Current AI Use Cases That Actually Work — A 2026 Field Guide

Forget the demos. Look at the receipts.

In 2023, every keynote promised an AI revolution by Christmas. Three years later the revolution is here in some places, absent in others, and very rarely where the press release said it would be.

This guide catalogues the use cases where AI is actually deployed at scale — measured in percentages of code shipped, tickets resolved, proteins folded, minutes transcribed — and the ones still stuck in the demo loop. Named companies, real numbers, no inflation.

Code generation: the one job AI has clearly taken

GitHub Copilot — Microsoft's coding assistant launched in 2021 — passed 1.8 million paying subscribers in 2024, the fastest-growing developer tool in the company's history. Microsoft has reported that developers accept around 30% of its suggestions, and that on some teams up to 40% of new code is now AI-generated.

Then came the second wave. Cursor — an AI-native fork of VS Code by Anysphere — reportedly crossed $100M in annual recurring revenue by mid-2024, faster than any developer tool before it. Anthropic's Claude Code (a terminal-based agent), Cognition's Devin, and a half-dozen others followed. The category went from "autocomplete on steroids" to "I'll write the function, you review it" in eighteen months.

Why does this work where so much else stalls? Code is structured, has tests, and runs in a sandbox where mistakes are caught instantly — perfect feedback loops. The model writes, the compiler tells it if it's wrong, the developer keeps the steering wheel. Stack Overflow's annual survey reported its question volume dropping by roughly a third between 2022 and 2024 — a quiet but enormous shift in how developers learn.

The interesting question is no longer whether AI can write code, but how much of the developer's job is now review, architecture, and taste rather than typing. That transition has already happened. It doesn't make headlines because there's no demo to film.

The corporate back office: support, search, summaries

In February 2024, Klarna announced that its AI assistant — built on OpenAI's GPT-4 — was handling 2.3 million customer-service conversations per month, equivalent to the workload of about 700 full-time agents. Resolution times dropped from 11 minutes to under 2. The company estimated a $40M annual profit improvement.

Intercom's "Fin" advertises resolution rates above 60% for tier-one support across thousands of customers. Decagon, Sierra, Ada — a whole cohort of companies has built businesses on the same wedge: automate the FAQ, order-status, password-reset bottom 70% of tickets, escalate the rest. It works because customer service is bounded, repetitive, and tolerant of imperfect answers when the alternative is an eight-minute hold.

Klarna walked some claims back in 2025, quietly rehiring humans after customer-satisfaction scores wobbled, which is the more honest signal. AI handles the easy half well, breaks on the hard half, and the blend needs constant tuning. But "the easy half" is still a multi-billion-dollar industry.

Adjacent to support is internal knowledge. Glean — a startup valued at around $4.6B in 2024 — sells exactly one thing: search across every SaaS tool your company uses, with an LLM on top to summarise. Notion AI, Microsoft Copilot for 365, and Atlassian Rovo do versions of the same. The use case is unsexy: "where did we land on the pricing review last quarter?" answered in three seconds instead of forty-five minutes of Slack archaeology.

What most people miss is that the corporate back office — support, internal search, sales-call summaries, contract review — is where most AI value is being captured today. Not in consumer chatbots. In the unglamorous middle.

Drug discovery: real wins, no approved drug yet

DeepMind's AlphaFold 2 (2021) solved a problem biology had worked on for fifty years: predicting protein structure from amino-acid sequence. AlphaFold 3 (2024) extended this to protein-ligand complexes — a critical step for drug design. The public database now contains over 200 million predicted structures, used by essentially every major pharma research group on Earth.

Isomorphic Labs — DeepMind's drug-discovery spinout — signed partnerships worth roughly $3B in upfront and milestone payments with Novartis and Eli Lilly in 2024 for AI-designed candidates. Recursion Pharmaceuticals merged with Exscientia the same year to build the largest AI-native discovery platform, with multiple compounds in early clinical trials. Insilico Medicine has a drug for idiopathic pulmonary fibrosis in Phase II — the first compound designed end-to-end by AI to reach that stage.

