Plain language

Which AI model for which job: a plain-language map

July 26, 2026 · 7 min read

Model names change every few months. The map underneath them doesn't. If you learn the categories instead of the brands, you can pick the right tool for any task — this year and next year — without reading a single benchmark chart.

The map: five kinds of work

Nearly everything a business asks of AI falls into one of five buckets. Each bucket has a different "right tool," and using the wrong one is how people end up disappointed — or overpaying.

1. Working with words

Drafting emails, summarizing documents, rewriting for tone, answering questions about a contract.

This is the home turf of the big general-purpose language models — the flagship chat models from the major providers. For everyday drafting and summarizing, honestly, every flagship is good enough; the differences that matter to a small business are price and comfort, not quality. One rule worth keeping: the more the words matter — a delicate customer situation, anything legal-adjacent — the bigger the model you should use, and the more a human should own the final read.

2. Working with code

Building scripts, automating workflows, fixing spreadsheets, wiring tools together.

The flagship models are all strong coders now, and the practical breakthrough isn't the chat window — it's agentic coding tools that write code, run it, read the errors, and fix them in a loop. This is the category that changed my life: a year of directing these tools took me from nearly zero coding knowledge to shipping real systems. If you only invest in learning one category on this map, make it this one — because it's the category that builds your automations for every other category.

3. Making images (and marketing assets)

Product mockups, social graphics, illustrations, ad variations.

Image generation is a different kind of model entirely — you're picking for style and controllability. The honest small-business advice: these tools are superb for internal drafts, concepts, and volume ("give me twelve versions of this ad idea"), and still hit-or-miss for anything requiring your exact product, exact text rendering, or brand precision. Generate the concepts cheaply; let a human finish the one that wins. And be careful with logos and brand identity — that's still designer work.

4. Reading images and documents

"What's in this photo?" "Pull the totals from these invoices." "Read this handwritten form."

Modern flagship models accept images and PDFs directly, and this quietly replaced an entire category of expensive document-processing software. Scanned invoices, forms, whiteboard photos — a model can read and extract from all of it. The caveat: extraction is judgment-shaped, so spot-check until you trust it, and for high-volume identical documents, consider whether the extraction step is the only AI joint in an otherwise scripted pipeline.

5. Analyzing data

"What's driving returns?" "Find patterns in this spreadsheet."

Here's the trap: asking a language model to do arithmetic across ten thousand rows is using the wrong tool. The right pattern is the one from bucket 2 — have the model write analysis code, then let the code do the math deterministically. The model supplies the approach; the script supplies the accuracy. That's exactly how a gigabyte of government complaint data became readable insight without a single hallucinated number.

Two sizing rules that never change

• Match the model to the judgment, not the hype. Within every category, providers sell big-and-smart and small-and-cheap. Shallow judgment — classify, extract, reformat — belongs on small models. Deep judgment belongs on flagships. Your bill is mostly a record of how often you got this wrong; the token post covers the money side.
• The best model is sometimes no model. If the task runs the same way every time, script it once and retire it from this map entirely.

Learn the buckets, not the brand names. Products get renamed every quarter; "words, code, make-images, read-documents, analyze-data" will still be the map in five years.