How AI Agents Automate Your Sales Pipeline

The promise sounds nice: an agent that researches your leads at 8 AM, drafts the cold email, books the call. In practice the difference between a slide deck and a working agent is about three weeks of careful integration work and a list of edge cases nobody talks about. This post is about what an inbound sales agent actually does, what we have shipped in production, and the things that go wrong.

What an "agent" means here

An AI agent is not a chatbot. A chatbot answers questions. An agent takes actions inside your systems.

A working sales agent does something like this:

1. Wakes up on a cron schedule (we usually pick 7 to 8 AM local time so the salesperson sees results by their first coffee).
2. Pulls overnight leads from your CRM (HubSpot, Pipedrive, Salesforce, whatever you use).
3. Enriches each lead from a few sources: company website, LinkedIn, sometimes Crunchbase or a competitor signal feed.
4. Scores the lead against your ICP. Not "good or bad", but a specific score with a stated reason.
5. For the top picks, drafts a personalized cold email that references something real from step 3.
6. Books a Cal.com slot or schedules a follow-up in your CRM when the lead replies.
7. Logs everything to an audit trail you can read with coffee in your hand on Monday.

That is the loop. It is unglamorous. It is also reliable, which is the whole point.

What it actually replaces

Most sales teams do this manually. Manually means a junior SDR opens HubSpot, copies a lead into LinkedIn, copies notes back, writes an email, sends it, and then forgets to follow up because something more urgent came in at noon. Total cost: 6 to 8 minutes per lead, plus 30 percent of leads that simply never get touched because the day ran out.

An agent does the same loop in under 90 seconds per lead and does not forget. Not because it is smart in some grand sense, but because it is a script that does the boring part the same way every time.

A recent build: what we shipped

Our most recent inbound-sales agent was built for a B2B services company on HubSpot. Roughly 40 inbound leads per week from their landing page and webinar funnel.

Stack we used:

• Anthropic Claude for the email drafting (Claude tends to write less obviously AI-flavored copy than GPT-4o for cold outreach).
• GPT-4o-mini for the lead scoring (~16x cheaper than GPT-4o, ~90% agreement on classification tasks).
• LangGraph for orchestration (state machine, retries, branching when the lead is missing data).
• FastAPI as the layer that receives webhooks from the client's CRM.
• Postgres for the audit log because the client wanted to see every decision the agent made.

Honest numbers from the first 60 days in production:

• Response time on inbound leads dropped from about 4 hours average to under 2 minutes.
• The agent drafts about 35 personalized emails per week. The salesperson approves them in under 10 seconds each.
• Roughly 60 percent of the agent-drafted emails go out untouched. The other 40 percent get a small edit.
• Booked meetings went up roughly 2.5x compared to the manual baseline. We do not claim 3x because two of those months were seasonal.

This is real and boring. There is no magic.

The honest list of what it does not do

I will save you a discovery call by saying this upfront:

• It will not write outbound cold email sequences that magically convert at 30 percent. Cold outbound is a different beast and most "AI cold email" tools produce slop.
• It will not replace your salesperson. It replaces the data work around your salesperson so they can spend their day on the calls that matter.
• It will not work without a clean CRM. If your HubSpot has 14,000 leads with no source tagging and three duplicate fields per contact, week one of the project is cleaning your CRM, not writing AI prompts.
• It will not handle complex objection handling on the call itself. Voice agents are a separate (and currently rougher) topic.

When this makes sense, when it does not

An inbound sales agent pays for itself when you have:

• More than 15 to 20 inbound leads per day. Below that, your salesperson can handle them faster than the cost of building the agent.
• A CRM with real data and a real sales process you can describe in writing. If your sales process is "Tomek decides", we cannot codify Tomek.
• Genuine competition on response time. If you sell something nobody else sells, slow response is not your bottleneck.
• A salesperson or founder who will spend 30 minutes a week reviewing the audit log and tuning the prompts.

If you have under 5 leads per day or your sales process lives in someone's head, skip this and hire a part-time SDR instead.

What this costs

A production sales agent in our shop runs 7,000 to 15,000 EUR fixed price. Timeline is 3 to 5 weeks. That includes:

• CRM integration (one CRM, additional ones extra).
• Enrichment integration (usually one paid data source plus scraping).
• The agent itself with audit log and emergency stop.
• A simple internal dashboard so the salesperson can see queue and approve emails.
• 30 days of post-launch support.

Running costs after launch are usually 80 to 200 EUR per month in LLM API spend for a team doing 40 to 100 leads per week. Enrichment data sources are on top of that.

When this is worth doing

If you have an inbound pipeline that is slowing because nobody can keep up with first-touch follow-up, an agent is the right answer. If your funnel is small or messy, fix the funnel first.

A 15 or 30-minute discovery call gets you a rough scope, a rough price, and an honest read on whether this is worth doing yet. No pitch deck.