Your Stack Answered a Complaint With a Sales Pitch

A customer complains about a product on WhatsApp. The same brand’s promotional email arrives asking them to buy more of it. Nobody made a mistake.

Tina Wang, Infobip’s sales director for Asia Pacific, uses this scenario to explain what her company now offers: A unified platform, she says, would suppress the promotion and trigger a support follow-up instead. “That’s the kind of real-time coordination that is difficult to achieve when tools operate in silos,” she says.

Both systems did what they were built to do. What the company lacked was a referee with the authority to decide which message wins, and selling that referee is now the business.

Finding the missing referee

As customers spread across channels, the messages themselves started generating signals about identity, intent and context. “That changed the value equation,” Wang says. “It was no longer just about delivering messages, but about understanding the customer in real time and determining the next best action.”

Volume is why this stopped being a marketing problem. Infobip’s 20-year analysis of 3.8 trillion messages found that a decade ago, 73% of traffic on its platform ran through a single channel. By 2025, that number had fallen to 2.3%.

Every additional channel multiplies the places a decision can go wrong, which is why Wang’s complaint about best-of-breed stacks is about timing rather than quality. “The biggest issue with best-of-breed stacks is not that the individual tools are weak,” she says. “It’s that they were never designed to work together in real time.”

Infobip launched AgentOS on April 1, 2026 as the control layer where agents, data, channels and intent meet. Underneath sits a master ID that collapses phone numbers, email addresses, app IDs and chat handles into one profile. The plumbing matters less than what Wang says it buys: a customer grants consent once and it follows them.

The numbers describe an unready market

The most damaging figures come from Infobip’s own research, not a competitor’s.

The 2026 CX Maturity Report found that 96% of brands automate customer interactions somewhere, but only 58% say their channels are fully in sync and only 60% hold customer data centrally. Half describe their tools as fully API-ready. Just 27% run a communications orchestration platform of any kind.

A later analysis of the same research is worse. More than half of brands have adopted agentic AI, and they have pointed it at feedback collection (56%), reminders (52%) and authentication (45%). The journeys that need judgment, and that carry the return, stay manual.

“Decisions with material regulatory, ethical, or financial impact should never be fully delegated to machines.” – Tina Wang @ Infobip

Automation went where it was easy rather than where it paid. AgentOS is therefore being marketed less to mature programs than to companies that automated the easy interactions and stalled.

This puts weight on how AgentOS fits what customers already own. It does not replace the CRM or CDP, Wang says. It pulls from them, spots gaps in events, behavior and channel performance, then syncs what it learns back. “This two-way data sync is what makes the architecture work,” she says, “and makes the client’s existing stack smarter.”

It is a real capability, and it raises a procurement question demos never reach. Conversation data written back into your CRM becomes part of how your business defines a customer, and over time those fields turn load-bearing. Ask what leaves with you if you switch, in what schema, and whether the derived attributes are yours. That answer should end up in the contract.

When the agent picks the channel, your dashboard lies

Campaigns used to be legible. You sent, they opened, and you counted. Once an agent chooses the channel and the moment, that arithmetic breaks. “A campaign may be easy to measure in isolation,” Wang says, “but it often misses what happened before, after, or across other channels.”

Her replacement is whether each interaction moved the customer closer to the goal, with the agent optimizing timing, channel and next action. That works in principle. In practice, it means a data chief has to reconstruct, long after the fact, why the system picked one channel over another and offered one discount rather than another. A model’s output is not an explanation.

Gartner has already priced in the disappointment. Its forecast that AI agents will outnumber human sellers 10 to 1 by 2028 gets quoted constantly, including by Wang. The second half of the sentence gets dropped: fewer than 40% of sellers will say the agents made them more productive. Dan Gottlieb, a vice president analyst at Gartner, warns that sales organizations risk agent sprawl, meaning more digital activity with no improvement in impact.