The honest number, though: zero AI-designed drugs have been approved by the FDA as of 2026. Drug-development cycles run ten to fifteen years; AI's contribution is concentrated at the discovery end, which compresses years one through three. We will not know whether AI changes the success rate of approved drugs until the late 2020s at the earliest.

Still, AlphaFold alone has, by some estimates, saved hundreds of millions of years of cumulative research time worldwide. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it. If you want to see AI doing something no human could do alone, this is the use case to point at — not the chatbot.

The commodity layer: transcription, translation, summary

OpenAI's Whisper — open-sourced in 2022 — effectively ended paid transcription for most workflows. Otter.ai, Fireflies, Granola, Read.ai, Loom AI: the entire meeting-notes industry runs on Whisper or its descendants. Real-time captioning on Zoom, Google Meet, and Teams is now a default feature.

DeepL — a German company that has been ahead of Google Translate on quality since 2017 — was valued at around $2B in 2024 and handles the bulk of professional translation in Europe. ChatGPT and Claude do the rest. The human translation industry has bifurcated into "literary work that needs a human" and "everything else, which is now machine-edited."

Summarisation is now invisible. Gmail summarises threads, Slack summarises channels, Apple summarises notifications, Notion summarises documents. The feature is no longer a product — it's a checkbox on every SaaS roadmap. When something becomes infrastructure, you stop seeing it.

The pattern across this commodity layer: AI works when errors are recoverable, the human stays in the loop, and the value is shaving minutes off something boring. Not replacing humans. Compressing time.

What's still not working at scale

Full agentic workflows — "tell the AI what you want and it does it end-to-end" — are not in material production anywhere. Cognition's Devin (launched March 2024, valued at $2B by 2025) was demoed autonomously closing GitHub issues; on the standard SWE-bench benchmark it scored under 14%, meaning it succeeded at roughly one real task in seven. Anthropic's "computer use," OpenAI's Operator, and the browser-using agent cohort are still early. They get stuck, hallucinate form fields, lose context, fail silently.

Customer-facing creative output has retreated. Sports Illustrated lost its top editor in late 2023 over AI-generated articles published under fake author profiles. Mass-produced AI imagery now under-performs stock photos in click-through tests. Brands tried it and learned that the floor of human-made content is higher than they thought.

Autonomous coding — the "no engineers needed" pitch — is similarly stuck. AI accelerates engineers; it does not replace them. Every claim of "we shipped a product with no engineers" eventually reveals at least one engineer.

The gap between demo and production is roughly the gap between "works once on Twitter" and "works ten thousand times in a row without paging anyone." That second number is still very hard.

What this means for you

If you're a builder: the wedge is still narrow, structured, automatable workflows where the human keeps the steering wheel. Find a $50-per-seat-per-month problem with a measurable success rate and pick it apart. The most successful AI products of 2024-2025 do one boring thing well — they do not try to be general assistants.

If you're an investor: the durable winners are not the labs and not the consumer chat apps. They are the picks-and-shovels (NVIDIA at roughly $3T market cap, the obvious one) and the vertical applications that own a workflow end-to-end — Cursor in IDEs, Glean in enterprise search, Harvey in legal, Abridge in clinical notes. Models commoditise. Workflows compound.

If you're a doctor, designer, journalist, or small-business owner, the practical move is unromantic: pick three repetitive tasks in your week that involve text, audio, or code. There is a tool for each, it costs around $20/month, and it will save you an hour. The economy-wide productivity gains from AI will look mostly like that — millions of small compressions, not a single visible revolution.

Ask Gradient

If you want to dig further, our Ask Gradient tool indexes every Gradient guide, briefing, and benchmark we publish. Try a question like "which AI use cases have proven ROI in regulated industries today?" and follow the citations. For background on why some tasks suit AI and others don't, our companion guides on what an LLM actually is and the scaling hypothesis explain the underlying mechanics in the same plain register.