Should you put an AI voice agent on cross-sell calls to existing customers? It depends on three things: whether your usage data shows a real fit, whether the offer helps the customer, and whether you can get consent and measure results.

Start with product-usage data, not a list

Build your call list from what customers have done in your product. Do not buy it or pull it from a general contact export. Limit AI calls to people who have already signed up. A usage signal gives you a reason to call, and that reason becomes your opening line.

Signals worth calling on:

  • Approved or eligible, but never used the product.
  • A feature adopted halfway, such as a funded account with no recurring deposit.
  • A recent milestone, such as a first repayment or a first year.
  • A lapsed account that was active until recently.

Exclusions to apply before any dial:

  • Open complaints or unresolved tickets.
  • Missed payments or active collections.
  • Any contact in the last few days, from any channel.
  • Customers who opted out, or who never gave consent.

In practice, a digital lender found thousands of customers who were approved for a loan limit but never took the loan.

That is a good call list. The reason for the gap is a guess about confusion, so a call that explains the product and asks a question can test it. A bulk offer cannot.

Keep the segment small enough to read. Tag each customer with the signal that put them on the list, and pass that tag to the agent as context. Then every call starts with the customer's own situation, not a pitch.

Frame the call around the customer's gain

Open with why you are calling, cite the customer's own usage, and ask a question before you make any offer. Calls feel pushy for three reasons. The offer comes first, the script is fixed, and the customer has no easy exit.

Cross-sell and upsell are different asks. A cross-sell offers a second product, such as a savings plan to a loan customer. An upsell offers more of what they already have, such as a higher limit. Either way, the opening line should sound like this:

"Hello Ada, I'm an AI assistant calling from Brightloan. Your loan limit has been ready since March and you haven't used it. Can I ask what you thought it was for?"

That line says who is calling, names the customer's own data, and asks. The offer comes only after her answer, and it matches what she said.

Make the voice sound like a person would talk.

  • Use plain words and short turns of one or two sentences.
  • Let the customer interrupt, and answer what they asked.
  • Pick a local voice, accent or language where you can. In our judgement, local voices tend to land better, and buyers often ask for local accents and languages.
  • Plan for off-script moments. Buyers fear a scripted, robotic agent.

Build the exit before the pitch. Give a clear "no thanks" path that ends the call politely and marks the contact as declined. Design the human handover too. Buyers raise on their own how polite it is, how a disputed amount is handled, and who chases a handover nobody picked up.

Gartner found that 87% of 3,566 customers say a company using GenAI must offer a way to reach a human. Plan for that option on every call, not only the angry ones.

Get consent and disclose the AI first

Check consent before the dialler runs, not after the first complaint. Store it per contact, and tie each record to a channel (voice, SMS, email) and a purpose (marketing, service). A yes to service texts is not a yes to AI-voice sales calls.

Being an existing customer does not settle this. In the US, AI voice calls need consent under the TCPA, and that follows the FCC's 2024 ruling. Treat "they already bought from us" as a reason to call, never as permission.

Disclosure rules differ by region. A short list:

  • US: The FCC's proposed "say it is AI at the start of the call" rule was still not final as of mid-2026. Some states already act. Colorado's Chatbot Safety Act takes effect on 1 January 2027 and requires public-facing conversational AI to disclose that it is AI. Disclose anyway.
  • EU: AI Act Article 50 has applied since 2 August 2026. Any AI that talks to people, including automated phone systems, must say it is AI from the first interaction.

Disclosure may draw a hostile reply on a sensitive call. Plan for that with a calm, short script and an easy path to a human. Disclose in the first sentence, as in the sample opening above.

Set guardrails on every campaign:

  • Suppress anyone on a do-not-contact list.
  • Call only inside local calling hours.
  • Cap retries, and stop after a refusal.

Insight7's voice campaigns are in beta. The spec lists consent marking, calling hours, retry rules and a do-not-call list. A duplicated campaign leaves out anyone added to the do-not-contact list since the original. You can test the agent in the browser, with no phone number needed. Separately, Insight7's call scoring evaluates 100% of customer conversations, not a sample.

