Most AI-in-CX conversation is about the enterprise side: AI agents handling calls on behalf of the company. A second shift is starting from the opposite direction. Consumers are beginning to send AI agents to contact companies for them.
The products are already in market. Google's "Ask for me" feature places AI calls to local businesses on a consumer's behalf, announcing itself as automated at the start of each call. Duplex has booked restaurants by phone for years. Contact center leaders are describing AI assistants that call in to negotiate for customers. Others, like ChatGPT's agent, work through websites instead, one more sign that voice isn't where agents want to be.
Adoption is early. One recent survey of nearly 2,000 consumers found that at most one in sixteen say an AI tool has contacted a company for them. But the direction is clear, and any company serving consumers should expect to field these calls regularly. The receiving side hasn't asked what that means for how CX is measured, how reps are allocated, or whether voice is still the right channel when both ends of the call are software.
A company's own AI agent is built around the company's goals: containment, resolution, cost. A consumer's agent works for the consumer. Cancel the subscription. Get the refund. It doesn't need warmth, has no patience to run out, and pushes until it gets its outcome.
Picture a consumer's agent calling a telecom provider to cancel a line. It authenticates, navigates the retention script, declines three save offers, and gets a confirmation number in six minutes. To the rep, it's a weirdly efficient, unusually cold caller. To the dashboard, it's a short call with a successful disposition. Neither captures what happened: the company spent a trained retention rep's time on a caller that no offer was likely to move.
Every component of a modern QA rubric was calibrated for human-to-human interaction. CSAT asks whether the customer felt heard. Empathy scoring looks for warmth. Hold tolerance is measured against human patience. None of it applies when the caller is a bot.
CSAT is noise. The bot has no feelings about the call. Any post-call survey will be ignored or auto-completed to satisfy its termination condition.
Empathy scoring misfires. A rep who was efficient and transactional with an AI caller did the right thing, and the rubric penalizes them for it.
Handle time flatters. A short bot call looks like a win, but the metric can't show that no rep needed to take it.
Worse, the aggregates lose meaning. Suppose, for illustration, AI-initiated calls grow from 5% of inbound to 20% in a quarter. CSAT, empathy and resolution all drift, and leadership can't tell whether operations changed or the customer base did.
First, a caller-type signal. Classify every inbound interaction as human, AI agent, or ambiguous, and carry that through every metric: CSAT, resolution and rep performance, each segmented by who or what the rep served.
Detection works best as a layered signal. Some agents identify themselves: Google's calling feature announces it is automated, and regulators, including in the EU, are moving toward requiring disclosure. For the rest, no single acoustic cue is reliable, and tells like flat pacing will fade as voice models improve. Dependable classification combines acoustic, behavioral and call-metadata signals, returns a confidence score instead of a verdict, and treats "ambiguous" as a real category so callers with speech differences aren't misrouted.
Second, a different definition of success. A human wants to feel heard. An AI caller wants a structured answer fast, so measure task completion accuracy and response latency, not CSAT.
Third, separate reporting. Blending human satisfaction and AI task completion into one aggregate is meaningless. Caller type has to be a first-class dimension in the data model and in leadership's views.
Once a company can reliably spot an AI caller, the harder question surfaces: why is this call happening over voice at all?
Voice is expensive. It occupies a rep, can't be parallelized, and adds transcription error and latency, all to communicate information that is almost always structured underneath: an order number, a date, a policy question with a bounded set of answers. Voice makes sense for humans. For an AI caller, it's the slowest, most error-prone way to exchange data both sides could trade directly.
AI-initiated contact belongs on a different channel: a documented API, an MCP endpoint, or any protocol where a consumer's agent can authenticate for its user and transact directly. Reps serve humans. A structured endpoint serves bots.
The math is stark. A live voice interaction typically costs a contact center somewhere between six and twelve dollars once rep time, telephony and QA are loaded, per industry-standard figures. An authenticated API call costs cents. Moving AI-initiated volume to an endpoint cuts that slice's cost by an order of magnitude and frees reps for the humans who need them.
Why would an agent use the endpoint? Agents default to the phone because it's universal. Companies will need to publish the structured option, route to it when they detect an agent, and lean on emerging agent-authentication standards. The phone becomes the fallback.
Routing agents to an endpoint raises real questions.
Churn friction. Some companies make cancellation hard on purpose, and click-to-cancel rules may force the issue.
Authorization and fraud. The endpoint must verify an agent truly acts for the consumer and screen out impersonation, including cloned voices.
Disclosure and fairness. Companies should be clear about whether and how they treat agent callers differently.
None of these argue against the shift. They are the reason to design it deliberately.
In the near term, consumer agents will be a small fraction of inbound volume and most companies won't notice. Reps will field the calls, dashboards will quietly drift, and auto-completed surveys will be treated as real CSAT.
In the medium term, the volume becomes impossible to ignore. A category of inbound traffic that behaves nothing like human customers starts to make up a measurable share of the queue. Reps call certain callers "weird," QA scores stop tracking performance, and leadership asks why CSAT is drifting when nothing in operations has changed. And as more businesses answer with AI of their own, more calls become agents talking to agents.
In the longer term, early movers have a detection signal, segmented analytics, and a structured endpoint at a fraction of the cost of a voice line. Everyone else is staffing call centers for bots.
The strategic question isn't whether to deploy AI agents on your own side; that decision is already being made. It's what you'll do when the next ten thousand callers are not people. The companies that answer it before the volume forces them to will set the architecture the rest of the industry has to live with.
CSAT, empathy scoring, and hold tolerance were calibrated for human callers and do not apply to AI agents, which have no feelings about the call and auto-complete surveys to satisfy termination conditions. Companies need a caller-type signal that classifies every inbound interaction as human, AI agent, or ambiguous, a different success definition for AI-initiated calls focused on task completion accuracy and response latency, and separate reporting so human-caller satisfaction and AI-caller task completion are never blended into one aggregate. Over time, the honest architectural answer for AI-initiated contact is a documented API or MCP endpoint rather than a voice line, which cuts per-interaction cost by an order of magnitude and frees reps for the humans who need them.
