Your texts and emails, answered against real data.
A customer texts "can you move Thursday to Friday?" The agent finds that job in your system, checks whether Friday works, moves it, and confirms. Not a canned auto-reply and not a chatbot that takes a message.
Same brain as the voice agent. The tools, the roles and the safety rails are shared code, so a fix on one channel is a fix on both.
Two channels, one agent
SMS
Comes in through the SMS number you already use, so opt-outs, metering and threading are the ones the platform already handles.
Comes in through a connected mailbox. The agent reads the message, finds the job, and replies on the thread. Most field service tools cannot answer an email at all.
It does the work, not just the talking
The agent has tools: it reads and changes your data mid-conversation, within the limits you set.
Knows who is texting
Resolves the number or address to the contact, so it answers about their jobs, not in general.
Books and reschedules
Reads availability, moves the job, writes it back to the record.
Respects business hours
A weekly grid in your own timezone, DST-correct, with holidays. Off means answered around the clock.
Answers from your FAQ
Hours, service area, pricing, policy. Authored once and shared with the voice agent.
Reads your documents
Optionally searches your ingested documents, the same store the document Q&A uses.
Hands off to a human
Escalates when it should, rather than guessing at something it should not answer.
What happens when it gets it wrong
The honest question about letting AI talk to your customers. Here is the actual mechanism, not a reassurance.
Nothing commits without a typed confirmation
Anything that moves money, cancels work, or notifies a customer stages a preview first and quotes it back. It commits only when the customer replies with the confirm keyword. The check runs in the runtime against the raw text, not in the model, so a model that believes it was confirmed still cannot commit. Matching is whole-word, so "send me a confirmation email" confirms nothing.
STOP never reaches the AI
A bare STOP, UNSUBSCRIBE or CANCEL is handled as platform business and the provider's own opt-out runs. Replying conversationally to STOP is exactly the compliance failure that keyword exists to prevent. Outbound replies also check the opt-out list independently, so an opted-out contact cannot be messaged even if a conversation is somehow still open.
Two bots cannot volley forever
Email has no hangup, so auto-replies, bulk mail, bounces, no-reply senders and out-of-office subjects are dropped before the agent sees them. Behind that, a per-hour reply cap counted by address, not by conversation, so closing and reopening cannot reset it. And a message that looks like it came from the agent's own mailbox is dropped outright.
It can only touch the tables you allow
The agent's queries run under a role whose grants are the tables you allowlisted. That boundary holds regardless of what the instructions say or what a customer tries to talk it into.
Every confirmation attempt is logged with what was asked, what arrived, and whether it counted. Reply caps and an optional monthly ceiling sit in front of the model, and crossing the ceiling suspends the channel rather than quietly spending.
AI and humans, same job
The agent takes the fiftieth "what time are you coming?" of the day. Your team takes the conversation that needs a person. Neither is doing the other's work.
Pair this agent with
Most teams adopt one agent first, then layer in two or three more in the same week.
Works in every trade