For most of my career, the answer to "we need more pipeline" was always the same: more budget, more tools, more people. When I took over Growth & GTM in a Founder's Office at an enterprise software company, none of those were on the table. We had a small team, a demanding market of automotive and manufacturing OEMs, and a founder who measured everything in revenue.
So we tried something different. Instead of buying another platform, we started building our own AI agents for the repetitive parts of marketing and sales. Two years later, a team of three runs the kind of programme I used to see at companies with ten marketers. This note is about how, and about the parts that didn't work.
What "AI-native" actually means
It does not mean "we use ChatGPT to write posts." Almost everyone does that now, and it hasn't changed anyone's results much. An AI-native team is one where work is designed around agents from the start: humans decide, agents prepare and execute, and every loop feeds data back into the next decision.
In our case that meant mapping the week of a marketer and asking one question about every task: does this need judgement, or does it need doing? The split was uncomfortable. A large share of what the team did each week was "doing": pulling reports, researching accounts, drafting first versions, chasing follow-ups, reformatting the same content for four channels.
The stack: small, boring and ours
We kept the tooling deliberately simple:
- n8n as the orchestration layer. It triggers, routes and connects everything.
- Claude as the reasoning engine inside those workflows, for research, drafting, classification and summarising.
- Claude Code to build the small internal apps we needed, including our own CRM.
- Render and Streamlit to host the agents and give the team simple screens to work from.
None of these are exotic. The advantage wasn't the tools; it was that we owned the workflows end to end and could change them in an afternoon.
The agents that earned their keep
1. The performance marketing assistant
Every morning it reviews campaign performance, flags anything that moved sharply, and drafts a short note on what probably caused it. A human still decides what to change. But nobody spends the first hour of the day building the same report.
2. The social and content engine
It turns one idea, such as a customer conversation, a product update or a point of view, into drafts for LinkedIn and other channels in our voice. We publish four to five posts a week now, and reach roughly doubled. The trick was giving it real raw material. Agents that write from nothing produce nothing worth reading.
3. The research and enrichment agent
Before a demo or an outreach sequence, it pulls together what we know about the account and the people involved, so the conversation starts at the second level, not the first.
4. The cross-sell finder
This one surprised us. It reads our existing customer base and flags which products each account isn't using yet. Its first run surfaced more than thirty expansion opportunities that had been sitting in our own data.
The biggest growth lever we found wasn't a new channel. It was the customer list we already had.
Together these save the team around twelve hours of manual work every week. That time didn't disappear; it moved into strategy, creative work and conversations with customers.
Three rules that kept it sane
Rule 1: Humans own the decision, agents own the draft
No agent publishes, sends or spends without a person approving it. This sounds slow. In practice approval takes seconds, and it is the reason the team trusts the system.
Rule 2: Every agent has one job and one owner
The early mistake was building one "marketing brain" that did everything. It did everything badly. Small agents with a single purpose are easier to test, fix and trust.
Rule 3: If it doesn't feed the tracker, it doesn't count
Every campaign and experiment is logged with its hypothesis and result. Without that loop, AI just helps you make the same mistakes faster.
What didn't work
Fully automated outreach. We tried it, and the replies told us exactly what buyers think of messages nobody really wrote. We also over-built: a few agents solved problems we didn't actually have. We switched them off without regret.
The other lesson was cultural. People worry, reasonably, that agents are a step towards replacing them. What worked for us was being explicit: the agents exist to remove the work nobody wants, so the team can do the work that gets them promoted.
Where to start if you're building one
- List every recurring task your team does in a week and mark each one "judgement" or "doing".
- Pick the most painful "doing" task and build one small agent for it. Just one.
- Measure the hours it saves and the quality of what it produces for a month.
- Only then build the next one.
AI won't make a weak strategy strong. But for a small team with a clear strategy, it is the closest thing to hiring that doesn't require a budget meeting.