Hiring someone to build AI email sorting and auto-replies: what to specify

The short answer

AI email triage is two jobs that people tend to order as one. Sorting reads each incoming message and decides what it is: a lead, a support question, an invoice, spam, something urgent. Replying writes and possibly sends an answer. Sorting wrong costs you a delay. Replying wrong puts a sentence in front of a customer, under your name.

So brief them separately. Decide which mail it reads, the handful of categories you want, and exactly which replies it may send without a person. Start with drafts that someone approves. And ask how you will see the messages it skipped, because the failure you will not notice is the email it never looked at.

Sorting and replying are different risks

Most of the value is in sorting. A message that lands in the right place, with a one-line summary and a priority, already saves the time of opening and reading it. If the label is wrong, someone moves it. Nobody outside your business sees the mistake.

Replying is where the risk is. An AI reply can promise a refund, quote a price it cannot see, or answer a question from the wrong customer’s thread. Once sent, it cannot be taken back.

That is why a good brief gives them different permissions:

	Sorting	Drafting replies	Sending replies alone
Who sees a mistake	You	You, before sending	The customer
How to undo it	Move the message	Edit or delete the draft	You can’t
Where to start	Day one	Day one	Only after a trial period, for a short list of message types

If the replies need to look things up, such as an order status or a booking, the build is closer to an AI agent than a simple workflow. The difference is explained in AI agent or workflow automation, and it changes both the price and the testing.

What I learned from my own auto-reply bot

I run a small bot that answers comments on social posts with a private message. It is not email, but it is the same shape: read incoming messages, decide which ones need an answer, send one. Three things went wrong with it, and each became a line I would now put in an email brief.

It only looked where it was told to look. The bot checked my twelve most recent posts. Because posts went out in pairs, that covered about a day and a half. Comments on an older post were never even attempted; I found two of them and answered by hand. The fix was to scan by time window instead of by count. For email, the same question is: which folders, labels, aliases and shared inboxes does it read, and how far back?

Some messages arrived by another route. When I paid to promote a post, the platform created a separate copy of it, and comments on that copy did not show up where the bot looked. Four comments went unanswered for three days. On the app they looked like one post, so it seemed the bot was ignoring people. In email, the equivalent is mail that reaches you through a forward, a second address or a contact form, which the automation never sees.

I checked it against itself. My first check said nothing was missing, because I compared the bot’s list of comments with a count from the same source, which had the same blind spot. The real comparison had to come from outside the bot: the app screen. For email, test against your actual inbox, not against the automation’s own log.

A fourth, smaller lesson: one reply failed for a single person six times in a row, every twenty minutes, because their account did not accept the message. Now a reply that fails twice is marked undeliverable and handled another way. Ask for a retry limit.

The brief: seven lines to hand over

Line 2 matters more than it looks. The chatbot outcome guide makes the same point for chatbots: describe what should be true afterwards, with examples, not a list of features. “Leads get a reply within an hour and land in the sales folder” is testable. “Uses AI to classify emails” is not.

Lines 1 to 5 fit into the goal, knowledge, never-do and handoff fields of the chatbot and AI agent brief. Line 7 is the one most briefs miss.

Drafts first, sending later

The safest way to start is a draft mode: the AI sorts everything and writes a reply as a draft, and a person reads and sends it. That gives you a record of what it would have sent.

That is how I run it for my own inbox. I have an AI job that sorts and summarises my incoming email and reports to me. It prepares reply drafts at most; it never sends a reply on its own.

For email, agree when a message type moves from “draft” to “send alone”. For example: after a set number of drafts of that type have been sent unchanged. Write the rule down. Then ask that sending can be switched off in one step, by you, without the freelancer.

What moves the price

I don’t have a reliable price list for email triage builds. Starting prices for the two kinds of freelancer who build it are on the workflow automation page and the AI agents and chatbots page. What moves the quote:

• Number of mailboxes and routes. Each one needs connecting and testing.
• Number of categories. Each needs examples and a test.
• Lookups. Replies that check an order system, a calendar or a CRM need those connections built.
• Sending alone. It needs stricter rules, a kill switch and more testing than drafts.
• Volume. AI usage is billed per message processed. Ask for an estimate per 1,000 emails, separately from the build price. Usage limits on the mailbox and the AI account matter at volume too.

If you already automate with n8n, Abdullah Ramzan lists AI agents inside n8n workflows that also take actions, which is the shape of a triage build that looks things up.

How to test it before you accept it

• Send it real mail. Twenty emails you received last month, forwarded or recreated, across every category. Count wrong labels as well as right ones.
• Use every route. One message to each address, alias and form. Each should appear in the result.
• Try the never-send list. Ask for a refund, a price, another customer’s details. It should draft or hand over, never send.
• Read the skipped list on day two and check it against your inbox yourself.
• Switch sending off using the method you were given, without asking the freelancer.

The method in how to test an AI chatbot before you accept it applies here: decide the pass mark before you start, and count wrong answers, not only good ones.

Wrap-up

Order sorting first and treat replying as something it earns. Name every place your mail arrives, list the few replies it may send alone, keep the accounts and the spending limit in your name, and ask for a daily list of what it skipped. The email it never read is the one that will cost you, and only a check from outside the automation will show it to you.