A customer emails support. The reply comes back signed by Jordan, their account manager. Jordan never saw it. An AI agent wrote it and sent it from Jordan’s account.
That setup is common. In our September survey, 756 respondents told us they run an AI agent that works with customers. 26% say the agent acts through a human employee’s account. Another 32% run a mix, where the agent sometimes uses its own identity and sometimes an employee’s. Put together, 58% of teams have a customer-facing agent that sends as a person at least some of the time. 39% give the agent a visible identity of its own.
Of 756 respondents with a customer-facing agent, 26% say it acts through an employee’s account and 32% run a mix of both.
Email is where agents blend in
Customer-facing agents work in email more than any other channel. Of the 813 respondents we asked, 78% named email, followed by chat (74%), SMS (44%), and phone (38%). A chat widget usually labels the bot. An email has a From line, a signature, and a reply-to address, and all three say whatever the sending account says. If that account belongs to an employee, the customer is talking to that employee.
Channels: 813 respondents asked about customer-facing channels. Incidents: 1,028 respondents.
Setup decides the default
We asked how agents are set up. We didn’t ask whether customers are told. A team can add an AI disclosure to a message sent from an employee’s inbox, and an agent with its own address can still sign as “Sarah from Support.” Setup matters because it decides what the customer sees when nobody adds anything. Send from an employee’s account, and the default is a person.
When the agent gets it wrong, a person takes the blame
Agents already make customer-facing mistakes. In the past 90 days, 18% of respondents had an agent send an incorrect or inappropriate email, and 18% had one respond inappropriately to a customer. (We covered the full list in what the agents did wrong.) Behind an employee’s account, the angry reply goes to Jordan, and so does the audit trail. Teams whose agents operate through OAuth on an employee’s account report full audit capability at 51%. Teams that give agents a dedicated identity report 66%.
Share reporting full audit capability, by how agents authenticate, n=1,029.
Respondents know where the risk sits. Asked for the biggest risk of giving an agent email access, 35% named exposure of PII or confidential information and 30% named the agent sending something incorrect or unauthorized. Between December and September, comfort with agents emailing customers without sign-off fell from 49% to 44%.
The mixed setup is the hardest to defend
The 32% running a mix have two problems. Customers can’t tell which messages came from a person, and unless every message is tagged when it’s sent, the team can’t tell either. One consistent identity per workflow fixes both, whichever identity the team picks.
Teams know where agent identity belongs
Asked where a dedicated agent identity would be most valuable, 41% chose customer support, second only to internal operations and reporting at 48%. Wanting one and getting one are different steps. Among organizations that haven’t given agents a dedicated identity, 56% say security review is what stopped them.
Regulators are asking too
Since August 2, 2026, Article 50 of the EU AI Act has required providers of AI systems that interact with people to make clear that those people are dealing with AI. Maine and New Jersey have their own rules requiring bot disclosure in commercial interactions. “The agent sent it from Jordan’s account” is a hard answer to give a customer, an auditor, or a regulator.
Five decisions to make before an agent emails a customer
Choose one identity per workflow and keep it for the whole thread.
Disclose where the reader looks: the From name, the signature, or both.
Tag every agent-sent message at send time, so your logs separate agent actions from employee actions.
Decide where customer replies go: to the agent, a person, or both.
Treat customer email as a write. 59% of teams already let agents read on their own and require approval for writes.
Where Nylas fits
Nylas Agent Accounts give each agent a real mailbox and primary calendar of its own, so the From line, the reply path, and the audit log all name the agent instead of a person. A policy engine, DLP, and a rule audit log give security reviewers something concrete to approve.
Email and Contacts API 
Calendar API & Scheduler 
Notetaker API 
Agent Accounts 
Sales & CRM 
Recruiting & ATS 
Healthcare
Financial
Legal
Real Estate
Education
Field Services
Docs
SDKs
Guides
Integrations
API status
Changelog
Support
Professional Services
