Newshot

Team's marketing AI spend is about to become your problem

Original text(via HN AI)
your marketing ai budget
your marketing ai budget

Next year, someone will ask you to justify your AI spend.

Maybe it’s your CFO during budget planning. Maybe it’s your CEO, asking why the AI line keeps growing when the tools were supposed to make everything cheaper. And the question might come in its hardest version: how do you protect both your headcount and your AI budget?

For most marketing leaders, this is new territory. Until recently, AI was a handful of subscriptions. Now teams build agents, marketing engineers show up in hiring plans, and the bill is growing.

Nobody has a playbook for this yet, including me. But it’s the conversation I expect to dominate marketing budgets in 2027, so let’s start untangling it.

This article has two goals:

  • A starting point for cutting AI cost strategically, starting from the marketing decisions you already make every day, then moving to the technical levers behind them.
  • The vocabulary you’ll need when this conversation reaches your desk, so you understand it and can take part when it turns technical.

It’s the first of several pieces I’ll write on this topic. It sits outside most marketing job descriptions, but none of us can afford to ignore it.

Why AI spend is so confusing

If your AI costs feel hard to read, that’s because they are. Here’s why.

It’s spread across too many places. Team ChatGPT or Claude seats. AI features inside your CRM, your design tool, your SEO platform. An API key someone set up to run an agent. A tool a colleague expensed on a personal card three months ago. Nobody sees the whole picture, and often nobody owns it.

Every tool counts differently. One charges per seat. Another sells credits. Another bills per action, or per resolved conversation. And underneath most of them sit tokens: small chunks of text, roughly three-quarters of a word each, which is the unit AI providers use to bill for what a model reads and writes. Comparing what your tools cost is like comparing prices in five currencies with no exchange rate.

In practice, companies pay for AI in four ways.Seats: a monthly fee per person for ChatGPT, Claude or Gemini, with usage limits that agents burn through fast. API access: pay-as-you-go per token, which is where most custom-built agents run. Cloud platforms: the same models bought through AWS, Google Cloud or Azure, so the cost lands on the company’s cloud bill, often far away from marketing. Credits inside other tools: HubSpot, Salesforce or Canva reselling AI in their own currency. Your agents can sit in any of these, and often in several at once.

You pay for work you never see. When you give an agent one task, it doesn’t make one request. It plans, searches, reads files, checks its own work and tries again. Each step is billed. One request from you can mean dozens behind the scenes. By Anthropic’s own measurement, an AI agent uses about four times more tokens than a regular chat, and a team of agents working together about fifteen times more.

The same task never costs the same twice. Ask an agent to research a competitor today and again tomorrow, and the cost can differ several times over, depending on how much it read and how often it looped. A 2026 study from Stanford, MIT and others ran AI agents on the exact same tasks several times: some runs used up to 30 times more tokens than others, and the expensive runs weren’t the better ones. That makes forecasting hard. Budgets built on last quarter’s numbers run out early.

AI gets cheaper, the bill gets bigger. The price of AI has been falling fast, so it feels like costs should go down. They go up, because cheaper AI gets used for more things, more often, by more agents. Economists have a name for this, the Jevons paradox: when something gets cheaper, we use so much more of it that the total bill grows.

Usage tells you nothing about value. The easiest number to track is how much AI people use. It’s also the least useful. A team burning lots of tokens might have found a brilliant workflow. Or it might have left an agent running all weekend.

Put these together and you get the situation many marketing leaders are in right now: a growing bill, no clear owner, and no good way to show what it’s buying.

Why this comes to marketing now

For a while, AI spend was IT’s problem. A few licences, a line in the software budget, done.

That’s changing quickly. Marketing teams now build with AI themselves. Marketing ops automates reporting and lead routing. Content teams run research and production workflows. Some companies hire marketing engineers to do this full time. Whoever does the building, the agents now live inside marketing.

But it changes the economics. Someone who builds agents creates a bill that keeps running after they go home. Agents run on schedules, overnight, in loops. And more and more of that bill lands in the marketing budget.

So when the question comes, how to protect both your team and your AI budget, it will be addressed to you.

How to cut AI cost through the marketing lens

Most advice on AI cost is technical: which model, which settings, which discount. We’ll get to that. But in marketing, the biggest savings come from decisions marketers already make every day. What to do, how well, and how often.

I group them into four.

1. Decide where AI is worth it

Not every task deserves the same AI.

