AI replies on LinkedIn: a human-in-the-loop setup

A human-in-the-loop setup means the AI drafts, and a person decides how much it may send alone. In Mako Reach you choose draft-only or auto-send per chat, set rules that hand a chat to a person, and let the agent ask you on Telegram or Slack when it lacks a fact. It runs on your own AI key.

Everything below describes what the product does today, so you can set it up in an afternoon and see where a person stays in the loop.

What should an AI reply agent do, and what should it not do?

It should answer the messages that leads send you, in your tone, and stop when it is out of its depth. It should not start conversations, guess facts it does not have, or keep replying to a chat that has ended.

The Mako Reach agent answers inbound messages only. It does not start conversations: your campaigns do that. Before it replies in AUTO, it checks whether the lead's last message needs an answer at all. A bare "you too!" after a goodbye does not, while a question after a goodbye does. If that check fails for any reason, the agent assumes a reply is needed, so a real question is never silently dropped.

It also has limits that stop loops. After a set number of AI replies in a row with no message from your team (4 by default), the agent hands the chat to a person. That stops endless back-and-forth with another bot.

Draft or auto-send: which one should you choose?

Start with draft. The two modes differ in who presses send.

Every chat is MANUAL by default. AUTO needs two switches: a workspace admin turns on Enable auto-send for the whole workspace, then you set a chat to AUTO or use the workspace-wide switch. Until an admin enables auto-send, nothing goes out without a person.

In AUTO, replies are not sent at once. They wait a randomized delay, 2 to 15 minutes by default, which you can change. During that time they sit under Pending AI replies, where you can cancel, edit or send now. You can also set working hours and a time zone, so a reply that falls outside them is deferred until the window reopens, and a daily cap on auto-replies per account (40 by default).

When should the agent hand a chat to a person?

Whenever you would not want a machine to answer. In Settings, under AI and AI agent, the Behaviour tab has Handoff rules. A rule you have enabled can stop AUTO on a chat and pass it to a person:

• Confidence. Before an automatic send, the agent checks how sure it is. If the confidence is at or below your threshold, it does not send. This rule is on by default for a new setup.
• Sentiment. If the lead's tone is at or below Negative, or Neutral if you choose that, the chat is handed over. Off by default.
• Keywords. A lead message containing any word you list, such as price, contract or lawyer, hands the chat over.
• Failed regenerations. If you regenerate a draft several times in a row without sending, the chat is handed over.

When a handoff fires, the chat switches from AUTO to MANUAL, the pending replies for it are cancelled, and a note in the thread says why. You also get an alert on every connected Telegram chat or Slack channel, so you know a person is needed.

How does the agent ask you a question?

When the agent lacks a fact it needs, such as a price or a date, it does not guess. It asks a person.

Under the Escalation tab you connect Telegram (a bot token from BotFather and a chat id) or Slack (with Add to Slack, which picks a channel). The agent posts the question there and pauses that reply. When someone answers, the bot offers buttons:

• Send (one-time) uses the answer in this reply only;
• Send and add to the knowledge base uses it and keeps it for future replies;
• Give another answer lets you rewrite it.

An answer is never applied on arrival: a person taps a button first. If the reply already went out, only the button that adds it to the knowledge base remains. Kept answers appear as Lessons in the AI agent settings, so the agent asks less often as it learns.

How do you set tone and context rules?

On the Knowledge tab, three fields shape every draft. Instructions (required) is the system prompt: who you are, how you write, what to avoid. Company info and Sales script are optional context that the agent can draw on.

Keep instructions short and concrete. Say what the agent may promise, what it must not (prices, dates, legal terms) and when to say that a colleague will follow up. Facts you do not give it are facts it should ask for.

On the Behaviour tab you choose the reply language. By default the agent answers in the language the lead wrote in. You can fix one language instead, and you can pick a separate language into which incoming messages and draft previews are translated so you can read them.

How do sentiment tags help you triage?

The AI marks each conversation as positive, neutral or negative with a fixed system tag, and the same tags show in Chats and in Leads. You can filter your inbox by them to read the warm replies first.

A person can overrule the AI. If you set a sentiment tag by hand, it is locked against the AI until you return it to the AI. If you remove it, the AI may rate the contact again after their next reply.

Stop words work before the agent does. A reply containing a stop word such as "stop" or "unsubscribe" adds the person to the suppression list before the AI agent runs, and no further step is sent. The built-in stop-word list is in English.

Whose AI key does it use?

Yours. The AI connection belongs to your workspace and only a workspace admin can set it, under Settings, AI, AI model. Mako Reach supports three providers: Anthropic (Claude), OpenAI and OpenRouter, which gives one key access to many vendors' models. The key is stored encrypted and shown masked afterwards. Without a key the agent does not run, and you pay your AI provider directly for what it uses.

Your drafts and the lead's messages go to the provider you connect. Read that provider's terms before you switch AUTO on for real leads.

How do you set it up in Mako Reach?

1. Connect your AI provider. In Settings, AI, AI model, choose a provider and add your key (admin only).
2. Write the knowledge. In AI agent, Knowledge, fill in Instructions, and add Company info and a Sales script if you have them.
3. Set behaviour. In Behaviour, choose the reply language, working hours, delay range, daily cap, and the AI replies in a row limit.
4. Enable handoff rules. Keep the confidence rule on, add keywords such as price or contract, and consider the sentiment rule.
5. Connect Telegram or Slack. In Escalation, turn on asking your team and connect a channel, so the agent can ask you and alert you.
6. Start in MANUAL. Open a chat and click Generate draft. Read ten drafts and fix the instructions where they miss.
7. Switch on auto-send for a few chats. Turn on Enable auto-send, set a handful of chats to AUTO and watch Pending AI replies for a week.

See the full list of settings on the AI reply agent page, see where replies land in the unified inbox, and set sender caps with the account safety settings. Compare notes with our buyer's checklist.

FAQ

Does the agent send messages without my approval?

Only in AUTO, which needs an admin to enable auto-send and a chat or the workspace to be set to AUTO. Replies first appear as pending replies that you can cancel, edit or send now. In MANUAL it only drafts.

What happens when the agent does not know the answer?

It asks you on Telegram or Slack and pauses the reply. Your answer is used in the reply, and you can keep it as knowledge for next time. A person taps a button before the answer is applied.

Can the agent start a conversation?

No. It answers messages that leads send you. Your campaigns send the first messages.

Which AI providers work?

Anthropic, OpenAI and OpenRouter, with your own key. A workspace admin adds it in Settings, AI model.

Does the AI reply in the lead's language?

By default, yes. A workspace can fix one reply language instead.