Our founders started working with AI over a decade ago, long before ChatGPT, Claude and Gemini became household names. We’ve watched it move from a research discipline to something that’s rapidly changing how businesses operate. We use AI every day, helping us to code, answer business questions or even help us decide where to go for lunch!
But we don’t trust it to do everything. And we think you shouldn’t either.
Why We Don’t Trust AI For Everything
The tools we’ve come to know as “AI” that write copy, answer questions, or make suggestions (the technical name is a large language model) don’t actually know anything. They work by predicting the most likely next word in a sentence, over and over, based on patterns learned from enormous amounts of text.
While this is an incredibly useful trick, it’s also the reason these tools will always, statistically, get things wrong sometimes. They’re not built to care whether an answer is true, only whether it sounds like something true would sound. Most of the time those are the same thing. Sometimes they’re not, and they’re not designed to tell you when they’re not.
This isn’t necessarily a danger when you’re using ChatGPT for a restaurant recommendation, but some businesses build “agentic workflow” chains where one AI’s output automatically becomes the next AI’s input, with no person checking the join. Think of it like a photocopy of a photocopy of a photocopy, except each copy is also occasionally confident and wrong. Each extra link in that chain doesn’t just add a small amount of extra risk. The chance of something going off the rails compounds fast, because every step can introduce a new mistake on top of the last one.
Human In Control
At Decima, AI is a tool we point at specific, well-defined jobs, inside guardrails we’ve built and tested, with a person always checking the output before it goes anywhere near your customers or your budget. We use it to do things faster and cheaper, not to do things unsupervised.
We also use more than one type of AI, because different jobs need different tools:
- Traditional machine learning: pattern-finding models trained to do one narrow job well, like predicting which customers are likely to convert. These are measurable, testable, and consistent, because they’re built for a specific question with a specific right answer.
- Generative AI: the newer, more flexible tools that can draft, summarise, and suggest. Brilliant for speeding up ideas and first drafts. Never left to make final decisions with your money on the line.
The result is straightforward: you get the speed AI offers, without gambling your marketing spend on a system without human supervision. Every experiment we run still sits inside the same Decisive Marketing approach, tested, measured, and refined by people who are accountable for the outcome.
How these models actually work
Large language models are, architecturally, next-word prediction engines built on an architecture called a transformer. During training, the model is shown enormous volumes of text and adjusts billions of internal parameters to maximise the statistical probability that a given word follows the words before it. There’s no internal model of “truth” being checked against, no fact database being queried, and no reasoning process in the way a person would recognise it. There’s a probability distribution over possible next words, and the model samples from it.
This is precisely why hallucination (the industry term for a model confidently producing something false) isn’t a rare glitch to be patched out. It’s an expected output of the underlying method: the model will, with some measurable frequency, produce a highly plausible-sounding continuation that is factually wrong, because plausibility, not accuracy, is what it was trained to maximise. Techniques like retrieval-augmented generation, fine-tuning, and reinforcement learning from human feedback all reduce the rate of these errors. None of them remove the underlying mechanism that produces them.
Why chaining multiplies the risk
If a single model step has some probability of being correct, chaining several steps together (what the industry calls agentic workflows, where one model’s output automatically becomes another’s input) compounds that probability across the chain rather than adding a small amount of risk each time. In simplified terms, if each step is right 90% of the time, five steps chained together without a human checkpoint are right roughly 0.9 x 0.9 x 0.9 x 0.9 x 0.9, around 59% of the time overall, not 90%.
Real-world agentic pipelines are more complex than this: errors aren’t always independent, some steps can catch or correct earlier mistakes, and reliability varies enormously by task. But the direction is well documented and consistent across the research: unsupervised, multi-step AI chains fail far more often than any single step’s accuracy would suggest, and the failure rate is difficult to predict in advance because it depends on how each step’s errors interact with the next.
Why we separate traditional machine learning from generative AI
Traditional machine learning models, regression, classification, and causal inference methods among them, are trained against a specific, measurable objective. That means we can test their accuracy directly against known outcomes and state, with evidence, how reliable they are for a given task. This is the type of model we lean on for things like predicting which customers are likely to convert, because the question has a defined right answer we can check our model against.
Generative AI models are trained for open-ended, fluent output, which makes them excellent for drafting, summarising, and generating variations at speed, but much harder to formally verify, because there often isn’t a single correct answer to check against in the first place.
Where the human checkpoint sits
In practice, this means every generative AI output in our process passes through a defined checkpoint before it can affect a client’s spend, a live webpage, or a reported result: a person reviews it against the original brief and either approves it, edits it, or rejects it. Nothing generated by AI reaches a client’s audience or budget without that step. It’s the difference between AI as an assistant with defined limits, and AI as a system left to run unsupervised.
