Tools
Which AI model should you actually use?
The mistake is treating “AI” and “LLM” as the same word. A large language model is a remarkable reasoner over language. It is not a forecaster, not an optimizer, not a purpose-built perception system, and not the economical way to classify at volume. The pattern that actually wins on cost, speed, and accuracy is quieter: specialist models do the work, and a language model takes the order, routes it, and reports back in plain English.
AI made output infinite. Judgement stayed finite. Tell us what you are starting with and what you need back, and this tool points you to the specialist that fits, in plain terms and in technical detail.
Find the right model
Two questions.
Start with what you have and what you need out of it. You get a plain-English answer and a real example, then the technical pick and the trap that catches most teams.
What are you starting with?
The short version
Twelve jobs where an LLM is the wrong tool.
Every row is something teams currently hand to an LLM. Every row has a better answer that is cheaper and faster, and once you have labelled data, usually more accurate.
The job
Predict churn or default from customer attributes
Reached for
An LLM with the row in the prompt
Actually fits
Gradient-boosted trees
Boosted trees beat deep learning on tabular data. Cheaper, faster, calibrated, and explainable through SHAP.
The job
Forecast next quarter's demand
Reached for
An LLM asked to analyze the data
Actually fits
ETS, Prophet, or LightGBM on lags
LLMs cannot do arithmetic reliably over a long series, and have no notion of seasonality or a confidence interval.
The job
Sort ten million support tickets into twelve buckets
Reached for
One LLM call per ticket
Actually fits
A fine-tuned DeBERTa or fastText
Orders of magnitude cheaper and faster, and usually more accurate once trained on your own labels.
The job
Find the documents similar to this one
Reached for
An LLM reading everything
Actually fits
Embeddings, BM25, and a reranker
Retrieval is a search problem, not a generation problem.
The job
Spot a defect on the production line
Reached for
A multimodal LLM
Actually fits
PatchCore, PaDiM, or a small CNN
Runs on the device in milliseconds. No round trip, no token cost, no network dependency.
The job
Read a scanned invoice
Reached for
A vision LLM on the raw image
Actually fits
OCR and a layout model, LLM for the tail
Purpose-built OCR is more accurate on dense text and far cheaper. Use the LLM to clean up, not to extract.
The job
Choose who gets the discount
Reached for
LLM judgement
Actually fits
Uplift or CATE modeling
You do not want who will buy. You want who will buy only if you discount. Different question, different math.
The job
Schedule 400 technicians across 3,000 jobs
Reached for
An LLM
Actually fits
A constraint solver or MILP
Combinatorial optimization with hard constraints. An LLM returns a plausible, infeasible schedule.
The job
Recommend the next product
Reached for
An LLM
Actually fits
Two-tower retrieval and a ranker
Recommenders learn from your behavioral log. The LLM has never seen it.
The job
Decide which of two headlines converts
Reached for
An LLM's opinion
Actually fits
A contextual bandit or an A/B test
Get the answer from reality, not from a prior.
The job
Transcribe a two-hour call
Reached for
An LLM
Actually fits
Whisper, then an LLM to summarize
Right tool per stage. Chain them rather than collapsing them into one.
The job
Catch an anomaly in server metrics
Reached for
An LLM
Actually fits
Isolation Forest or an autoencoder
Streaming, unlabeled, and real time. An LLM cannot sit in that loop.
The wider field
Language models are one family out of 14.
Six of the 145 approaches in this guide are language models. The other 139 are how you forecast demand, catch fraud, read an invoice, spot a defect, route a ticket, and decide who to call. Filter by family or by the shape of the answer you need, or describe your problem in the search below.
Family
Shape of the answer
What kind of answer you get back: a number, a category, a decision, and so on.
Select a family, an answer type, or describe your problem above. Rows are families, and the shape filter is the kind of answer you need. Nothing here needs reading end to end.