Spend by value. Be generous with the work AI makes possible for the first time, things that used to be very hard or simply out of reach: enriching your persona profiles with real-time insights from forums, communities and review platforms; complex lead scoring across dozens of signals; mining hundreds of sales calls or support tickets for the objections and language your buyers actually use.

Positioning and strategy are not on this list. They should come from people, fully. Spend as little as possible on high-volume, low-value work: tagging, reformatting, alt text, meta descriptions. And some work shouldn’t be done at all. If an agent produces a weekly report nobody reads, stop it.

Not every task needs your most powerful model. An internal email draft doesn’t need the same model as your flagship product campaign launch. Many teams use their most powerful and expensive model for everything, simply because it’s the default.

Some calls belong to people. A marketer’s five-minute judgment call, drawing on context, relationships and timing, often does what an agent looping for an hour can’t, and at a fraction of the cost. The question of tokens or headcount goes both ways. Knowing which work belongs to people and which to agents is a skill in itself.

2. Stop waste before it starts

Most wasted AI spend happens before the AI writes a word.

A sharp brief is a cost lever. A vague brief produces thirty variants. Nobody likes them, so the agent runs again. And again. A clear brief, clear positioning and a well-defined audience cut those rounds. AI multiplies what you already do, including what you do badly. That’s true for your budget too.

Fewer, better pieces. AI made it cheap to publish a hundred articles a month. Agents make that volume expensive again, and much of it wasn’t worth publishing anyway. Fewer pieces, better targeted and higher in quality, were always the better content investment. Now they also help keep your AI bill down.

Keep your brand context lean. If your agents load a 40-page brand book into every task, you pay for those 40 pages every single time. The shared “content brain” I wrote about in my GitHub article does two jobs: it makes outputs better, and if it’s concise, it makes every task cheaper. Tight, well-maintained positioning documents are now a budget item as well as a brand asset.

3. Build assets instead of paying for the same work again

A recurring AI task is a recurring cost. Every month you pay for the same work, from scratch. So for each task you repeat, ask: can this become an asset?

Turn repeated work into scripts. Take a monthly performance report. You can ask AI every month to pull the data, do the calculations and write the summary. Or you can ask AI once to write a small script that does the calculations, run that script every month for free, and use AI only to interpret the results. Christopher Penn makes the same point: the cheapest AI work is the work that no longer needs AI. You don’t need to write the script yourself. Your part is spotting which tasks repeat. Setting it up in the most cost effective way is a job for marketing ops, your marketing engineer or a technical colleague.

Save your instructions as skills. Skills are an effective way to get consistent output, and a cost-effective one too. Your AI reads only a short description of each skill until a task needs it, so you stop paying for the same long instructions in every conversation.

Decide how often each agent really needs to run. Does competitor monitoring need to run every week, or is once a month enough? Does the dashboard need a daily refresh if you review it on Mondays? Many marketing agents run far more often than the decisions they inform and spend more money than they should.

Switch off the agents nobody uses. Scheduled agents and automations rarely get switched off. The person who built one moves to another project, nobody reads its weekly report anymore, an early trial version keeps running next to the final one. Audit agents the way you audit software subscriptions: who owns it, what it produces, when someone last used the output.

4. Make the spend defensible

The goal here is a bill you can explain to anyone who asks.

Separate experiments from operations. Give experiments their own capped budget, apart from the agents that have already proven their value. You keep encouraging your team to try new things, without handing out a blank cheque.

Measure what you can, even roughly. Total AI spend says very little on its own. Exact cost per article or per lead is hard to get today: most AI bills don’t break down by task, and marketing attribution was messy long before AI. But you can start small. Pick one workflow you can isolate, like an agent with its own account or a tool with its own credits, and divide its monthly cost by what it produces. Then include your total AI spend in the marketing costs you already report, so it sits next to your other investments. The numbers will be rough, but they give you a far better starting point for the budget conversation than no numbers at all.

Consolidate. If you’re paying for AI credits in eight different tools on top of your own AI subscriptions, you’re probably paying several times for similar work. Where you can, consolidate.

The technical levers worth knowing

These will mostly be handled by IT, your marketing engineer or your vendors. You won’t be configuring them yourself. But a marketing leader needs to start building literacy here, so you can ask the right questions and understand the answers.

Here are the four that matter most, and ones you’ll hear about. Ask the questions below of whoever builds and runs your agents: IT, marketing ops or your marketing engineer.

