Generative AI

The Technical Definition

Generative AI is a category of machine learning models that produce new output — text, images, audio, video, code — rather than just classifying or predicting from existing data. The model learns statistical patterns from massive training sets, then samples from those patterns to generate something that looks like it belongs to the same distribution.

Most of the systems your vendors are pitching fall into this category. ChatGPT generates text. Midjourney generates images. GitHub Copilot generates code. The underlying mechanic is the same: pattern-matched probability, not understanding.

What This Actually Means for Your Business

Generative AI is the reason every vendor deck looks the same right now. It’s the engine behind the marketing copy assistant, the contract drafter, the customer service bot, the design tool, and the “AI co-pilot” that got bolted onto the software you already pay for. When your CMO says she wants to “use AI for content,” she means generative AI.

Here’s the part that gets glossed over. These systems are optimized to produce output that looks right. They are not optimized to produce output that is right. Those two things overlap most of the time, which is exactly why the failures are dangerous — they arrive wearing a confident tone of voice and a clean format.

A legal team using generative AI to draft contracts will get drafts that read like contracts. Some of them will cite cases that don’t exist. A finance team using it to summarize earnings calls will get summaries that sound like analyst notes. Some of them will invert a number. A marketing team using it to write product copy will get copy. Some of it will describe features your product doesn’t have. The model isn’t malfunctioning when this happens. It’s doing exactly what it was built to do — generate plausible content.

The cost structure also surprises people. Generative AI looks cheap per query and gets expensive at scale. A customer service deployment that costs $0.02 per response sounds like nothing until you multiply it by two million tickets a quarter. Token-based pricing rewards short prompts and short outputs, which is the opposite of what most enterprise use cases actually need.

The other thing nobody tells you: generative AI is non-deterministic by default. Ask the same question twice, get two different answers. For brainstorming, that’s a feature. For anything that has to be auditable, it’s a problem you have to engineer around.

Reality Check

What the vendor says: “Our generative AI platform creates personalized content for every customer at scale.”

What that means in practice: The system produces a plausible-looking draft that still needs human review for accuracy, brand voice, and compliance — unless you’re comfortable shipping content you haven’t read. The personalization is template-and-variable, not magic.

What Operators Actually Do

The companies getting real value from generative AI treat it as a draft engine, not a publish engine. The output enters a workflow where a human edits, verifies, or approves before it reaches a customer, a regulator, or a board deck. The productivity gain is real — drafts are faster, blank pages are shorter — but it shows up in time saved, not headcount removed.

The other pattern that’s working: narrow the use case. Instead of “use AI to write marketing content,” it’s “use AI to generate three subject line variants for A/B testing.” Instead of “use AI for legal,” it’s “use AI to extract every renewal date from this folder of contracts.” Bounded tasks have bounded failure modes. Open-ended tasks have open-ended ones.

Companies also decide upfront where generative output is allowed to be the final word and where a human signs off. That decision lives in policy, not in the tool. The vendors won’t draw that line for you.

The Questions to Ask

1. What’s the failure mode when the model hallucinates here? If a wrong-but-plausible output ships to a customer or a regulator, what does it cost — and who finds out first?

2. What’s the human review step, and who owns it? Generative output that goes straight to production without a checkpoint is a policy decision, not an oversight. Make sure someone made it on purpose.

3. What does this actually cost at our usage volume? Per-query pricing looks trivial in a demo. Run the math against your real ticket count, document count, or content volume before you sign.