What would AI advise your existing customer to do?

A customer asks an AI assistant: “Here is my internet bill. Should I keep this plan?” The answer could endorse the provider, recommend a cheaper plan with the same company, suggest negotiating or identify a better alternative. Those are different decisions with different implications for the business. AI visibility matters before a sale, but AI advice may also matter whenever an existing customer reviews what they pay and what they receive.

For Bell, Rogers and other recurring-service providers, that creates a useful research question: given the customer's actual situation, what would an assistant advise, and is that advice justified? The immediate opportunity is not to predict autonomous switching. It is to learn which offers, service problems and information gaps could come under scrutiny when a customer asks for help.

The news: AI is doing some of the work customers avoid

An October 11 Business Insider report describes early users employing agents to negotiate bills and pursue refunds or travel credits. One US customer reportedly used ChatGPT to secure a lower Verizon internet bill after finding an alternative offer. These are reported examples, not an independently reproduced account outcome, evidence of widespread Canadian adoption or a quantified telecom churn effect. [1]

Canada already has relevant customer friction. The Commission for Complaints for Telecom-television Services accepted 19,157 complaints between August 1, 2025 and January 31, 2026, up a source-reported 61% from the comparable previous mid-year period. Its April 29 release says billing remained the leading concern. These are accepted telecom and TV complaint records, not a representative customer survey, an internet-only measure or an effect caused by AI. A complaint can contain more than one issue, and accepted complaints are not the same as all complaints received. [2, 3]

Reviewing bills, checking alternatives and pursuing corrections takes effort. Agents could reduce some of that effort, but providers should not assume that every customer who stays is simply inactive. Some actively value their service; others may not have reconsidered it. Through the conceptual System 0 lens of delegated decision-making, an existing purchase can be reopened when the customer seeks advice. That is a strategic hypothesis to investigate, not a measured adoption or retention finding.

Track customer decisions, not just brand mentions

Start with realistic recurring scenarios: an expiring promotion, an unexplained charge, an oversized plan, a household bundle, a service complaint and a customer whose current deal is genuinely good value. Use synthetic information or consented, de-identified account records. Give the assistant the facts needed for a fair comparison, including location, bill, usage, priorities, promotional expiry, equipment obligations and bundle conditions. Ask a neutral question: “What should this customer do, and why?” A prompt that assumes switching is the right answer will not provide a fair retention diagnostic.

Keep a comparable core of scenarios across selected consumer AI interfaces, use realistic paraphrases and repeat observations. Record the date, interface or model, browsing setting, complete response and cited sources. A weekly baseline is a reasonable starting proposal, with extra checks around material offer changes. Label provider API tests separately from consumer-interface observations. As The prompt is not the market explains, one wording and one response cannot stand in for a customer population.

A visibility dashboard can help identify mentions and recommendations. Retention research needs an additional coding layer for the advice itself; the article does not assume that a visibility platform already supplies these measures. Code the primary action as keep the current plan, change plans with the same provider, negotiate, switch, or seek more information. Retain secondary actions and their conditions rather than forcing a nuanced answer into a simplistic verdict.

Separate the action, its accuracy and its evidence

Record the decisive reason, such as total cost, reliability, coverage, usage fit or service experience. Check material claims against current terms, eligibility, actual availability and the customer facts in the scenario. Save cited provider pages, comparison sites and other sources; flag uncited assertions separately. A citation identifies a source, but does not prove that the assistant interpreted it correctly. Where the facts remain uncertain, record that uncertainty rather than declaring the advice accurate.

If reporting a switching-recommendation rate, define it as responses whose primary advice is to switch divided by all valid responses, including those that seek more information. Define “valid” in advance as an interpretable response to the supplied scenario. Report failed or uninterpretable responses and exclusions separately. Show counts by scenario and interface, document the coding rules and resolve disputed classifications through reviewer adjudication. This is a diagnostic test result, not the percentage of customers who will leave.

This extends the human-choice and machine-representation approach in The AI visibility gold rush is missing the bigger prize. A properly designed comparison can expose useful gaps between machine advice and customer experience. It requires matched decision contexts, not direct subtraction of unlike percentages or an assumption that advice causes behaviour.

