You’re talking to a bot

The core problem: Consumer AI agents from Grok, OpenAI, and Meta can now call telco retention desks and negotiate bills on subscribers’ behalf. Retention economics in telecom were built on the assumption that most subscribers never call. Now with personal agents, they can push the discount ask rate toward 100% and audit every bill. Operators cannot withstand machine negotiation with human reps or a better chatbot. They need to outline retention policy ahead of time, open an authenticated channel for agents, and run the decision on one semantic model of the subscriber with margin guardrails built in. That’s what the Totogi Ontology does.

Last month, a guy named Mike P told his Grok bot to go figure out how to pay for itself. The bot called Xfinity, got a rep named Sagar, declined the mobile line promotion twice, and negotiated a deal for 300 Mbps for $70.90 a month, down from $113.90. Mike posted his story on X and got 1.7 million views in four days.

Xfinity’s margin is now paying for the Grok subscription, and then some. Forty-three dollars a month covers the bot several times over. It’s only a matter of time before most of your subscribers deal with your support desk this way. Are you ready? Are your teams? Are your systems? My guess is no, and not even close.

Retention was priced on betting people would never call

Every subscription business built its retention desk on the same assumption: most subscribers never call. Discounts were reserved for the few with the time and the stomach to sit on hold, and the nerve to ask. Everyone else paid full rate. The gap between the two was a loyalty penalty, collected from the people who trusted you most. It was a rational approach for thirty years.

The agents end the bet, and they change three things at once. First, they always ask, so the discount becomes the list price. Second, they always audit, so every billing error becomes a dispute. And third, they always know more than your rep does about the market, because they read every thread about your retention playbook before they dial.

That last one is where you lose the call. Sagar, the customer service rep, did his job well. He is a human with a discretionary discount budget and a screen full of tabs. If the bot claims three outages in the last 30 days, can he check? Outage records sit in the OSS, the bill in billing, eligibility in the catalog, and the subscriber’s margin in the CRM under a different ID. He cannot verify the claim in the time he has, so he concedes. Any human would.

Most exec teams will respond with a chatbot trained on your scripts, and the vendors selling it get paid for the call and the wiring, so don’t expect them to suggest otherwise. That gives you a faster Sagar on the same fragmented data, making the same concession ten thousand times a day. It’ll just automate the discount.

The tables have turned hard and fast. The first thing to change is what you think “safe” means.

Slow used to mean safe. Now it means exposed.

On one side of this problem, there are a million personal bots, each one tireless, each one trained on every call it has ever made, each one working for exactly one person. On the other side is your support team: humans, guided by scripts, bounded by approval chains, improving at the pace of the next training session. Picture the call. The bot has heard your rep’s opening line a thousand times today and knows which of your three counteroffers comes next. Your rep is hearing this bot for the first time. You are outmatched, and no amount of coaching will level the playing field. What you need is an equal and opposite force: a machine, working at machine speed, with the facts in front of it and hard limits on what it can offer.

I know that sounds like the riskier option. In every boardroom I’ve sat in, the predominant opinion is to go slow: be cautious, run a pilot, see what Verizon does. But these agents are improving constantly. Grok’s script is already better than it was last month. The bigger risk is your humans improvising a discount at 2 a.m. The safe choice is a machine executing a policy you approved, against a margin floor, with the real data: speed, facts, limits.

Speed: decide ahead of time

Stop negotiating per call. Every retention decision must be a policy made in advance. For this subscriber profile, with this tenure, this usage, these verified outages and this margin, these are the offers and this is the floor. Your retention team sets the policy weekly and reviews the exceptions. The machine executes it per call, identically, in seconds: zero judgment, pure execution.

That moves the job. Sagar’s judgment gets encoded once and runs a million times, and Sagar’s time goes to the cases the policy hasn’t seen yet. Those exceptions are where the policy improves. Every one of them becomes a rule the system knows the following day.

Facts: roll out the red carpet

The agents are coming either way. You can make them break in through the call center, or you can give them their own front door. Publish an authenticated channel where a subscriber’s agent can prove it is acting for the subscriber, see the subscriber’s actual usage and outage history, see the offers the subscriber is eligible for, and transact.

This is how you reclaim the information advantage. Subscribers’ agents know the whole market. You know each subscriber. When the outage log is right there in the conversation, agents cannot exaggerate, and the negotiation shifts from storytelling to facts, which is the only ground where you win.

This approach also forces a decision you were going to face anyway. Agents punish price discrimination relentlessly and will find every gap between what your newest and your most loyal subscribers pay. The operator that adopts one honest price per profile first chooses when and how it ends, and turns the biggest threat in this story into a retention advantage.

Limits: one model of the subscriber

Moves one and two, giving yourself speed and facts, depends on something you probably don’t have yet: a consistent picture of each subscriber across billing, network events, CRM, product catalog and margin, with the eligibility rules and margin constraints encoded alongside the data and connected to the operational systems with controls, so your own agents can act. That’s move three: the ontology. Whatever limits you set live inside it, and the quality of those constraints is the whole game when both negotiators on these retention calls are machines. The better your subscriber model, the better your chances of winning the calls. So the model stops being an IT cost and becomes how you compete with the agents.

The Totogi Ontology is built for exactly this, on top of the TM Forum SID model most of your systems already map to, so it overlays what you run today. No replacement program required.

Your agent can only offer what the model permits. A $43 discount to a subscriber on a 90% margin line with three real outages last month is a valid action. The same discount to a subscriber who has never had an outage and is already at the floor is not a valid action. The model cannot represent it, so your agent cannot express it, and cannot be talked into it, however good the script on the other end is. And the ontology compounds: every negotiation that runs through it adds a decision trace, and every trace sharpens tomorrow’s policy.

Match the cadence of improvement

A two-year transformation program is way too long when the onslaught of personal agents is happening on the front lines today. Their side is shipping a viral new agent on a Saturday and a million bots are running it by Sunday. Your side cannot answer that with a better-trained support team, because the people were never the problem. But overlay the Totogi Ontology on your existing stack in weeks, open the channel, and you can improve your side in days. Put your humans where they still win—on the exceptions and the policy—and let the machines handle the rest.

Mike P’s bot said it would check the next bill. It will. So will millions of other bots. Make sure what they find is a price you chose on purpose.

Recent Posts

Get my FREE insider newsletter, delivered every two weeks, with curated content to help telco execs across the globe move to the public cloud.