What AI is good at

Strip away the hype and the warnings. Here is the honest list of what a current AI model does well, and the property they all share.

The honest list

Drafts of anything written. Emails, briefs, descriptions, proposals, sales copy, internal documentation, summaries, rewrites. Anywhere the value is “get me from blank page to something I can edit.”

Transformation between formats. Long to short. Casual to formal. Bullet points to paragraphs. English to Hindi. Tables to prose. Code in Python to code in JavaScript. Models are extraordinary at restructuring information without losing the substance.

Extraction from messy text. Pull invoice numbers from a stack of PDFs. Pull customer sentiments from a year of support tickets. Pull the action items from a meeting transcript. Models read fast, don’t get bored, and surface structure from noise.

Classification and routing. Is this support ticket about billing or product? Is this email urgent or routine? Is this CV worth screening? Models can categorize text at scale, replacing rule-based logic that used to take a team to maintain.

Brainstorming and expansion. Give it a half-thought, get back ten developments of it. Give it three product names, get back thirty more in the same flavor. Models are good at “more like this.”

Compression and summarization. Long meeting transcript to key decisions. 50-page report to a one-page brief. Months of customer feedback to recurring themes. Compression is where models often beat humans on both speed and consistency.

Code. Code is mostly text, and the model has seen a lot of it. Drafts, autocompletes, explanations, translations between languages, finding bugs in small snippets. Not flawless, but a force multiplier for anyone who writes code or wants to read code.

Conversation about anything. “Walk me through this concept.” “What questions am I not asking?” “What’s the counter-argument?” Models work well as a thinking partner for thinking that benefits from being articulated.

What all of these have in common

Each task above shares two properties.

The answer can be checked. Either you can read it and judge (“does this draft sound right?”), or it has a verifiable structure (“does this extraction match the source?”). The model produces, the human verifies — fast.

Being approximately right is useful. A draft that’s 80% there is a great starting point. A summary that captures 9 of 10 themes is a huge time-saver. The model doesn’t have to be perfect; it has to save more time than the verification costs.

Where one of these properties breaks, the use case usually breaks. Anywhere the output is unverifiable, or where 80% accuracy is the same as zero accuracy, look elsewhere or add structure — the territory covered in what AI is bad at.

Where this shows up in a business

The shape of the list maps directly onto common operational work. Writing similar things from scratch — sales emails, internal updates, customer responses. Manually categorizing — tickets, leads, applications. Reading volumes to extract specifics — PDFs, contracts, transcripts. Reformatting existing content across formats or languages. The blank-page tasks that wait on someone finding time and headspace.

Each of these is shaped like “model produces, human verifies, time is saved.” Off-the-shelf tools usually clear the first version of these tasks without custom software. The compounding gains come from doing this in places where the work already happens, at the volume it already happens at.