business owner.
The bottom line
AI callers turn a forecasting blind spot into a service level problem, and the answer is neither a block nor a purchase. Measure the volume, make the IVR finish more of the job, give the flagged numbers a lane, and move their demand away from your peak. Each step is cheap, each can be tested on its own, and together they recover most of the lost service level.
Model your own AI call share with the free Erlang C Calculator, and test roster options with the Capacity Planning Calculator.