AI Receptionist Spam Calls: What 493 Real Calls Showed Us, and the 4-Layer Fix We Now Build (2026)

Short answer: Yes, robocalls will eat into an AI receptionist's minutes, and the hit is bigger than most owners expect. In one multi-business Retell workspace we audited in October, 158 of 493 inbound calls in 30 days were robocalls. Another 188 had no caller speech at all. Robocalls the agent politely talked to ran about a minute each. Robocalls dropped by detection ran about four seconds. The fix is not one setting. It's four layers: reject known junk before the agent answers, detect recordings in the first seconds, give the agent a fast exit script, and tighten the default timers. Then keep spam out of your CRM so your team never chases it.

We build AI voice receptionists at Wrk Less, mostly on Retell with n8n behind them, from Orlando. Here's what the call logs actually showed and the spam setup we now ship on every build.

How much spam does an AI receptionist really get?

More than your old voicemail made you notice. A voicemail box hides robocalls. An AI receptionist answers every one, transcribes it, and bills for it.

Here's the 30-day breakdown from that audit (Sept 12 to Oct 12, 2026). The workspace held three businesses: a home-service contractor, a professional-services firm and the reseller's own front desk.

Call type	Calls	Minutes	Spend	Avg. cost per call
Real callers	147	141.2	$40.25	about 27 cents
No caller speech (silent or instant hang-up)	188	41.3	$14.69	about 8 cents
Robocalls the agent talked to	45	45.5	$14.40	about 32 cents
Robocalls dropped by IVR detection	113	7.6	$2.67	about 2 cents
Total	493	235.6	$72.01

Fewer than one in three calls was a real person. Real callers drove $40.25 of the $72.01 spent.

Three findings changed how we build:

1. The expensive spam was the polite spam. The 45 robocalls the agents engaged averaged about 61 seconds. Nearly half ran 55 to 60 seconds, because the agent waited out the whole recording and then offered to take a message. One agent told a loan robocall, "I've recorded your message about the preapproval." That message then went to the owner like a lead.
2. Blocking numbers barely helps. Those 45 robocalls came from 44 different numbers. A blocklist only catches repeat offenders.
3. Detection is cheap when it fires. On the one line where Retell's IVR detection was catching robocalls, 113 of them ended in about four seconds each, for about 2 cents a call.

The scripts were what you'd guess: "approvals desk" loan offers, business-funding pitches, and "your Google Business listing has a verification issue." That lines up with the national data. YouMail's Robocall Index estimates 3.8 billion robocalls in September 2026, and its October 7 release says September's biggest campaigns were pre-approved debt-consolidation loan offers sent from tens of thousands of numbers. Orlando ranked 13th nationally with an estimated 43,033,300 robocalls, up 4% from August (YouMail Robocall Index).

Is the FCC fixing this for you?

Slowly, and not in a way you'll notice on your bill. The FCC's Enforcement Bureau keeps cutting off bad carriers. On October 5 it ordered downstream providers to block Digital Solutions Inc. and pulled it from the Robocall Mitigation Database. On October 16 it removed Mexico IP Phone, giving providers two business days to stop accepting its traffic. That pressure is real, but campaigns that rotate through thousands of numbers keep getting through. Your receptionist needs its own filter.

How do you stop an AI receptionist from answering robocalls?

Stack four layers. No single layer catches everything. Together they push most junk under five seconds.

1. Reject junk before the agent picks up

Retell lets you point each phone number at an inbound webhook. It fires before the call connects. If your endpoint returns reject: true, Retell hangs up "before any agent or voice setup," and no call object is created. We run that webhook in n8n. It looks up the caller in the CRM, rejects numbers your team has already tagged as spam, and passes known customers to the agent with their name and open job. That last part is the same pre-call lookup we use for repeat-caller context.

Two cautions. Retell only rejects on the boolean true. A string "true" is ignored, and the call goes through. And if your webhook fails all three attempts, Retell falls back to the number's default agent. Build it so a failure still answers real customers.