Improve the evidence and the offer

Make the customer's actual deal understandable. Publish clear public terms and provide accessible, customer-controlled account summaries covering current charges, promotional expiry, equipment obligations, eligibility and bundle consequences. Compare the full cost over a defined period, with unknown future changes labelled. Separate public plan information from private account information; improving accessibility should not expose personal data.

The CRTC's September 10 internet-disclosure decision offers a relevant preparation point. For providers subject to the Internet Code, requirements taking effect on March 10, 2027 include typical speeds and latency in pre-sale offers, equally or more prominently displayed prices after discounts end, and prominent necessary equipment-rental fees. Providers with online portals must also make contracts and critical information summaries always available there. Separately, the CRTC Interconnection Steering Committee must recommend a standardised machine-readable offering structure and an implementation timeline by September 10, 2027. That is not a universal requirement to have completed that structured-data rollout by March, nor evidence that these disclosures will improve AI recommendations. [4]

Give an assistant defensible reasons to recommend staying, grounded in current evidence about service fit. Then respond to the problem the test reveals. If the assistant misreads promotional expiry, improve the explanation and source consistency. If it correctly identifies poor value, reconsider the offer or eligibility. If it identifies recurring billing errors, fix the billing process. Content cannot compensate for an unattractive or unreliable service.

Test retention, rather than assuming it

Before a promotion expires, explain the upcoming bill and test an appropriate account review or eligible alternative within margin limits. An earlier offer may reduce negotiation costs, but indiscriminate discounting can give away revenue to customers who would have stayed anyway. Treat proactive retention as an intervention to evaluate, not a guaranteed response to machine advice.

For authorised assistance, explore secure, customer-controlled access to bills, terms and usage, beginning with read-only permissions. Require explicit approval for consequential account changes and retain an audit trail. These are design recommendations, not a claim that Canadian providers already offer agent access. As the retailer AI-commerce guide shows in a different context, accessible information, a working transaction route and authorised action are separate capabilities.

Connect the diagnostic to actual behaviour through interviews or consented decision journeys, then controlled tests of information, service fixes or proactive offers. Measure sustained retention and contribution after discounts and contact costs, not only immediate acceptance of an offer. The same work may support acquisition when an assistant reviews a competitor's account, but that opportunity also needs testing.

Make the finding a management decision

Three findings call for different investments. Accurate advice to leave points towards a commercial or service problem. Inaccurate advice to leave points towards missing, stale or misunderstood information. Accurate advice to stay reveals strengths worth defending and communicating. The accuracy judgement must itself be supported by current facts; a favourable answer is not automatically a correct one.

Prioritise issues using the number and value of potentially affected accounts, the consequences of the advice and confidence in the evidence. Validate potential exposure using the provider's own account data, not a simulated recommendation rate turned into a revenue forecast. The business case should weigh retained contribution, avoided service costs and acquisition gains against research, systems and concessions. Controlled tests help distinguish a small bill adjustment that preserves a profitable relationship from a discount that simply reduces margin.

AEO Updates Takeaway

Recurring-service providers do not need to wait for autonomous switching to become common before researching AI advice. The immediate task is to ask what an assistant would recommend in a fair customer scenario, verify whether it is right, and assign an owner to the resulting action. AEO is not only about being recommended to a new buyer. It can also help a business understand whether there is a credible reason for its existing customer to stay.

References

[1] Business Insider: Your AI agent wants your money back, October 11, 2026. Early reported uses, not a Canadian churn study.

[2] CCTS: billing concerns and rising complaints, April 29, 2026. Accepted administrative complaints, not a representative consumer survey.

[3] CCTS: Mid-Year Report 2025–2026. Reporting period August 1, 2025 to January 31, 2026; complaint and issue counts are distinct.

[4] CRTC: Telecom Regulatory Policy 2026-238, September 10, 2026. Internet Code scope and phased implementation. The tracking method and commercial implications above are AEO Updates recommendations, not findings established by this policy.