Every provider in this comparison now pitches at agent builders — Sendblue leads with it, Blooio sells an inbound-only plan designed for reply-first agents, LoopMessage prices a cheaper shared sender specifically for assistants. The infrastructure question is settled. The design question is not.
What an agent is genuinely good at here
- Speed. Replying in eight seconds at 11pm on a Sunday is something no small business can staff, and in lead response that is most of the value.
- Structured replies. Confirming, rescheduling, checking a booking, answering opening hours.
- Triage. Working out what a message is about and routing it to the right person with context attached.
- Drafting. Proposing a reply for a human to send, rather than sending it. Underrated and much safer.
What it should not do
| Situation | Why a human takes it |
|---|---|
| Anything about price beyond a published rate | Negotiation is a relationship act, and errors are expensive |
| A complaint, or any message with frustration in it | An agent handling anger correctly is not a bet worth taking |
| Clinical, legal or financial specifics | Liability, and often a regulatory line |
| Anything it has already failed once | A second wrong answer converts confusion into distrust |
| A message it is not confident about | Silence for two minutes beats a confident wrong answer |
The two-message rule
An agent gets two messages. If the thread is still open after that, a human owns it.
This is the most useful constraint we know for small business deployments. The first message buys response time. The second handles the common structured case — a confirmation, a reschedule, a simple question. Past that, the conversation is doing something that needs a person, and the marginal value of automation goes negative fast.
Northgate Realty
iMessage
Hi Sam, this is Elena at Northgate — got your note about Oak St. Free for a walkthrough this weekend?
Saturday morning works. also is the roof original?
10am works — I'll check on the roof and come back to you shortly.
Disclosure
Do not have the agent claim to be a person
Several US states now require disclosure when a consumer is interacting with an automated system, and the rules are tightening. Beyond the legal exposure, a customer who discovers mid-thread that 'Elena' was software feels deceived in a way that damages the relationship permanently. Using a real staff member's name for an automated opener is the specific pattern to avoid — either disclose, or have the named person actually be the one replying.
Practical build notes
- Opt-out handling runs before the model, always. It is a deterministic branch, never a judgment call. Implementation in the automation walkthrough.
- Log every generated message. You will need to read them back when something goes wrong, and something will.
- Set a confidence floor and escalate below it. Escalation is cheap; a wrong confident answer is not.
- Rate-limit the agent per contact. A loop that sends four messages in a minute is how a line gets flagged — see why numbers get flagged.
- Use typing indicators. A reply that lands in 400ms reads as a machine even when the content is perfect.
The measurement that matters
Not containment rate. Containment — the share of conversations the agent handled alone — rewards exactly the behaviour you do not want, which is an agent refusing to hand off. Measure outcomes instead: bookings, resolution, and the share of threads where a human had to apologise for the agent.