AI Agents Got More Channels and More Nerves This Week

Two things happened on May 27 that sit in awkward tension with each other.

Aircall launched AI Messaging Agents, extending its voice AI platform to handle inbound SMS and WhatsApp conversations autonomously on existing business phone numbers. The same day, Talkdesk rolled out proactive AI agents for retail and financial services, designed not to answer calls but to make them, recovering abandoned carts, pushing loan applications, and chasing delinquent accounts.

Also the same day: a survey of 2,500 senior decision-makers found that 75% of enterprises have rolled back, paused, or shut down AI agent deployments they already shipped. Not because the tech lacks potential. Because hallucinations and data exposure are burning them in production.

That is the real story this week. The platforms are expanding (more channels, more initiative), and the buyers are contracting (more caution, more governance). Both are happening at the same time, to the same market.

Aircall: voice becomes multichannel

Aircall's move is straightforward in concept. Its AI Voice Agents, which launched earlier, handle inbound phone calls autonomously. AI Messaging Agents do the same thing for SMS and WhatsApp, on the same phone numbers, in the same workspace. When a customer texts a business, the agent reads the message, pulls from the company knowledge base, replies in natural language, and can execute actions like looking up a Shopify order or creating a HubSpot deal.

Tom Chen, Aircall's CPO, framed it as collapsing channels into a single surface: "AI Voice, SMS, and WhatsApp now run through one platform, one knowledge base, and one workspace."

The target use cases are the ones you would expect: lead qualification for sales teams (the agent asks qualifying questions and books meetings), routine support resolution (order tracking, status updates), and after-hours coverage. Aircall says 23,000 companies already use its platform for over 2 billion interactions per year. Adding messaging to voice on the same number is the kind of incremental expansion that makes existing customers stickier.

What is worth noting: the agent hands off to human agents when the conversation gets complicated, and the full message history transfers with it. That handoff quality, not the AI's ability to reply to "where is my order," is what determines whether the feature actually works or just annoys people who get trapped in an AI loop.

Talkdesk: inbound meets outbound

Talkdesk is doing something more ambitious, and riskier. Its new proactive AI agents sit within the Customer Experience Automation (CXA) platform and are designed to initiate outbound contact rather than just respond to it.

For retail: cart abandonment recovery (the AI calls or messages the customer in real time, walks them through checkout) and product recall management (automated high-volume outreach with compliance tracking).

For financial services: loan pre-qualification outreach, deposit growth campaigns, and early-stage collections. Each of these involves regulatory constraints around lending disclosures, collections compliance, and product recommendations. Talkdesk says its workflows embed these constraints. The company did not disclose how compliance is audited in production.

The shift from inbound to outbound is the part worth paying attention to. A customer who calls support has signaled intent. An AI that calls a customer is guessing at intent. The line between "helpful proactive outreach" and "AI telemarketing" is not theoretical. Talkdesk will need to show that its agents stay on the right side of it, and that means publishing error rates, not just capability lists.

Talkdesk, valued at roughly $10 billion, counts Canon, Sysco, and Kimberly-Clark among its customers. Several financial institutions, including Credit Union 1 and Emprise Bank, are already on the platform.

The 75% rollback number

Here is where the week's narrative gets complicated.

A global survey of 2,500 senior business decision-makers, reported on May 27, found that nearly three-quarters of enterprises have pulled back live AI agent deployments after encountering problems in production. The biggest concern, cited by 33% of respondents, was customer data exposure: AI agents revealing sensitive information during conversations. The second concern, at 22%, was hallucination-driven brand damage: AI systems generating confident but wrong information in customer-facing contexts.

In banking, healthcare, legal services, and telecom, a single bad AI response can trigger regulatory consequences. The survey found that many enterprises are now restricting AI agents to narrowly defined tasks, adding approval layers, and moving toward retrieval-based architectures to ground responses in verified data rather than model inference.

This is not a rejection of AI agents. It is a recalibration. Companies are still buying. They are just deploying more carefully, and in some cases pulling back to internal use cases where human oversight is easier to maintain.

What this means for buyers

If you run customer operations and are evaluating AI agents, this week captures the exact moment you are living in. Vendors are expanding what agents can do (messaging, outbound, financial services workflows) while the data from early adopters says "proceed, but with actual guardrails."

Three practical takeaways from the convergence:

Handoff quality matters more than channel count. Aircall adding SMS and WhatsApp is useful only if the AI-to-human handoff actually works. Test the handoff before you test the AI's ability to answer questions. A customer who gets stuck in an AI loop after texting a simple question is worse off than one who waited on hold.

Outbound is harder than inbound. Talkdesk's proactive agents are interesting for cart recovery and collections, but the regulatory surface is different from inbound support. If you are in financial services or healthcare, ask the vendor exactly how compliance is verified in production, not just what the agent is capable of doing.

The 75% rollback is not a reason to wait. It is a reason to scope. The enterprises that pulled back are the ones that deployed too broadly too fast. The ones that are succeeding are narrowing the agent's scope to specific, measurable tasks: intake qualification, appointment scheduling, order status. Scope down, deploy, measure, then expand.

The platforms are moving fast. The data says the buyers who win will be the ones who move deliberately.

---

Sources: Aircall AI Messaging Agents announcement | Talkdesk proactive AI agents | Enterprise AI rollback survey | Talkdesk GlobeNewsire press release