Accenture asked 25,590 people in 16 countries what they would hand to an AI agent. 74% said they would have one negotiate a deal, resolve a complaint, renew a subscription or reorder something. You can read it in the Consumer Pulse Research 2026.

The line in that report I keep rereading is further down. Consumers are most willing to give an agent autonomy in two stages of the journey: negotiation and post-purchase.

That is your retention desk and your complaints line. The two places in most companies where the process quietly counts on the customer running out of patience.

Friction was part of the plan

Plenty of support lines were built to wear people down. Sometimes deliberately, sometimes through ten years of small decisions. The menu that loops. The hold music. The save offer that shows up exactly when you have decided to leave. All of it works because humans have a dinner to get to.

An agent does not have a dinner to get to. It can sit on hold for an hour and call again the next morning. State of AI’s summary of Business Insider’s reporting mentions bots that call back just to confirm a hotel reservation one more time.

So if part of your cost model assumes a percentage of callers hang up, that assumption is aging fast.

It is already on the phone

TTEC, one of the large contact center outsourcers, started hearing odd-sounding callers about 18 months ago. Its quality team eventually worked out they were AI. James Bednar, TTEC’s VP of product and innovation, told Business Insider he expects some brands to stop accepting agent calls altogether, because the volume “could be astronomical.”

A lot of the early traffic is business to business: medical practices sending AI to call insurers for preauthorization, bill-negotiation services replacing human negotiators. A retention rep at a credit card company said some bots seem to be mapping the phone tree, asking how callers reach her team.

And my favorite detail from the whole story: a claims adjuster at a California insurer gets the same bot several times per shift, calling for accident attorneys, asking about airbags and body shops. Her company’s quality system scores her empathy on each call. It has docked her for not being nice enough to the bot.

What breaks

Your metrics, first. Handle time, CSAT, quality scores, containment. Every one assumes a person on the other end, and you staff and pick vendors off those numbers. Repeat callers who never get angry drag them in directions nobody can explain in the monthly review.

Then authentication. A bot can’t tell you the name of your first pet, and you have good reason to be suspicious, because scammers get the same tools and can place thousands of calls a day. But every extra check you bolt on to stop bots lands on your human customers too. Alorica’s co-CEOs are already talking about agent registration, possibly through ID.me, so a company can confirm a bot actually works for a real person.

Then retention. A save offer is a bet that the caller will take the first reasonable thing to end the call. An agent instructed to get the best deal will not.

Where I would start

Kahlil Lalji argued on X that agents will push banking activity toward the most programmable institutions, since “agents do not care about branches, brand affinity, or decades-old banking relationships.” That is his forecast, not a measured fact. I think the logic travels well beyond banks, though. If the customer’s agent can choose who to deal with, it will choose whoever answers through an API instead of a nine-option menu.

At IQ Source the first pass is a short list. Which requests can an agent safely complete, and which need the account holder? Do you detect agent calls today, and do you report them separately? What is a customer’s agent allowed to commit to, and what is yours allowed to commit to? Do you give agents a verified channel, or make them use the phone?

That last one carries the most weight. A programmatic channel with explicit permissions keeps the phone line for people and gives you something you can audit. We saw the same security lesson when Cloudflare and 37 other companies built payment rails for agents and every facilitator researchers tested had violations. And to know whether the channel works you will need evals written against your own processes, because an empathy score clearly won’t tell you.

If your support team has started mentioning calls that sound a little off, design the door for agents this quarter, before the agents pick one for you.

Let’s design your agent channel