On Monday morning, the salesperson has to choose which companies to contact, prepare an angle, check profiles, start the first actions and catch up with replies that arrived over the weekend. The problem is not only the number of messages. It is the state of the process: who was contacted, who replied, who is waiting, and who deserves a human follow-up.
In the setup I use for consulting firms, digital agencies and service companies, the agent does not replace the salesperson on LinkedIn. It prepares the campaign, organizes tasks, tracks replies and updates the CRM. LinkedIn invitations and messages remain manual actions unless a specific integration is explicitly authorized by the platform. The architecture uses Sales Navigator to target, Lemlist to sequence and a Hermes agent to operate the follow-up.
The channel is not the strategy
LinkedIn is useful in this setup because it lets a team filter profiles by function, industry, seniority or company size. It also makes it possible to work from a professional profile without starting by finding an email address. That does not make LinkedIn the best channel for every company, and volume does not fix poor targeting.
Cold email can be relevant when a company has reliable business contact data and a proposition that fits the recipient. A contact form captures inbound intent. Calling can work when there is context or a prior signal. The right choice depends on the market, the sales cycle and the quality of the data.
The first decision is therefore to define the target and the signal that justifies the contact. The tool comes next.
A three-piece architecture
| Piece | Role | What it actually does |
|---|---|---|
| Sales Navigator | Target | Company and person filters, saved lists, job-change alerts |
| Lemlist | Sequence | Email, enrichment, manual LinkedIn tasks, reply tracking and webhooks |
| Hermes agent | Operate | Prepares the campaign, reads conversations, creates next tasks, archives in the CRM and writes the morning report |
Hermes replaces neither Sales Navigator nor Lemlist. It sits above them. Its MCP connection, an interface that lets the agent use documented tools, lets it prepare actions and track their result. It does not automatically grant permission to act on a user's behalf on LinkedIn.
The weekly rhythm
On Monday, the salesperson defines the week's campaign: target, positioning, angle and stop rules. Hermes turns that brief into filters, a review list and a sequence. The salesperson then reviews the list in Sales Navigator and approves the companies that genuinely deserve a first contact.
The sequence can include:
- a task to send a LinkedIn invitation manually;
- a task to draft or review the message after acceptance;
- an email search and verification;
- an email follow-up when the context and legal basis allow it;
- a phone-number lookup, without assuming that the person can automatically be contacted on WhatsApp.
The agent can prepare these steps in Lemlist and remind the salesperson about deadlines. The salesperson keeps control of LinkedIn actions and reviews messages before sending. The list must also contain stop conditions: reply, unsubscribe, invalid address, request not to be contacted or lack of relevance.
The reply loop
When a reply arrives, a Lemlist webhook can feed a dedicated Slack channel, directly or through an n8n automation. Hermes then reads the full thread through the Lemlist API instead of acting on the latest event alone.
It can then:
- classify the reply and suggest a next step;
- create a follow-up task for the salesperson;
- prepare a proposed time slot or an email draft;
- stop the sequence when a prospect declines, asks not to be contacted or raises a question that requires human judgment.
A positive reply should not mechanically trigger another solicitation. The salesperson checks the context before proposing a meeting. This boundary between automated preparation and human commitment follows the same principle described in the article on human approval for production agents.
Traceability that helps the salesperson
Every touchpoint, reply and stop decision can be archived in the CRM. The morning report should not be another dashboard. It should answer a few questions: which conversations changed, which ones need a reply, which tasks are late, and which contacts must be excluded.
For an SMB owner, the benefit is knowing the state of prospecting without asking each person to reconstruct their history. The value comes from continuity and follow-up, not from sending one more automated message.
What the agent does not do
The list of what I refuse to automate matters as much as the architecture itself:
- Define the target. The owner or salesperson chooses the segments, companies and signal that justify contact.
- Automatically send on LinkedIn. Invitations and messages remain manual in the setup described here. LinkedIn's terms prohibit bots and unauthorized automated methods.
- Invent the core message. The angle and positioning are framed at the start of the week. The agent personalizes within that frame.
- Handle a sensitive conversation. As soon as a question becomes commercial, contractual or relational, the salesperson takes over.
- Decide to stop on its own. The agent can apply explicit rules and flag a case. The relationship owner defines the exceptions.
This discipline prevents the worst failure mode of prospecting automation: continuing to contact someone who declined, or replying badly to a genuinely interested prospect.
What it costs, and when it makes sense
Prices should not be presented as a stable package. On September 8, 2026, Lemlist's pricing page lists its Multichannel plan at $109 per month on monthly billing. Enrichment credits and some add-ons, including WhatsApp, then depend on usage. Sales Navigator varies by country, plan and billing method.
Hermes can run on existing infrastructure, but its actual cost depends on volume, models and connected tools. The useful comparison is not the subscription against zero. It is the system cost against the time spent researching, following up, updating the CRM and taking over conversations.
This architecture starts to make sense when:
- the target can be identified by function, industry or company size;
- several dozen new prospects need to be reviewed each week;
- a salesperson can handle interesting replies quickly;
- stop, objection and data-retention rules are defined.
These are practical markers, not a universal threshold. For fifteen contacts a month, a manually researched and written email may still be simpler.
Pitfalls to handle before deployment
LinkedIn limits are not a volume setting. LinkedIn prohibits third-party software or unauthorized automated methods from accessing the service, adding contacts or sending messages. A Lemlist sequence that offers LinkedIn actions is not, by itself, permission from LinkedIn.
Enrichment is not truth. An email address or phone number can be outdated, belong to the wrong person or be used in a context where outreach is not expected. Each data point should be checked and its source documented.
A notification is not a conversation. Hermes must read the full thread before suggesting an action. A reply such as "not interested" should stop the sequence, not trigger an automatic follow-up the next day.
Compliance belongs in the workflow. B2B electronic prospecting requires people to be informed and given a simple way to object. CRM retention must support deletion or anonymization when necessary. Moving to WhatsApp or phone requires a separate check of the channel and context.
What it changes for an SMB
A sales team of one to three people does not need a robot talking to everyone. It needs a system that prepares the work, keeps the history and surfaces the conversations that deserve human time.
The useful metric is therefore not the number of messages sent. It is the number of relevant conversations relative to the time spent preparing, following up and taking over. The machine organizes the process. The salesperson remains responsible for the relationship.
Sources consulted
- LinkedIn User Agreement, rules on bots and unauthorized automated methods, checked September 8, 2026.
- LinkedIn Help: Automated activity on LinkedIn, checked September 8, 2026.
- Lemlist: pricing and features, pricing, credits and add-ons checked September 8, 2026.
- CNIL: commercial prospecting, information and objection rights, checked September 8, 2026.
If you want to examine this kind of process in your company, start with the AI agents and automation page and a limited first workflow with explicit stop rules.
Also available: Read in French