It is now embarrassing for a marketing vendor not to claim AI in some form. The result is a category-wide lexical inflation: every template engine is now "AI-powered," every dashboard is "AI-driven," every freelancer with a ChatGPT tab open is suddenly an "AI specialist." Below, the five species I have personally identified in client engagements over the last twelve months. Names changed; behaviors not.
Behavior: a static content template (email, ad, landing page) with three to five variable fields. The variables are filled in by a human or by a simple Mad Libs script. The vendor calls this "AI personalization."
How to identify: ask for two outputs from the same input. They will be identical. Ask for an output on an edge case the template wasn't designed for. The vendor will say "that's not what the system is for."
What it costs you: a premium of 3–10× over what a sober vendor would charge for "we'll write you a template."
Behavior: a thin UI in front of a public LLM API (OpenAI, Anthropic, Google) with no proprietary retrieval, no domain fine-tuning, no evaluation layer. The product is the chat box. The intelligence is whatever the underlying model has on the day you query it.
How to identify: ask the vendor what proprietary data the system uses. If the answer is "the conversation history" or "public web scraping," you have a wrapper. Ask what happens when the underlying model is updated. If the answer is "we get the upgrade automatically," you have a wrapper that will silently regress whenever its supplier ships a change.
What it costs you: $200–$2,000 a month for what your engineer could build in two weekends — and a complete dependency on a third-party vendor's stability.
Behavior: a freelancer or junior employee uses a public LLM as a draft accelerator and bills the output as "AI-augmented service." The work is fine. The pricing is opaque. The skill being charged for is an undergraduate's ability to read a prompt.
How to identify: the deliverables have the cadence of a public LLM (em-dashes, three-bullet structures, the word "leverage" as a verb). The vendor cannot articulate what makes their use of AI different from a customer doing it themselves.
What it costs you: less than the previous species, but you are paying full-service rates for what is now commodity capability.
Behavior: an impressive-looking system that works perfectly inside a sandboxed sales demo and breaks immediately on real customer data. There is always one more integration, one more migration, one more pilot to clear before it is "production-ready."
How to identify: ask to see a deployed instance handling production traffic, with the previous month's KPIs. If the answer is "every deployment is custom," you are funding a permanent R&D project.
What it costs you: 6–18 months of time, plus an internal team that gradually loses the will to push back.
Behavior: a system whose outputs are presented as objectively correct because "the AI decided." No interpretability, no audit trail, no kill switch. Common in lead-scoring and bid-management products.
How to identify: ask why a specific decision was made. If the answer is "the model determined…" with no further detail, you are looking at a black box. Ask what happens if the model is wrong on a specific case. If the answer is "it self-corrects over time," you are funding your own future incident.
What it costs you: invisible, until a high-value customer is mishandled and you cannot explain why.
Before you sign with any AI vendor — ad platform, marketing tool, agency, "agent" — run them against these five. If the vendor is none of the above, ask them to articulate what the underlying capability actually is, in one sentence, without using the word "AI." If they cannot, you have your answer.
This is not a moral judgment. Templates are useful. Wrappers ship faster than custom builds. Juniors with tabs open are how most work gets done now. The problem is not the species; it is the labeling. You should know what you are buying so you can pay an honest price for it.