Why AI agents change B2B growth
What moves when an agent does the repetitive work of B2B growth, what still needs you, and where to start in your own team.
AI agents change B2B growth in one specific place: the repetitive work that sits between your tools. Researching an account, enriching a list, drafting a first message, pulling the weekly numbers, checking that a page still ranks. That work used to need a junior, an agency or your evenings. An agent can now do most of it, at any hour, provided someone has written down how it should be done and someone checks what comes out. The choice of customer, the strategy and the decision to send anything in your name stay with you.
I ran growth at Deux.io for eleven years, with more than 500 clients, and I now build agents with Claude for B2B companies. What I see every time is that an agent moves the work more than it removes it. You spend less time producing and more time specifying and checking.
What an agent does that a script did not
A script follows a path you drew in advance. An agent receives a goal, a set of tools and written rules, and decides the next step itself: read this page, call that API, notice a field is empty, look somewhere else, stop when the output matches the spec.
That matters for growth because most growth work is made of small judgements on messy inputs: whether a company fits your ICP, which of three people actually runs sales, whether a homepage says the company sells to HR teams, which recent event is worth mentioning in a first message. Before, you either hired someone to make those calls one by one or you skipped them. Now a model can make them at volume, with evidence attached, and you check a sample.
The next chapter draws the line between agents, automation and chatbots. Keep one test in mind until then: if every case fits a flowchart drawn in advance, you need automation; if each case needs a small decision, an agent may be worth it.
Where the time goes in a growth team
Here is how I split the usual tasks. The rows are typical of the B2B teams I have worked with, not measurements.
| Task | Repetitive part | Judgement part | Agent candidate |
|---|---|---|---|
| Building a prospect list | Imports, enrichment, deduplication | Defining the ICP, accepting the list | Yes, with a hand-checked sample |
| Outbound messages | Finding a usable signal per lead, first draft | Offer, tone, pressing send | Yes, a person sends |
| SEO content | Briefs, drafts, internal links, checks | Which topics, what gets published | Yes, behind a quality gate |
| Weekly reporting | Exports, joins, tables | What the numbers mean | Partly, with a second instrument |
| LinkedIn posts | Drafts in a calibrated voice | Opinions and stories | Drafts only |
| Pricing and positioning | Very little | Almost all of it | No |
The last column is never a plain yes. Every candidate comes with the condition that makes it safe, and that condition is where your time goes once the agent runs.
The bottleneck moves to the spec and the check
I co-founded Derrick, a B2B data enrichment and prospecting tool, and its marketing runs on Claude agents across seven areas: SEO, LinkedIn, video, analytics, newsletter, outbound, and a “feature factory” that produces new data functions for the product. Each area has a pilot, an agent or a skill, that checks the output before anything goes out. Building that taught me where the effort really sits. A draft costs almost nothing. Saying what a good one looks like, and catching the bad ones before a reader does, is the job. Chapter 17 describes that system in full.
A real example: in November 2026 a LinkedIn post said “filter, then enrich”, and only a check against the real product showed that the gesture also enriched, and billed, hidden rows (chapter 10 tells the story).
So the work that remains looks like this:
- a written spec for each agent: role, rules, examples, and what done looks like;
- a check that does not rely on the agent’s opinion of itself (at Derrick, every outgoing text goes through a mechanical lint and then a judge run in a blank context);
- a measure of what the output brought in, separate from how much was produced.
Cheaper attempts change how you decide
When a try costs less, you can run more of them, provided each one is bounded. At Derrick, an experimenter agent runs SEO tests on a limited set of pages, ten for instance, and every test gets a dated verdict after four weeks. The verdict is written down whether the test won or lost.
Stopping is part of it: at Derrick, a Reddit lever was started and then cancelled in October 2026.
The other side of cheap volume: an agent working confidently from a broken input produces a lot of wrong output quickly. In November 2026, our Head of SEO agent found that one of its own measurement exports counted only about a third of the real clicks, and a production stop had been decided on a false zero (chapter 12 tells how it was caught).
What still needs you
Some decisions do not get better with more volume, and those are yours:
- knowing who your customer is, from your own deals, before any agent searches for more of them;
- choosing what to try and what to stop;
- accepting an output, by reading a sample against criteria written beforehand;
- anything irreversible or sent in your name: emails, posts, pushes into a live campaign;
- noticing that a number looks wrong, and asking where it came from.
An agent can help with each of these. It can propose the experiments, prepare the sample, flag an anomaly. It should not own them.
Audit your own week before you build anything
Before choosing a tool, list what you and your team actually did last week. A table like this one is enough:
| Task | Hours per week | Tools touched | Done looks like | How I would check it | Reversible |
|---|---|---|---|---|---|
| Enrich inbound sign-ups for sales | 3 | CRM, spreadsheet, LinkedIn | Company, size, role filled for every sign-up | 20 rows read by hand | Yes |
| Weekly SEO report | 2 | Search Console, analytics, slides | One table, five comments | Totals match the source | Yes |
| Follow-up emails after demos | 2 | Calendar, CRM, mailbox | Draft ready in CRM within a day | Sales reads before sending | No once sent |
The rows above are an example. Fill in your own, then keep the tasks that repeat every week, touch at least two tools, have an output you can check against a rule or a sample, and can be undone. The first agent worth building is usually the dullest line on that list.
Your next step: fill the table with ten real tasks, circle one, and write down how it is done today, step by step, as described in start from a process that already exists.