AI tools for marketers: how to build a working stack and get out of prompt chaos
In short
- Which AI model you pick is secondary; the real problem is the missing system: the team has no shared prompts, knowledge bases, or processes, so a pile of subscriptions delivers zero effect.
- Working with AI goes through four maturity levels: personal experiments → shared prompts and guides → tools built into processes → AI agents inside the workflow.
- A marketer's minimal stack: one strong general-purpose LLM + an image generator + the AI already inside your work tools (analytics, ad platforms) – not ten subscriptions.
- There is one criterion of usefulness: time and money saved on a specific process, not the wow effect of a demo.
According to various surveys, more than half of marketers in Russia already use AI tools daily. Yet the typical picture inside teams looks like this: everyone has a personal ChatGPT or Alice (Yandex's assistant) account, their own random prompts, results are never saved, quality jumps around, and the combined effect for the business is close to zero. I call this prompt chaos: the tools are there, the system is not.
This article is about getting out of that chaos: which maturity levels a team passes through when working with AI, what minimal tool stack a marketer actually needs, and how to tell that AI has started making money instead of just entertaining you.
Why "subscribing to AI tools" does not work
An AI model is a tool with no process built in. Buying a subscription gives you a capability, not a result: the model does not know your product, your audience, your tone, or your quality standards. All of that has to be handed over – through prompts, knowledge bases, examples. If every employee does this from scratch and in their own way, the team pays the "first-step tax" forever.
Hence a paradox I keep running into in client work at Matveo: a company pays for 5–10 AI services, and at best one of them produces a measurable effect. The services are not to blame: none of them is built into a process – the task has no owner, the output has no quality criterion, the team has no shared library of what already works.
The four maturity levels of working with AI
In the "One Percent" white paper I described the maturity model in detail; here is its essence. Each next level gives a multiple of the previous effect, and you cannot skip levels.
- 01Level 1 – personal experiments. Everyone tries things on their own: texts, images, "write me a post". Effect: minutes saved locally, nothing reproducible.
- 02Level 2 – shared assets. The team builds a prompt library, tone-of-voice guides, brand references, shared accounts. Results become reproducible: any employee gets an equally decent draft.
- 03Level 3 – AI inside processes. Models are built into specific processes with owners and metrics: a content pipeline, call reviews, ad account cleanup. This is where measurable savings appear – hours and days per week.
- 04Level 4 – AI agents inside the workflow. Multi-step tasks are handed to agents with tools and boundaries; people manage the system. More on this in the article about AI agents in marketing.
An honest self-check
If all of the company's AI subscriptions were deleted today, what would break? If the answer is "nothing", you are at level 1, no matter how many services you have bought. Real maturity is measured by the processes that would stop working without AI.
A marketer's minimal tool stack
The good news: you need very few tools. A working stack for 2026 has three layers:
- One strong general-purpose LLM (Claude or ChatGPT class) – for texts, analysis, strategy, reviews. It is cheaper and more effective to master one tool deeply (projects, knowledge bases, long documents) than ten superficially.
- An image generator for creatives and content – with established brand styles, so that no image ever starts from zero.
- AI inside your work tools – ad platform smart bidding, speech analytics in your phone system, AI features in your analytics. These do not need to be "implemented"; they need to be switched on and configured.
Everything else – narrow "AI for X" services – add only for a specific process, once you hit a limit of the base stack. In 8 out of 10 cases the task is solved by a general-purpose LLM with a good prompt and your data.
How to measure the payoff: one metric instead of the wow effect
The only honest criterion of an AI tool's usefulness is time or money saved on a specific process: "publishing an article took 6 hours, now 2", "cleaning up ad placements took half a day, now 20 minutes", "we never had enough people to review calls, now 100% get reviewed". A vague "it got more convenient" does not count.
In practice: set up a table of the processes where you use AI, with two columns – "time before" and "time after". Review it with the team once a month. Processes with no effect are candidates for switching the tool off or rewriting the prompt; processes with a large effect are candidates for the next level of automation, up to an agent.
And remember: AI amplifies whatever system already exists. If your marketing has no end-to-end numbers and no regular decision cycle, start with the foundation – what systematic marketing is – and plug AI into a workflow that already works.
Frequently asked questions
Which AI tools does a marketer need first?
The minimal stack: one strong general-purpose LLM (Claude, ChatGPT) for texts and analysis, one image generator for creatives, and the AI features inside the tools you already use – ad platforms, phone systems, analytics. You do not need ten subscriptions: most tasks are covered by a general-purpose model with a good prompt and your data.
Why is AI producing no results in our team?
Almost always the cause is a missing system, not weak models: the team has no shared prompt library, no knowledge bases about the product, no processes with owners and metrics. Everyone experiments alone and results never accumulate. The fix is to move from personal experiments to shared assets: common prompts, guides, quality criteria.
How do I calculate the effect of AI in marketing?
For each process, record the time (or cost) before and after adoption: producing a piece of content, cleaning up ad placements, handling a lead, assembling a report. The sum of hours saved, multiplied by the team's hourly cost, plus the effect on conversion, is an honest estimate. The wow effect of a demo is not a metric.
Ruslan Matveev
I build marketing as a system. Founder of Matveo, shipping AI products.
More on the topic