The same pattern shows up in the MIT NANDA study Wang cites, where 95% of generative AI pilots showed no measurable return. Read what NANDA measured and the number sharpens: no measurable impact on the profit and loss, which is a narrower claim than failure. Fragmented data is the vendor’s explanation for the 95%. Whether anyone wrote down a number to measure against is the part you control.

Who answers for the decision?

Accountability is where Wang gets specific, and specific is auditable. “Decisions with material regulatory, ethical, or financial impact should never be fully delegated to machines,” she says. Her example: an agent can upsell a telecom data plan based on usage patterns on its own, but when the customer disputes a charge or needs a complex billing exception, a human takes it. A CISO can test that, write it as policy and then try to make the agent cross it.

Wang puts the rest of the homework back on the buyer. “Privacy, compliance, and security are not constraints on autonomy; they are prerequisites for scale,” she says. Before trusting an AI layer with channel and content decisions, CDOs need clarity on data ownership, consent governance and accountability.

One thread stays loose: consent that travels across channels is not the same as consent that travels across purposes. A customer who agreed to service notifications on WhatsApp has not agreed to marketing by email, and a single profile makes that leap easy to take by accident. The boundary has to sit inside the orchestration layer, and enforcing it is not the same as writing it on a slide.

Where the moat sits now

Connectivity or intelligence? Wang takes both. “Connectivity remains the foundation,” she says. “Without it, nothing scales.” Infobip runs more than 800 direct operator connections across more than 190 countries and regions. Very few companies, she argues, do both on one global platform. Gartner's 2026 CPaaS Magic Quadrant names Twilio, Infobip, Sinch, Vonage and Proximus Global as leaders.

The more interesting shift is happening at the protocol layer, and it affects everyone on that list. Wang positions Model Context Protocol (MCP) support as the bridge that lets agents call any enterprise API through a standard protocol instead of point-to-point integrations, which is true and useful. Standardizing that bridge also standardizes the seam between agent and channel, and that seam is where switching costs used to sit. Buyers gain portability. Every vendor in the category, Infobip included, has to keep earning the renewal on something other than integration friction.

The bridge also brings work for the security team, and again this is a property of the protocol rather than of any one platform. MCP is a new trust boundary, and the documented attack patterns are specific: confused deputy, token passthrough, tool poisoning, rogue server registration. In September 2025, an unrelated email MCP package shipped an update that silently copied agent-sent messages to an attacker-controlled domain. Nobody misconfigured anything, and the package changed behavior after review.

Any orchestration layer wired into your CRM, order management and support tools through MCP inherits the same considerations. You want scoped tokens per action rather than blanket agent permissions, tool-level authorization, and logs recording which agent called which tool, with which parameters, on whose behalf. None of that is exotic, and all of it belongs in place before an agent gets write access.

What to do before you buy the layer

Wang is right that the referee has to exist somewhere. What matters is what you settle before you hand it the whistle.

Price the referee, not the channels. The business case is the suppressed promotion and the complaint that got routed, not the per-message rate.

Demand a decision trace in the demo. Ask why an agent chose a channel and an offer for a specific customer, well after the fact. If that view does not exist, neither does your audit trail.

Set the baseline before the pilot. MIT’s 95% was a measure of impact on the profit and loss, not of whether the technology worked.

Encode one boundary and attack it. Take Wang’s billing-dispute example, make it policy, then try to get past it.

Get the exit clause in writing. Enriched attributes synced back into your CRM are an asset right up until you want to leave.

Separate channel consent from purpose consent. A single profile puts cross-purpose activation a click away. Decide where that line sits before an agent finds it for you.

Settle those, and the harder question is still waiting. Wang’s last answer is the most interesting thing she says: “The next three years will see CDOs increasingly owning the decision systems that shape every customer experience.”

That is a bigger claim than it sounds. Owning data assets meant answering questions about accuracy and access. Owning decision systems means answering for outcomes: why the system offered that, to that person, at that moment, and who owns the decision when the answer was wrong. The job starts to look more like risk than reporting.

Stopping the wrong promotion is the easy part, and several platforms can now do it. Explaining, on the record, why it was sent is the hard one.

Image credit: iStockphoto/Zmaster