From Garden Hose to Fire Hose: Notes from MEF CONNECTS San Jose

AI agents are about to send a lot more messages than humans ever did, and the messaging infrastructure needs to be ready before they do. That was my biggest takeaway from MEF CONNECTS Enterprise Communications USA, held October 5 in San Jose.

Telgorithm was proud to sponsor the event, and I joined a panel on agentic AI, conversational commerce and the future of customer engagement. I shared the stage with Chris Brunner of Authvia, Nitin Seth of Conversive and Ramy Riad of Webex CX, with Matt Ekram of the Mobile Ecosystem Forum moderating. Each of us sits in a different part of the ecosystem, which made for a great conversation. My lens is infrastructure: delivery, registration and compliance. Here's what stood out.

Outbound is where agents hit the wall

Most of the agentic AI use cases we hear about are inbound: a customer messages a business and an agent answers. Inbound rarely strains capacity. Outbound is different. Once an agent starts sending follow-ups, or every inbound conversation automatically triggers an outbound one, throughput becomes the constraint.

On average, one in ten messages runs into a throughput barrier. Picture an agent mid-conversation with a customer when a message simply fails. The customer doesn't know why, and neither does the business. The cause is capacity that wasn't planned in advance.

Content is the other half. Agents are getting very good at holding a conversation, but if what they send starts to deviate from the use case a brand registered? In A2P 10DLC compliance we call that campaign drift, or drift content, and it can create serious compliance problems. Our job as a provider is to get ahead of both issues as this technology rolls out, not react after the fact.

Who owns the mistake when an agent goes off script?

Matt asked the question everyone in the room was thinking about: when an AI agent sends something it shouldn't, who is accountable? Today the answer is the brand. The brand registers the campaign, puts its EIN on it, declares the use case and sends the message. But it’s not black and white, and it gets more complicated with AI.

Detection is the challenge. In any single message, a good brand that slipped and a bad actor look almost identical upstream. One message can’t tell you which is which—what matters is the pattern over time.

Scale makes this urgent. When a human agent goes off script, a supervisor can pull them aside between calls and fix it with a conversation. One AI agent can handle what used to take a hundred human agents, so a problem grows much faster, but fixing it is slower. There's no quick word between calls; it usually takes an engineering change. We're going from a garden hose to a fire hose.

The answer is a chain where everyone understands their role: the brand, the CSP supporting it, the DCA and the network operator. On top of that sits a trust layer built on patterns. A sender with a consistent, healthy history earns street cred, and that track record can justify lighter filtering upstream. That kind of trust will matter more as AI volume climbs.

Let's not repeat the 10DLC rollout

Like I said in my closing remarks on the panel, our industry is good at Monday morning quarterbacking. The A2P 10DLC rollout was a Wild West, and ultimately unregistered 10DLC traffic was cut off altogether. It was so bumpy because it was cleaning up spam that had festered for years. Agentic AI volume will grow far faster than that, so we need to get ahead of it now. What encouraged me in San Jose is that industry leaders are already focused on AI while it's still rolling out. 

Capacity planning is a good example of what needs to change. Today, brands have to guess their capacity before they send a single message. When you register a brand with T-Mobile, for instance, the daily limit starts at 2,000 messages for the whole brand. Some teams assume their agent won't talk much and register for too little. Others pay for extra vetting to avoid a wall they may never hit. Neither guess is right.

The better approach is a system that monitors usage in real time and alerts the software provider, and then the brand, or the agent itself, before it reaches a limit, so vetting can be expanded ahead of time instead of after messages start failing.

Why Conversational RCS fits agentic AI

Right now, Telgorithm's focus is helping businesses adopt and launch RCS. In July we launched our RCS Business Messaging suite, with Rich and Rich Media formats plus Conversational RCS; we're among the first providers to offer session-based Conversational RCS billing.

That billing model is what makes RCS especially compelling for AI agents. Remember that one agent can handle what used to take a hundred human agents. Every one of those parallel conversations draws on the same brand-level daily allowance, and with per-message pricing, every clarifying question and helpful follow-up across all of them adds cost. The bill climbs fast, which quietly pushes businesses to make agents less conversational. Conversational RCS charges a single flat fee for unlimited back-and-forth within a 24-hour window. The economics finally match the way agents actually talk.

It also connects to the reliability issues we discussed on stage. If a device can't receive RCS, our system detects that and falls back to SMS automatically. For high-volume and conversational traffic, our patented Smart Queueing manages throughput so messages get delivered without rate-limit failures. That's the proactive approach to capacity I described above, built into the product.

Getting ahead of the fire hose

Thank you to the Mobile Ecosystem Forum for a great event, and to my fellow panelists for a sharp, honest conversation. Agentic AI is going to change how businesses talk to their customers, and the companies that plan for capacity, compliance and trust now will be the ones whose agents actually get through.

If you're building AI agents that message customers, or you're ready to launch RCS, we'd love to talk. Contact us to get started. You can also watch the full panel on YouTube.