You have six months

This week I built a complete marketing program in about twenty minutes. A webinar, its landing page, a guide, a blog post — and, at the technical core of it, the top five AI optimisations the whole thing is about.

I am not qualified to write those five optimisations. That is the part worth stopping on.

Here they are. Read them and decide for yourself whether a CMO produced them:

1. Complete visibility and forecasting of AI workloads. Allocate every token, GPU-hour and managed-AI dollar to a team, a product and a feature. Forecast per workload. Alert on anomalies before the invoice does. This one is first because it is the multiplier — without allocation, every fix below is a one-off that regresses, because nobody owns the number.
2. Model selection policies. Enforce which model tier each use case may call. Route classification, extraction and eval traffic to smaller models, enforced at the gateway, with evals so exceptions have to earn their cost. Frontier models are the default because nobody set a different default.
3. Token efficiency. Prompt caching on repeated system prompts, context pruning, and batch endpoints for work that was never interactive in the first place.
4. Bedrock commitment and capacity optimisation. Right-size provisioned throughput against p95 utilisation, move batch-tolerant jobs to batch inference, and use cross-region inference profiles to absorb burst instead of over-provisioning one region.
5. GPU and inference infrastructure right-sizing. Reclaim idle GPU capacity, autoscale endpoints to observed demand, checkpoint training onto spot. Classic infrastructure hygiene — it just has not reached the AI stack yet.

The rest of the campaign I could have done slowly and badly on my own: the topic, the story arc, the landing page, the copy, the branding. That list above is a different category. “Right-size provisioned throughput against p95 utilisation” is not a sentence a marketer writes from general knowledge. It needs somebody who genuinely knows how AI workloads get optimised — the kind of subject-matter expert whose calendar sets your launch date, because you book a week of calls and then wait on a review cycle. I do not have that knowledge. I have never had it. I now have work that carries it anyway, grounded in our own systems rather than in something a model remembered.

One thing I have left out, deliberately: each of those five carries a savings range in the campaign, and I am not publishing them here. They are directional industry ranges pending measurement against live data, and the campaign's own checklist says replace them before anything ships. Publishing them as fact in a post arguing for counted numbers would be the exact failure I describe further down.

And the campaign is not just those five. It is a recurring twenty-minute series with five episodes mapped, episode one's run of show timed minute by minute, landing page copy and the built page with the form slot marked, a four-email sequence, a blog post, and the slide designs I am building as I write this. Conventionally that is two to three weeks and five people minimum — ideate, write, design, build, and someone to run the schedule keeping the other four in sync — plus the SME whose calendar sets the actual timeline. There were no meetings. No planning session. No developer. No creative round. No waiting on an expert. One person, one session, and a set of agents with access to every system that holds our numbers.

I want to be precise about what that means, because “AI made me faster” is the most boring sentence in marketing right now, and it is not what happened. Speed is the least interesting part. The ceiling moved from how fast can I produce to what do I know — and then that ceiling moved too.

The number

My team is five people. We have production access to our own website. Since January we have merged 108 pull requests to production — from marketing, not engineering — across 249 live pages, with fourteen systems of record wired into the working session.

I keep a ledger of the significant work, with the conventional headcount and duration I would have quoted after twenty years of running these functions, set against what it actually took. Six tracked projects. Conventionally: 1,313 person-days. Actually spent: 31.75.

That is a factor of 41.

The stress test: you are entitled to think my conventional estimates are generous. So halve them. Halve every single one and it is still 20.7×. Cut them to a third and it is still nearly 14×. There is no version of this arithmetic that lands anywhere near the 10 or 20% gain the industry is comfortable talking about.

The gap between those two numbers is the story, and it is why I think most marketers have about six months.

The number I would actually lead with

One more figure, and it is the one I would put first if I were trying to convince a sceptic rather than sell a headline.

Of the 275 commits my marketing team has made to production this year, 103 are mine. 172 belong to Andrew Brown, who runs growth. He out-commits me by two thirds.

I could have left that out. It is the most useful fact I have, because it kills the easiest objection to everything above: that this is one enthusiast with an expensive hobby, a CMO with a side project and a flattering spreadsheet. It is not. It is a way of working that transferred to the person sitting next to me, and he is better at it than I am.

That is also why this is a six-month problem rather than a five-year one. If it only worked for the person who enjoys it, you could wait it out.

Using AI is not the same as being AI-native

Almost every marketer I meet is using AI. Almost none of them are AI-native, and the difference is not enthusiasm or prompt quality. It is write access.

An assistant that drafts copy saves you an hour. An agent that reads the warehouse, builds the page, opens the pull request and then measures what happened replaces a workflow. One of those makes you slightly faster at your existing job. The other removes the job as it is currently constructed.

So the skill that matters is not being good at AI. It is working like an engineer. Version control. A single source of truth. Grounding before generating. Freezing a dataset before anyone writes a word against it. Shipping a commit instead of circulating a draft. None of that comes from a marketing playbook, and I do not think it can be faked from the outside.

The uncomfortable corollary: this is not a story about tools levelling the playing field. It is a story about a small number of people in every marketing organisation becoming dramatically more valuable, and the rest becoming difficult to justify.

I have written the org chart. We do not need five brand designers. We need one, with AI. Not five operations people — one. Not a web team, an SEO agency, a paid agency, a marketing ops function and an analyst — one growth person with agents. Multiply that across a department and you are not looking at a productivity gain. You are looking at a budget line that halves, twice, and does not come back. Both people and programs.

You are either excellent with AI, or you are out of a job. I have stopped saying that gently, because every marketer I say it gently to hears “carry on.”

The part nobody is ready for

Here is what I did not expect, and the reason this is not a triumphalist post.

Running this way, I could not tell you exactly how many agents are working on my behalf. Somewhere near a hundred. Not a hundred sessions — a hundred running things, doing work against live systems while I am doing something else.

And the imprecision is the point. This entire post argues for counted numbers over estimated ones, and I cannot put a firm number on my own agent count. Nobody handed me a budget for it. Nobody approved it. There is no cost centre, no owner of record, no dashboard where a CFO could see what any of it costs or whether it is worth it.

I run marketing for a company whose entire product is finding the waste in cloud spend, so I recognise this shape immediately. It is 2013 and someone just discovered they can spin up instances on a credit card. The productivity was real. The bill, three years later, was also real — and by then nobody could remember who started what, or why.

Agent sprawl is the next cloud bill. It will arrive faster, because there is no procurement gate in front of it. Any competent marketer can start ten agents this afternoon and nobody will notice until the invoice does something interesting. The organisations that come through this will be the ones that put visibility, control and budgeting around agent capacity before the capacity becomes the problem.

That is not a reason to slow down. It is a reason to instrument.

What I would do on Monday

Stop evaluating AI tools. Give your best marketer write access to something that matters, and see what comes back. If nothing comes back, you have learned something important about your team. If a merged pull request comes back, you have found the person your department should be rebuilt around.

Then count your agents. If you cannot say how many are running, who owns them, and what they cost, you do not have an AI strategy. You have an unmetered utility and a very good quarter.

One caveat before you copy any of this, and it is the entry on my own list I would most want you to read: build capacity outran data quality. Two weeks from research paper to full launch program was deliverable. Fact-checking it in two weeks was not. Speed does not remove the bottleneck. It moves it somewhere you were not looking.

You have six months to find out where.

The argument, with the full arithmetic: daveanderson.com.au/six-months. The complete working record — the ledger, the counts, the method and the four things it does not solve, all counted from git history rather than self-reported: daveanderson.com.au/ai-native.