Track conversions with a clean control group

Decide what counts as a win before the first call. Pick one outcome, such as offer accepted, product taken or account funded. Then set a window, for example 14 days after the call. Without a window, late conversions get credited to the wrong campaign.

Build the comparison into the list. Two designs work:

  • Voice-only sublist. Carve out a group that only the AI agent will contact, so any conversion from it belongs to voice.
  • Randomised holdout. Leave a random slice of eligible customers uncontacted and compare them with the called group.

In our judgement, avoid referral codes on live calls, because customers rarely can note a code mid-call.

To compare channels, run voice beside SMS on the same customer set. If you already send SMS offers, add a voice arm on the same customers and compare conversion rates. That gives your sponsor a like-for-like result.

Metrics to report for every arm:

  • Reach rate. A call where only the AI spoke is not counted as reached.
  • Opt-outs, escalations to a human and complaints.
  • Conversion per contact, and cost per conversion.

Then read the calls. Transcripts show where customers got confused and where they said yes or no.

This matters because many AI service projects cannot show a return. Gartner's analysis of 432 use cases found about a quarter with positive ROI and a quarter with negative ROI. For 42%, leaders could not tell what value was produced. A control group keeps your campaign out of that last bucket.

Pilot small, review, then scale

Start with a small sample and agree the success measure before the first dial. Regulated buyers often want a short, low-risk pilot, with payment only after results show a measured lift. Write down the outcome, the window and the control group from the last section, and have your sponsor sign off on them.

Before any real number is called, talk to the agent in the browser. Try the awkward replies: "not interested", "who is this?", "put me through to a person".

Then review the first 50 calls:

  • Read the transcripts for the places where customers got confused or hung up.
  • Check that each customer matched the signal that put them on the list.
  • Tune the opening line, the script and the exclusion rules.

Widen the list only after a clean review. Each round teaches you something, so keep the rounds small.

Related reading: Subscription Renewal Call Scripts for AI Voice Agents With Objection Handling, AI Debt Collection Call Script Templates That Get Customers to Pay Without Feeling Chased and 9 Best AI Debt Collection Software Platforms for Lenders and Subscription Businesses.

Questions people also ask

What is the difference between cross-selling and upselling?

Cross-selling offers a different product, such as savings for a loan customer. Upselling offers more of the same product, such as a higher limit. Track them as separate campaigns, because they need different scripts and success measures.

Do I need consent to make AI cross-sell calls to existing customers?

Usually yes. In the US, AI-voice calls need prior consent under the TCPA. Collect a separate, clear opt-in at sign-up, store proof per contact, and have counsel check each region you call.

Should an AI voice agent say it is an AI?

Yes, in the first sentence. Several regions already require it, or soon will. If a customer later asks "am I talking to a machine?", the agent must answer truthfully.

How do you cross-sell to existing customers without sounding pushy?

Make one offer per call and ask one question before it. Accept the first "no" without a second pitch. If the customer sounds unsure, offer a text with details, then end the call politely.

Which customers should you call first for a cross-sell?

Start with one signal that has a clear next step, such as approved but never used. Keep the group small enough to review by hand. A few hundred well-matched customers teach you more than thousands.

How do you measure whether AI cross-sell calls work?

Compare called customers with a holdout group, using the outcome and window you set in advance. Add intermediate wins, such as an app download or a finished identity check. They show whether drop-off comes from the customer or the product.

Is it better to cross-sell by phone, SMS or email?

It depends on the offer. In our judgement, voice suits offers that need explaining, and SMS suits simple offers and links. Some customers prefer messages. Run both on the same customers and let conversion per contact decide.

Call only when usage gives you a reason

Call a customer only when their own usage gives you a reason, and be able to prove the result. A reason you can name and a result you can measure keep a cross-sell call from sounding like telemarketing. If this campaign vanished tomorrow, which number would show you missed it?

Book a demo of Insight7