Nylas

Two products can both call themselves an AI email assistant. One drafts a reply and waits for you to approve it. The other sends it, books the follow-up, and updates the CRM without asking someone. Same label, different systems, and the controls aren’t interchangeable. When a workflow pauses for review, a bad output has somewhere to get caught. When it doesn’t, the mistake is out in the world before anyone notices. That’s also where prompt injection stops being theoretical. The agent is built to read and act on whatever lands in the inbox, and strangers decide what lands there. Our Director of Information Security published a new post on where that line sits, and how permissions, agent identity, human review, monitoring, and audit trails need to change once a workflow crosses it. 👉 https://lnkd.in/gUGiTRUT

"We take data privacy seriously" is not an answer to "Do you use our data to train your models?" That's what our new guide addresses: 10 specific questions enterprise buyers can ask AI vendors during security and procurement review, along with the red flags that indicate a vendor has a policy instead of a process. Questions cover subprocessor disclosure, what data AI features actually access, whether AI activity is distinguishable from user activity in audit logs, and how vendors handle prompt injection and excessive permissions. Includes a downloadable checklist for documenting vendor responses. 👉 Read the full guide: https://lnkd.in/g8uiWMSn

Every recruiting product team building on Nylas has shipped features for human identities: recruiters, interviewers, hiring managers. Then someone proposes an AI recruiting agent, and you hit a question the product doesn't have an answer for. Whose account does the agent send from? An autonomous agent is the first actor in your product that isn't a person. It needs an identity of its own, which is what Agent Accounts provision: an inbox and a calendar on one grant. Because both sit on one identity, the agent handles the coordination. When a panelist RSVPs no to an interview, it sees the decline and rebooks before the candidate ever knows there was a gap. 👉 New post: https://lnkd.in/g-n4bVmt

An application that only checks calendar availability shouldn't have full account access. That's a common OAuth scope mistake in email and calendar integrations, and it's the first thing our new Privacy by Design post walks through. Two more principles from the piece: → Encryption is a baseline, not a complete security posture. It works alongside access controls, monitoring, and vendor oversight. → Token revocation should be immediate on user disconnect, not deferred to the next sync cycle. Nylas also ships Privacy Mode for eligible plans, which can scope calendar access to events created through the Nylas APIs. Full post covers the shared responsibility table between platform and application, plus an 8-item developer checklist: https://lnkd.in/dXpRCGDE

Nylas doesn't train AI models on customer email or calendar data. That's the first principle of our new AI governance post. Two more: → No proprietary foundation models. AI features run on third-party providers subject to our vendor review process. → AI features are off by default. Teams enable them per account, tenant, or workflow. ExtractAI is a good example. It categorizes and routes email using rules customers define, not general-purpose model behavior on unscoped data. What each AI feature touches, plus where our responsibility ends and the deploying team's begins: https://lnkd.in/gz5UWzfc

Ten years ago, building an all-in-one platform for venture capital investors meant raising $5M and hiring a team of twenty engineers. Lighthouse did it with three founders and a coding agent. Sourcing, CRM, LinkedIn outreach, an email client, scheduling, workflows, all in one vertical product. The one part they didn't build themselves is the communications layer. Email threading that works across Outlook and Google. Meeting notetakers wired straight into the calendar. Scheduler powering their own white-labeled booking pages. "You guys are the only third party we use. There's a lot of functionality in it; it's not that you can just build it." Lean teams lose the most time on plumbing. Here's the pattern worth studying. 👉 Read how Lighthouse did it: https://lnkd.in/gQVE9tm9

Agent Accounts gives an AI sales agent its own email address and its own calendar on one identity, so the same agent that sent the cold opener can read the reply and finish the job. That last part is where most outbound agents stall. The prospect writes back, "Sure, how does next week look?" and the agent has no calendar to check and nowhere to put a meeting. So it hands the thread to a human, and the speed you just built drains away while the lead cools. One account carries the whole thing, from the first email to the booked meeting. 👉 New post: https://lnkd.in/gG7acMqg

A Nylas Agent Account calendar sends invitations people accept in Gmail and Outlook the same way they'd accept one from a colleague. It also receives invites at the agent's own address, and RSVPs so organizers see the agent's response next to every other attendee's. Same Events API you already use for connected grants. Same webhooks. Full walkthrough on the blog: creating and canceling events without breaking the invite, routing response logic through event webhooks, checking free/busy before proposing a slot, and the pitfalls specific to running a calendar through automation. 👉 https://lnkd.in/gyBnG8Jv

Most enterprise AI security conversations start with the model. That's the wrong starting point. The model matters, but just as much risk lives in the workflow around it: what it can access, what actions it can take, and who reviews those actions before they run. Same model, different exposure. Much of that difference comes down to governance. This matters most for communications APIs. Email and calendar integrations hold one of the highest-trust positions in enterprise software, connecting to executive communications, customer conversations, schedules, and contacts. An AI workflow that summarizes inbound email needs read access only. Grant it send permissions "just in case," and it can now email customers on behalf of your users. Our new post breaks AI security into four governance problems (permissions, auth, monitoring, vendor risk) and gives you a risk-tiering framework to classify any AI comms workflow before it ships. How are you scoping permissions for AI workflows on communications data? 👉 https://lnkd.in/gS3Qhdg2

Under a day. That's how long it took Wealthie New Zealand to build a working agent inbox on Nylas Agent Accounts. Co-founder Riki Carston built the software his financial advice firm runs on, and email was the bottleneck. Every thread with a client's lawyer or accountant is a compliance record that has to land against the right household, usually with no context attached. Now his routing layer reads each message and files it on its own. Clear references route instantly. Everything else gets inferred, scored, and held for review when it's unsure. He did it solo, in a day, on one API. Anyone building inbound email routing will recognize this problem. Here's how Riki solved it. 👉 https://lnkd.in/guDYWYrA