2. Turn on IVR detection with a custom prompt

Retell agents have voicemail and IVR detection settings. The agent API lets you add a detection_prompt that describes what counts as an IVR, with a hang-up action, and checks during the first 3 minutes of the call. Our front-desk agents carry a prompt that names the current scripts: Google listing verification, "press zero to speak with an agent," warranty and loan recordings.

Be straight about one thing. Retell's October 16 guide to answering machine detection says these detectors run on outbound calls and have "no effect on inbound calls." In the logs we reviewed, inbound robocalls on the line with detection configured still ended with Retell's ivr_reached reason in about four seconds. Treat it as a bonus layer, test it with your own recorded robocall, and don't make it your only defense.

3. Give the agent a one-line exit

Most agents are written to be helpful, so they take a message from anyone. Add a rule near the top of the prompt: if the caller is a recording, or is pitching loans, funding, listings or SEO, say one sentence ("This line is for customers of [business]. Goodbye.") and call end_call. Don't offer to take a message. Don't write a CRM note. That rule goes straight at the 45 engaged robocalls, which averaged about a minute each and produced fake "messages" for the owner.

Keep the opposite case in mind too. Some legitimate callers are machines now. Google's assistant can call a business for a customer, and it says who it's calling for. Your exit rule should target sales pitches and recordings, not every robotic voice.

4. Change the default timers

Retell's API defaults are generous: end_call_after_silence_ms is 10 minutes and max_call_duration_ms is 1 hour. A front desk needs neither. We set silence to 10 to 30 seconds and cap calls at 5 minutes, then hand anything longer to a person through a warm transfer.

The spam-proofing checklist we run on every build

• Inbound webhook on every Retell number, with a CRM lookup and a spam-tag reject list
• Webhook failure path tested (calls still reach the default agent)
• IVR and voicemail detection on, with a detection prompt naming current local scripts
• Exit rule in the prompt plus the end_call tool, tested against a recorded robocall
• Silence timeout at 10 to 30 seconds, max call length set on purpose
• Post-call webhook in n8n tags each call real, silent, spam or IVR from the disconnection reason and transcript
• Spam and silent calls never create a lead, a task or an owner alert
• Monthly report: spam share of minutes, new spam numbers, real calls missed

The post-call piece runs through our Retell webhook integrations. It's the difference between a clean morning callback list and a list padded with loan offers. We use the same tag-and-drop rule in our real-estate after-hours build.

What about the silent calls?

Don't treat them as spam. In the audit, 188 calls had no caller speech, and 89 of them ended within 10 seconds. Some are dialers checking that a line is live. Some are real people who hung up when they heard an AI. You can't block your way out of that.

Two fixes help. First, a shorter greeting that names the business in the first second, so a real caller knows right away they reached the right place. Second, a follow-up for unknown silent callers who might be customers, sent only where your consent setup allows it. Read our missed-call text-back guide before you automate texts.

Does spam count against your AI receptionist bill?

It depends on the vendor. Smith.ai says every answered call counts by default, with an allowance to remove "up to 10% of their calls" as spam each billing cycle (Smith.ai AI receptionist pricing). In our audit, robocalls were 32% of calls. Jobber says calls it tags as spam don't count as Receptionist conversations (Jobber help). Usage-based platforms like Retell bill every connected minute, which is why the four layers above matter. We compare the billing models side by side in AI receptionist pricing 2026.

Bottom line

An AI receptionist answers everything, so it has to know what to hang up on. Reject what you can before the call connects. Detect recordings in seconds. Give the agent permission to end the call. Keep spam out of the CRM. If your call logs are full of loan offers and your owner is getting "messages" from robocalls, book an AI systems assessment. We'll pull your last 30 days of calls, show you the spam share in minutes and dollars, and fix it. See how we handle Retell AI voice agent setup.