AI buzzwords: connotation, denotation, and precision
In fraud and AML conversations, people often share AI buzzwords without sharing meaning. Someone says "model," "training," "confidence," or "hallucination," and everyone nods. Later they get to "risk score," "self-learning," "adaptive," or "explainable AI." Same problem, second layer: shared words, unshared definitions.
This post is a field guide for that mismatch. First, what AI buzzwords hide: denotation versus connotation. Then clearer evaluation questions. Finally a deeper dive to unpack the vocabulary underneath. First the general AI/ML terms that show up early on calls, then the fraud and AML terms that show up in program design and vendor reviews.
What AI buzzwords hide: denotation vs connotation
Denotation is the technical meaning. In machine learning, "train a model" has a fairly specific denotation: think labeled data, fit weights, evaluate on held-out data.
Connotation is what people tend to mean in conversation. When someone asks "how do you train your model?" they may mean product improvement, feedback loops, data retention, or whether the system gets smarter over time.
Buzzwords sound precise while remaining fuzzy. Generative AI borrowed a large vocabulary from classical machine learning without inheriting the same mechanics. Fraud and AML then layered another vocabulary on top of that.
We need both the denotation and connotation to understand each other. If you jump straight to "adaptive behavioral analytics" without agreeing what "model" and "training" mean, you'll start having a conversation about the industry on unsure footing.
Clearer questions beat shared buzzwords
If you only remember one thing from this post, remember this table. When a vendor answers a buzzword with more buzzwords, ask the clearer question instead.
Buzzwords are fine as entry points. They are poor as evaluation criteria. If two people use the same words with different denotations and connotations, they'll never know whether they actually understood each other. Shared vocabulary matters.
Layer 1: the AI/ML terms that show up first
These are the terms I hear earliest on customer calls. They are not "too general." They are the shared vocabulary problem. Without them, the industry terms below become too fuzzy.
"Confidence" and "probability score"
I've seen a lot of questions that assume a product returns something like a classical ML confidence or probability score.
That framing makes sense for traditional machine learning. An ML model maps inputs to a bounded output space. "Fraud probability: 0.82" is a natural artifact of that design.
It is a weaker default assumption for generative and agentic AI. Those systems often produce explanations, dispositions, and structured narratives. They may surface confidence-like signals, but those signals are not automatically the same kind of calibrated probability you get from a trained classifier.
So when someone asks for a confidence level, useful follow-ups are:
- Do you mean a calibrated probability from a scoring model?
- Or do you mean how strongly the agent supports a conclusion given the evidence it was given?
- Or do you mean how much of the required evidence and data was actually available to the agent when it ran?
"Model" is doing too much work
Depending on the speaker, "model" can mean:
- A traditional machine learning model (scores, classifiers, rankers)
- A large language model (the foundation model behind generative text)
- An agent or agentic system (model + tools + harness + orchestration)
- The product capability as a whole ("can your model find emerging fraud trends?")
That last one is especially common. Someone asks whether "the model" can do something, but what they really want to know is whether the product can discover new and emerging fraud patterns. The word "model" quietly shrinks the scope of the question.
In practice, generative products are rarely "just a model." They are closer to:
model + harness = agentic AI (or an agentic architecture)
Prefer the more specific term: machine learning model, large language model / foundation model, agent / agentic system, or product capability / investigation workflow.
For a deeper split between AI and ML, see AI vs ML: what's different, what's similar. For tool surfaces around these systems, see What is MCP? The Model Context Protocol, explained as one example of tooling in the agentic AI sector.
"Hallucination" needs a definition every time
"How do you minimize hallucination?" is not a bad question. It is an underspecified one.
Hallucination usually means: the system produced an inaccurate conclusion relative to what a person understands to be true. A person's sense of the truth often rests on a wider set of inputs than the system was given.
In practice, one of the most common "hallucinations" reported in the real world is not the model inventing facts from nowhere. It is the system doing its best with incomplete or incorrect inputs. Every AI system has a goal. Unless you constrain it carefully, it will try to answer even when data is missing.
Before asking how a vendor minimizes hallucination, define what you mean:
- Fabricating entities, amounts, or citations that were never in context?
- Overstating certainty when evidence is thin?
- Reaching a disposition a human disagrees with?
- Filling gaps instead of saying "I don't have enough information"?
If the conclusion looks wrong, ask first: what evidence did the system actually have? A surprising number of reported hallucinations are missing-input and missing-data problems.
"Training," updates, and customer data
"How often is your model trained / updated?" and "Do you train on our alerts?" are where ML vocabulary collides hardest with generative AI and agentic systems.
Most AI product vendors are not training frontier LLMs from scratch. A relatively small set of foundation models get trained on broad internet-scale corpora and released on provider schedules, often every one to six months. Open weights vs closed weights matter for distribution and control, but they are usually not the buyer's real question.
The clearer questions are:
- How often is your agentic AI updated?
- How often do you ship new capabilities?
- What informs those updates?
- How do you validate a change before production?
- Do you train on our customer data?
Foundation-model pretraining is not the same as training on your alerts. A foundation model may have been trained on a huge public or licensed corpus. That is different from a vendor absorbing your alert data into model weights.
