Most agent-scale stories you read this year will be projections, vendor decks, and analyst forecasts. This one is twelve months of actual production.
Legacy is a healthcare growth team that went live on Atlas one year ago. On day one they had a single thing turned on: email-only support deflection. Atlas drafted responses to inbound customer email; their team approved each draft before it left the building. Nothing else. No chat widget. No voice. No integrations. The smallest possible version of the system, with the operator firmly in the approval seat.
Twelve months later, Legacy is running four channels on the same Atlas substrate: email, web chat, voice agents on inbound calls, and integrated agents that write to their CRM and kit-ordering system. Total case volume processed across all channels, month twelve versus month one: roughly 8x. Their team has been in the approval seat the entire twelve months.
The order matters
Most teams trying to scale agents do it in the opposite order. They start with the demo-friendly surface (a chatbot, a voice agent, an autonomous deal-closer), wire it into everything they have, and discover the failure modes in production. Legacy went the other way.
Month one. Email. The hardest channel to embarrass yourself on. Email is async. Mistakes get caught before they ship because the operator is reviewing each draft anyway. Email built the muscle: the team learned what Atlas does well, what it gets wrong, what their source documents say versus what they thought they said.
Months two to four. Web chat. The next obvious step once email proved itself. Visitors on the site asked questions in the moment instead of writing email and waiting. The knowledge base from email transferred directly. On the same traffic that existed before, the chat surface produced 6.1x more inbound conversations, a 25.2% buyer-intent rate, and ~94% deflection on repetitive questions. The substrate compounded: the documents Atlas had been quoting in email were the same documents it quoted in chat.
Months five to nine. Voice agents. The phone is where regulated industries lose customers when there is no answer at 9 PM on a Friday. Voice agents on inbound calls let the team handle inquiries in the moment with the same knowledge base and the same approval boundaries. Voice is harder than chat because the channel is real-time; the work the operator does is reviewing what was said after the fact and flagging anything the agent should not repeat.
Months ten to twelve. Integrated agents.The big lift. Agents that read from Legacy’s CRM and write back to it. Agents that talk to the kit-ordering system. Both built from connectors in the Clarm catalogue, not from custom engineering. This is where most enterprises stall, because integrated agents are the ones that can break things. The reason it worked here is the previous nine months. By the time the integrated agents shipped, Legacy already trusted the approval queue, already trusted the source receipts, already had a year of audit-log evidence about how the system behaves. The architectural bet was paid back on every previous channel.
Why this is the agent-scale playbook
~8x more enterprise apps will integrate task-specific agents by the end of 2026 than at the start of 2025. 88% of organizations that have shipped agents in the last year have already reported a security incident. About 14% of agents reach production with full security and IT approval. The gap between executive confidence in agent controls (82%) and the reality (14%) is the conversation every CIO is having in May 2026.
The standard answer is to put an “agent governance layer” in front of the agents you already shipped. That works the same way putting a firewall in front of a leaking application works: not as well as you hoped, and you still have the leaking application.
The Legacy answer is the substrate-first one. The governance is invariant: source receipts on every answer, approval gate before any agent action reaches a customer, audit log on every retrieval, draft, and model call, tenant-isolated retrieval at the database layer. When a new channel ships, it inherits the invariants. When a new agent ships, it inherits the invariants. The team scales the workflow surface without re-litigating the security posture.
What this lets a team do next year
The honest version: Legacy in month twelve is doing things they did not plan to do in month one. Voice agents were not on the year-one roadmap. The CRM integration was not on the year-one roadmap. The kit-ordering integration definitely was not on the year-one roadmap.
That is the point. The substrate is what lets you say yes to the agent your team needs next quarter without re-doing the security work, without re-doing the audit work, without re-doing the integration work. The Clarm catalogue keeps growing; every connector is one more thing your agents can read from or write to without engineering touching it.
The math on the next year, if the pattern holds: more channels, more integrated agents, more throughput per operator-hour, same governance invariants. That is what scaling agents safely looks like from the inside.
See the architecture if you want the substrate details, or book a pilot discussion to talk about which channel makes sense to start with for your team.