Addys.AI CEO says meeting assistants will evolve into follow-up agents

Addys.AI CEO Alex Ge says the next wave of workplace AI will move beyond transcripts and summaries into proactive agents that draft emails, update CRMs and book meetings after the call. The shift matters because it could cut the gap between a meeting and completed work across sales, operations and knowledge work.

Why it matters: - Workplace AI is shifting from capturing information to completing follow-up work. - Proactive agents could reduce the manual coordination that still sits between a meeting and a finished workflow. - The change matters most in recurring, multi-step tasks such as sales follow-up, CRM updates and scheduling.

What happened: - Alex Ge, founder and CEO of Addys.AI, said the first wave of AI tools handled transcripts, summaries and drafting, but left the hardest work for people. - Ge defined a proactive AI agent as software that maintains context across tools, infers next steps from a conversation and prepares actions for human review. - Addys.AI said its proactive meeting agent is designed to turn meeting commitments into drafted emails, tickets and CRM updates. - The company is based in New York and San Francisco and was founded in 2026 by Ge.

The details: - First-generation meeting tools cut the cost of transcription, summarization and drafting, but they do not decide who owns an action, gather history from other systems or update records. - AI note takers such as Otter and Fireflies produce summaries, but users still have to interpret the notes and complete the follow-up. - Chat assistants can draft content, but most interactions still begin with a prompt and require the user to copy the output elsewhere. - Ge described the shift from passive AI to proactive AI as moving from conversation to context synthesis, proposed actions and then review and execution. - Ge broke proactive agents into four parts: context across applications, next-action inference, staged execution and privacy with secure compute. - Context across applications means connecting email, calendars, chat, documents, project trackers and CRMs so a meeting can be read against prior messages, decisions and account history. - Next-action inference means spotting commitments, implied tasks and open decisions and turning them into proposals tied to the original conversation. - Staged execution means consequential actions wait for user review, while lower-risk actions can gain autonomy inside explicit permission boundaries. - Privacy and secure compute require least-privilege access, retention rules, audit logs and revocation. - Trusted execution environments, including Intel TDX and confidential GPUs, can reduce exposure during processing, but Ge said they are only one layer. - Rena Labs, the company behind Addys.AI, builds TEE infrastructure as a shared confidential computing layer for such agents. - Addys.AI says its sales workflow drafts a follow-up email from the meeting and account history, updates the deal stage, logs notes and contacts in HubSpot and Salesforce, and books the next meeting, with each action held for approval by default. - Ge said the same pattern can help founders track decisions across recurring meetings and help students manage networking and applications.

Between the lines: - The argument is not really about generating more text. It is about coordinating work across systems with a clear trail from meeting to action. - The biggest technical risk is that an agent may confuse a suggestion with a commitment or act on stale information. - Ge said every proposal needs a source trail and a way to correct errors. - Ge also said access to multiple apps does not equal permission to combine all of their data. - He said summary quality alone is a weak measure of usefulness. - Better metrics, in his view, include accepted action rate, correction rate, time saved after review, task completion and the severity of errors caught. - Ge said the most effective agents will not automate every decision, but will handle the predictable coordination points and keep the line between recommendation and execution clear.

What's next: - Ge expects proactive agents to become a defining direction in AI over the next three to five years. - He said software will move from waiting for instructions to completing work, and buyers will move from paying for access to paying for results. - Ge credited the open-source agent OpenClaw with making agentic software more tangible, while noting that its command-line setup and configuration remain difficult for most users. - Ge said the Model Context Protocol, or MCP, should make connections between AI and everyday software routine. - He expects that shift to create a new kind of super app: a trusted layer that understands context across software, prepares the next step and coordinates the work. - Addys.AI said more information is available on the company's AI meeting assistants page.