Unit21 does not train its agents on customer data. Customer data is used at inference time to investigate the case in front of the system. Product improvement happens through evaluations, harness changes, prompts, tools, and configuration, not by embedding your alerts into a foundation model.
Classical ML training vs agentic improvement
Classical ML training (denotation example)
Suppose you have 1,000 alerts:
- Hold out 500 for evaluation.
- Train on the other 500.
- The algorithm fits weights from patterns in that training set.
- You run the resulting model on the held-out 500 and see whether it separates true positives from false positives.
In that world, training data is inside the model in a meaningful sense. The weights encode patterns learned from those examples.
You can see a related phenomenon with foundation LLMs themselves. There are well-known examples where a short prompt containing a few sentences from a copyrighted book leads the model to continue with large passages nearly word for word. Even after attempts to remove that training, it is still found to exist in the foundation model itself. That is a reminder that pretraining corpora can leave strong traces inside a foundation model.
This is still not the same thing as your SaaS vendor "training on your alerts." It does help explain why people reach for the word "training" so quickly when evaluating which products to use. The next section gets into more detail on what it means to "train" agentic systems.
Agentic improvement ("training" in the AI space)
When people say they are "training" an agentic AI, they often mean something much closer to software development iterations:
- Build a set of test cases
- Run the agent against those cases
- Score whether it reached the right conclusion for the right reasons
- Update prompts, tools, retrieval, or internal knowledge
- Re-run the evaluation suite
The agentic system is not absorbing those alerts into its weights the way a classical ML model does. You are steering behavior at runtime and validating that steering with tests.
Better words for most product work: evaluation, test suites, prompt and harness updates, steering at runtime, capability changes informed by failures.
At Unit21 we do not have a "one agent to rule them all" mentality. This is why "training" an AI does not make sense at face value. Instead we have an entire system built around agentic systems. Simply put, we recommend thinking in terms of LLMs + Harness = Agentic System. This harness-heavy view is explained at more length in Inside Unit21's AI Suite and The New Unit21: why we rebuilt everything around AI agents. We talk less about one magic model and more about the system around it.
Layer 2: fraud and AML terms that need the same precision
Once the foundation vocabulary is clear, the industry terms become useful instead of slippery.
Risk score
Can mean a supervised ML probability, a rules-engine severity, a composite case priority, or an agent's confidence-like recommendation. Ask what a score means and what it is for. Then ask what exactly is scored, how calibration works, and how an analyst should use it.
Self-learning / adaptive
Can mean continuous retraining on labeled outcomes, threshold tuning, rule recommendations, prompt and harness updates, or marketing language with no automatic change loop. Ask what changes automatically, what needs human approval, and what proves the change helped.
Behavioral analytics / anomaly detection
Can mean peer-group baselines, customer-level behavior models, unsupervised outlier detection, supervised fraud models, or an agent reviewing patterns in context. Ask whether the system can automatically detect new or unexpected patterns in your data. Then ask what baseline is used, against whom, and what happens when an anomaly is found.
Explainable AI
Can mean feature importances from an ML model, rule traces, evidence citations, work logs, or a narrative rationale. In regulated fraud and AML work, the useful question is usually: can an examiner follow the evidence and steps that produced this recommendation?
Real-time
Can mean payment-time decisioning in milliseconds, near-real-time alert creation, fast triage after an alert lands, or streaming feature updates underneath a batch investigation. Ask how long a full alert review takes, from transaction in to AI decision recommendation. Then ask what decision is made in real time, at what latency, and what work happens asynchronously.
A simple rule of thumb
Before accepting answers, raise questions:
When you say that word, do you mean the classical machine learning denotation, the generative or agentic connotation, or a fraud and AML program outcome?
If the answer is "I'm not sure," that is useful information. The conversation needs a more specific direction. Set the context and jointly define the words everyone is using: score, foundation model, agent, harness, evaluation set, baseline, latency budget, or product capability.
Summary
- Define denotation vs connotation first, then use clearer questions. Shared buzzwords are a weak evaluation method.
- Foundation terms (model, training, confidence, hallucination) show up first on calls and need clarity.
- Industry terms (risk score, self-learning, adaptive, behavioral analytics, anomaly detection, explainable AI, real-time) have the same problem because the foundation is muddy.
- Classical ML training embeds patterns into weights. Agentic improvement more often steers behavior at runtime and validates with tests.
- Unit21 does not train agents on customer data.
- Precision of language is part of due diligence.
If you want to see what clearer language in an agentic system looks like in a real fraud and AML workflow, take a look at Unit21's AI Agent and ask about the harness, evaluations, and evidence handling, not just "the model."
FAQs
Is an LLM a model?
Yes, in the narrow sense. In product conversations, "model" often secretly means the whole agentic system. Ask which one the speaker means.
Does Unit21 train on customer data?
No. Unit21 does not train its agents on customer data. Data is used to investigate the case at inference time. Improvement comes from evaluations, harness changes, and configuration, not from absorbing customer alerts into foundation-model weights.
Does reducing hallucination mean the AI will never be wrong?
No. It means you define the failure mode you care about and design controls for it: grounding in evidence, refusal when data is missing, citations back to source records, and evaluations that catch unsupported claims.
Should we stop using ML vocabulary entirely?
No. Use it when you mean it. Do not let a shared buzzword hide an unshared definition.
