AI agents or a part-time head of growth
What to hand to agents, what to keep for a person, and how to split the two.
Agents take over execution that can be checked against a rule: drafting, enriching, classifying, reporting, publishing a page that passed its checks. A head of growth decides what is worth doing at all, owns the numbers, and says stop when a channel is not working. For most B2B companies that have outgrown founder-led marketing but cannot justify a full-time hire, the useful question is in what order to combine the two. My answer is the person first, for a short time, to set direction and write the rules, then agents on the repetitive work, with the person reading the results every week.
I spent 11 years as co-founder and CMO of the growth agency Deux.io, and I now build agent systems, including the one that runs Derrick’s marketing (I co-founded Derrick). Both sides of this chapter come from that.
What agents already do well in growth
The test is simple: can someone check the output without redoing the work? If yes, an agent can probably do it. If checking it means redoing it, you are paying twice.
| Task | Can an agent do it | How you check it |
|---|---|---|
| Enrich and clean a list of accounts | Yes, with a data tool | Rows in against rows out, a sample read by hand |
| Draft posts, emails and pages from your material | Yes | Lint, an independent judge, your final read |
| Produce pages from data | Yes | Checks before shipping, tests after |
| Classify replies to a campaign | Yes | The quote is kept next to each label |
| Build the weekly report | Yes | Totals match the source tools |
| Choose which channel to bet on | No | There is no rule to check it against |
| Decide that a number is wrong | Partly | It can flag the gap; someone has to decide |
| Talk to customers about why they bought | No | Only by doing it |
The detailed use cases are in this guide: a B2B prospecting agent, a programmatic SEO agent, a LinkedIn agent, and a full marketing team.
What stays with a person
Four things, in my experience.
Choosing and stopping levers. In Derrick’s system, a Reddit lever was started and then cancelled in October 2026. A routine does not stop itself: ending a lever is a judgement call, and it belongs to a person.
Trusting the numbers. An agent can catch that two exports disagree, as Derrick’s Head of SEO agent did in November 2026 (the full story). Deciding which past decisions to revisit, and owning that, belongs to someone accountable. The same reflex is at the heart of making tracking reliable in a growth audit.
Customer conversations. Why people buy, what they compare you with, which words they use. That material feeds every agent’s instructions, and it only comes from talking to them.
Arbitration. Sales wants more leads, product wants fewer bad sign-ups, the budget is fixed. Someone accountable has to choose, and explain the choice.
Three setups, and when each one fits
| Setup | Fits when | Main risk |
|---|---|---|
| Agents run by the founder or a marketer, no growth lead | The strategy is clear and stable, and someone has time to write and maintain rules | Nobody questions the direction; the system optimises the wrong thing efficiently |
| Part-time head of growth, no agents | Direction is unclear, or the data is not reliable yet | Recommendations wait for people to execute them |
| Part-time head of growth who builds and runs the agents with your team | Direction needs work and execution capacity is short | The system depends on one person if it is not documented in your tools |
The third setup is the one I work in, so weigh my view accordingly. Its risk is real, and I ran into it myself: until 9 November 2026 the whole Derrick agent organisation existed in one copy, on one laptop (how it is protected now). If you hire someone to build agents for you, the instructions, routines and logs must live in your repository and your accounts from day one.
How to split a week
An illustrative split for a B2B company with a part-time head of growth two days a week and agents running routines. The days are an example; the principle is that agents propose and the person decides at fixed points.
| When | Agents | Head of growth |
|---|---|---|
| Monday | Produce the weekly scorecard: output (what was produced) next to outcome (what it brought in) | Reads it, decides what to continue, change or stop |
| Monday to Thursday | Run routines: drafts, enrichment, pages, reply classification. Each routine proposes, nothing goes out without approval | Approves or rejects in batches, answers what the agents flagged |
| Thursday | Re-run frozen test cases if any instructions changed | Writes new rules from the week’s corrections |
| Ongoing | Log what was built and why | Talks to customers and sales, updates the facts file |
What the head of growth hands back to the agents each week is written: a rule added, a lever paused, a fact corrected. If their week produces only meetings and opinions, the agents keep doing last month’s work. Measuring what an agent brings in covers the scorecard.
Questions to ask before you hire either
Whether you buy an agent setup, hire a head of growth, or both in one person, ask for evidence on these points:
| Question | A good answer looks like |
|---|---|
| What did you stop recently, and why? | A named channel or test, a date, the number that decided it |
| How do you know a number is right? | A second, independent source, and an example where they disagreed |
| Who checks the agent’s output? | A mechanical check plus a reviewer that is not the agent that wrote it |
| What happens if you leave? | Everything lives in your repository and your accounts, documented |
| What does an agent never do alone? | Publish, send, spend above a ceiling, touch customer data without a rule |
| Can I see a log of what you built and why? | Dated entries with the lesson learned |
If I can help you set this up, that is the work I do as a freelancer.
Your next step: list the growth tasks your team did last month, mark each one “checkable by a rule” or “needs judgement”, and give the first column to agents only once someone owns the second.