AI Agent for Outbound Prospecting with Human Approval: Nothing Goes Out Without Your OK

Outbound prospecting rarely fails for lack of will. It fails because every message takes research: understanding the company, finding the right person, finding a credible reason to write to them now. When there is no time for that, you send a generic message that goes unanswered, or you send nothing. The other route, automating everything, has its own risks: an agent that writes alone to strangers on behalf of your company puts your reputation on the line with every message. An autonomous AI agent can do the research and writing, provided a human keeps control of the sending.

The inspiration: prepare in bulk, send under control

The starting point is a use described in accounts from people who run autonomous agents: the agent studies a list of accounts, fills in the CRM, writes the messages, and asks its human to approve before any sequence starts. What matters in these stories is the human's place: they do not write the messages, they approve them, which takes a few minutes instead of several hours, without removing responsibility for what goes out.

Atako takes this principle literally and adds a simple rule: a single grouped validation question that lists all the messages, rather than a series of requests that interrupt. You see the whole batch and decide in one go.

What the agent does

The scenario deploys a single agent, the outbound researcher. For each target account, it follows the same method.

It first researches the company on the web with its browser: the official site, the news of the last 90 days counted from today's date, the job posts and the public product pages. It then judges its fit with the ideal customer you described, in one line justified by signals. If the company does not fit, it says so and prepares no message.

For the accounts that fit, it names the role of the right person to contact, and looks for a verifiable public contact. If it cannot verify a name, it keeps the role alone. It writes a short personalized message: one concrete reason to contact this company, one benefit, one simple question. Finally, it searches in HubSpot, then creates or updates the company and the contact, with a note holding the research, the sources and the proposed message. It never creates a duplicate.

When all accounts are processed, it asks you a single validation question, which appears in your Atako notices and lists each message, account by account. It waits for your answer.

What the agent does not do

The agent never sends a message itself. It adds no lead to a campaign before your approval. It invents no figure, no name, no email address and no quote, and it says what remains uncertain. As soon as a prospect replies, it stops all contact with them and tells you.

After your approval, it acts only on the approved messages. If Lemlist is connected, it adds each lead to the campaign you chose. Otherwise, it creates a "Send" task in HubSpot for the record owner, with the message as its body. A refusal from you is enough to stop everything: nothing is sent. This is the principle behind human-in-the-loop design, applied at the moment it matters most, before contact with a third party.

How to set it up in Atako

From this use case's page, deployment takes four steps.

1. Connect HubSpot, and Lemlist if you use it. Create a private app in HubSpot with read and write access to companies, contacts, notes and tasks, then paste the token when Atako asks for it. Lemlist connects with its API key and stays optional.
2. Fill in your context. Describe your offer and your ideal customer, list your target accounts (ten at most, one per line, with the site if you have it) and name the Lemlist campaign if you have one.
3. Deploy. A single action creates the agent and the "Prospecting" project. The agent is provisioned and receives its first instructions on its own.
4. Approval, then a weekly rhythm. The companies and contacts appear in HubSpot, and the validation question reaches your notices. Every Monday at 9:00, the agent rereads the replies and stops the sequences concerned.

If you do not have an account yet, creating one starts your trial and the scenario is configured right after. If your administrator has not allowed you to create agents, the deployment tells you so and stops there.

Where to see the results

You follow the work in three places. In HubSpot, each target account is a company with its contact, and the agent's note holds the research, the sources and the proposed message. In your Atako notices, the validation question lists all the messages: this is where you approve, edit or refuse. On the Atako floor, the agent appears with its current activity, which tells you whether it is still working on the accounts or waiting for your answer.

You can also talk to the agent in the chat: ask it to rephrase a message, add an account, change the tone or revisit the fit criterion. Requests are handled in the same conversation, with the memory of the accounts already processed.

Safeguards

Access is deliberately narrow. The HubSpot connection is granted to the agent with a limited list of nine tools: searching, creating and updating companies and contacts, creating notes and tasks, and associating records. If Lemlist is connected, five tools are granted: listing and reading campaigns, adding a lead to a campaign, finding a lead by its address and listing activities. An agent that sees a short list of tools uses them more reliably than an agent lost in a catalog of hundreds. Nothing is granted by default, and you can withdraw access at any time.

Human validation is not a comfort detail: it is the condition of the use case. You keep the decision on every message, hence on every time the company speaks. Remember to respect the rules that apply to your prospecting, in particular consent and the right to object depending on the recipient's country: the agent prepares, but responsibility for sending stays with you.

What to expect

Expect short, sober messages, more relevant than a generic template, but still improvable on tone. The first batch mostly serves to calibrate: after a few corrections in the chat, the next messages get closer to your way of writing. The gain is not sending more, it is sending better-grounded messages while spending far less time on them. For a small sales team, it is the difference between prospecting that happens in bursts, when someone has an hour, and prospecting prepared continuously, where all that is left for you is to approve the messages.