Smart Replies — Let Postbox Answer for You

Most data collection tools stop at storage. A submission comes in, it lands in a database, and you get a notification. What happens next is your problem. You open the email, read the message, type a reply, hit send. Multiply that by every contact form submission, every support request, every “what’s your pricing?” question you’ve answered forty times this month.

We built smart replies because that loop is wasteful. If the answer already exists — in your FAQ, your pricing page, your support docs — there’s no reason a human needs to type it again. Postbox can read your knowledge base, match it against what someone asked, and either draft a reply for you to review or send it automatically.

No training. No fine-tuning. No prompt engineering. Just plain text.

What a knowledge base looks like

A knowledge base in Postbox is exactly what it sounds like: text you write in your own words. Your FAQs, pricing details, return policy, support guidelines, office hours, whatever you want Postbox to be able to reference when crafting replies.

There’s no special format. No YAML. No decision trees. You write it the way you’d write an internal doc for a new team member — “here’s what we charge, here’s our refund policy, here’s how to reset a password.” Postbox reads it as-is.

You manage your knowledge base from the form settings page. Each form gets its own, because the context for a contact form and a support form are usually different. A contact form’s knowledge base might cover pricing and general questions. A support form’s might cover troubleshooting steps and known issues.

Two modes: draft and auto

Smart replies run in two modes, and you pick which one per form.

Draft mode generates a reply and saves it as a draft. You see the submission, you see the proposed reply, and you decide whether to send it, edit it, or discard it. This is the safe default — useful when you’re still building confidence in your knowledge base, or when the stakes of a wrong answer are high.

Auto mode sends the reply automatically, but only when the knowledge base covers the question. If the LLM determines the answer isn’t in your knowledge base — if someone asks something your docs don’t address — Postbox flags the submission for manual reply instead of making something up. This is the key constraint. Auto mode is not “answer everything.” It’s “answer what you know, escalate what you don’t.”

The distinction matters. A lot of AI-powered reply systems hallucinate when they don’t have the answer. They generate a plausible-sounding response that’s factually wrong. We deliberately chose not to do that. If the knowledge base doesn’t cover it, Postbox tells you it needs a human. That’s the only responsible default.

Where smart replies sit in the pipeline

Smart replies are the last step in the processing pipeline. The order is intentional:

1. Spam filtering — junk submissions get blocked before anything else runs
2. Auto-translation — non-English submissions get translated to your language
3. Smart reply — the cleaned, translated submission gets matched against your knowledge base

By the time smart replies run, the submission has been validated against your form schema, spam has been filtered out, and the message is in your language. This matters more than it might seem.

If smart replies ran on raw submissions, they’d waste credits on spam. They’d fail to match knowledge base entries because the submission is in Portuguese and your knowledge base is in English. They’d try to generate replies for submissions that should have been rejected at the schema level.

The pipeline order isn’t arbitrary. Each step makes the next one more effective. We wrote about this design philosophy in our thesis — every layer of processing has context from the layers before it.

How the LLM works

Under the hood, smart replies use an LLM with a temperature of 0.7. That’s high enough for natural, varied responses — you won’t get robotic copy-paste answers — but low enough to stay grounded in the knowledge base.

The model receives the form’s schema context (what fields exist, what types they are), the knowledge base content, and the submission data. Schema context is what makes smart replies actually useful. Because Postbox knows which field is the “message” and which is the “email,” it can focus on the right content and address the reply to the right person.

This is another reason structured forms matter. A schemaless endpoint that accepts arbitrary key-value pairs gives the LLM nothing to work with. A structured form with named, typed fields gives it everything.

Use cases

Contact form deflection. A visitor asks “what’s your pricing?” Your knowledge base has your pricing breakdown. Smart reply sends a detailed response in seconds. The visitor gets an answer before they’ve closed the tab. You never had to open your inbox.

Support automation. A user reports a known issue — say, a password reset email not arriving. Your knowledge base documents the fix (check spam folder, whitelist the domain, try again). Auto mode sends the solution immediately. You only see the ticket if it’s something new.

FAQ handling. Common questions get answered instantly. Uncommon questions get flagged for you. Instead of wading through fifty “do you offer refunds?” messages to find the one edge case that actually needs your attention, you handle only the edge cases.

On-demand via MCP. If you’re using Postbox through an MCP client, the draft_reply tool lets you trigger reply generation on specific submissions programmatically. Useful for agents that want to triage submissions and selectively generate replies.

What it costs

Each smart reply costs 1 AI credit. On the free plan, you get 50 lifetime credits — enough to try the feature, not enough to run it at scale. On Pro ($19/month), you get 500 credits per month that replenish on your billing cycle. After that, replies are metered at $0.01 each.

For context, spam filtering and auto-translation each cost 0.5 credits. Smart replies cost more because they generate original text rather than classifying or translating existing text.

Getting started

If you already have a form set up — whether through an HTML form, a curl command, or an agent — adding smart replies takes about a minute:

1. Open your form settings
2. Write your knowledge base (or paste your existing FAQ)
3. Pick a mode: draft or auto
4. Save

The next submission that matches your knowledge base gets a reply. If you chose draft mode, you’ll see it waiting for your approval. If you chose auto, it’s already sent.

That’s the whole workflow. No training step. No waiting for a model to fine-tune. No prompt engineering. Write what you know in plain text, and let the pipeline do the rest.