Model routing. AI providers offer families of models: large, expensive ones for hard reasoning, and small, fast, cheap ones for simple tasks. Routing means sending each task to the cheapest model that can still do it well, often automatically. Classifying 2,000 survey responses can go to a small model. Writing the insight summary from them can go to a large one. This is the technical side of deciding where AI is worth it.

The question to ask: do we match models to tasks at all, or does everything run on the same one?

Context discipline. Context is everything the AI has to read before it answers: your instructions, the files you attach, the conversation so far, the results of its previous steps. Agents re-read their context at every step, and you pay for it every time. Long conversations and oversized files make every step more expensive. Starting a fresh chat for a new task, and attaching only what the task needs, already helps. At team level, the bigger issue is whether there’s a principle behind what your agents are given at all. Without one, the default is to give them everything just in case, and that default is expensive.

The question to ask: do our agents get only the information each task needs, or do we give them everything just in case?

Prompt caching. When the same block of text is sent again and again, like the same instructions or the same brand guidelines, providers can remember it and charge up to 90% less for reading it again. The catch is that the repeated part has to stay unchanged and sit at the start of the request. A lean, stable content brain is a perfect candidate.

The question to ask: is caching switched on for our agents, and is it actually being used?

Batch processing. If a task can wait a few hours, the major providers offer a discount, typically half price, for processing it in bulk instead of instantly. Think overnight refreshes of old articles, monthly reports, or meta descriptions for a few hundred pages.

The question to ask: which of our tasks really need an answer right now?

A note on self-hosting. Self-hosting, also called running on-prem, means running an AI model on your company’s own servers instead of paying a provider each time you use it. It’s becoming a conversation in larger companies with high, steady volumes of AI work. It’s a complex decision, though. It depends on how much work would move, the full cost of hardware and the people who run it, whether an open model is good enough for your tasks, and, especially in Europe, whether your data is allowed to leave the company. If IT brings it up, you’ll at least know what they mean and why the answer is rarely simple.

What to check this week

You don’t need a developer for any of these. An afternoon and a spreadsheet will do.

  • List every AI tool and credit your team pays for. Include the ones on personal cards and the free trials that quietly turned into paid plans.
  • Find the owner of each agent and automation, and how often it runs. Anything without an owner is a candidate to switch off.
  • Pick your three most frequent AI tasks. Could any of them become a script, a template or a skill?
  • Check whether automatic overage billing is on. Some AI tools keep charging past your plan’s limit unless you switch that setting off.
  • Work out a rough cost per output for one workflow. Pick one you can isolate, like an agent with its own account or a tool with its own credits. Divide its monthly cost by what it produces: articles, reports, briefs.

Frequently asked questions

Why are AI costs so hard to track?

Because they’re spread across too many places, and every tool counts differently. Your team might pay for ChatGPT or Claude seats, API keys, the company’s cloud bill and AI credits inside tools like HubSpot or Canva, all at the same time. Agents add another layer. One task can trigger dozens of billed steps you never see, and the same task can cost a different amount every time it runs.

How do companies pay for AI?

In four ways. Seats are a monthly fee per person for ChatGPT, Claude or Gemini. API access is pay-as-you-go per token, and it’s where most custom-built agents run. Cloud platforms sell the same models through AWS, Google Cloud or Azure. And tools like HubSpot, Salesforce or Canva resell AI as credits in their own currency. A single agent can draw on several of these at once.

How can marketing teams reduce AI spend?

Start with four moves. Decide where AI is actually worth it, and use smaller models for simple work. Stop waste before it starts with clear briefs and fewer, better pieces. Turn repeated work into assets like scripts and reusable skills, and switch off the agents nobody uses. Then make the spend defensible: separate experiments from proven operations, and estimate the cost per output.

What is model routing?

Model routing means sending each task to the cheapest model that can still do it well. AI providers offer large, expensive models for hard reasoning and small, cheap ones for simple tasks. Classifying 2,000 survey responses can go to a small model. Writing the insight summary from those responses can go to a large one.

Where should I start with managing my team’s AI spend?

With an afternoon and a spreadsheet. List every AI tool and credit your team pays for, including personal cards and trials that quietly became paid plans. Find the owner of each agent and how often it runs. Look at your three most frequent AI tasks for anything that could become a script or a skill. Check whether automatic overage billing is on. Then estimate the cost per output for one workflow you can isolate.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by HN AI
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