Future AI Use Cases: What's Coming vs Still Science Fiction

Capability is the easy part

Self-driving cars were five years away in 2016. They are still, by most honest measures, five years away in most cities. The capability — a car that can drive itself — already exists. The diffusion does not.

This guide applies the same lens to everything else AI is supposed to deliver. Below is a forecast in three buckets — plausible in 2-3 years, plausible in 3-5, plausible in 5-10 — and a fourth for the things that will keep being announced and keep not arriving. The aim is to leave you with a calibrated map, not a hype playlist.

2-3 years: the unglamorous wins

The next two years are not about AGI — artificial general intelligence, the still-undefined milestone where a single model matches humans across most cognitive work. They are about three boring deployments that will quietly absorb hundreds of billions of dollars.

First, narrow AI agents. An agent — an LLM wired up to tools, browsers, and APIs — can already book a flight, reconcile an invoice, or refactor a spreadsheet. By 2028, expect agents handling end-to-end workflows in tightly scoped domains: customer-support tickets resolved without a human, expense reports filed and approved, recruitment screens batched, supplier onboarding completed. Not "your AI assistant." Specific, rule-bound jobs with measurable error rates and a clear human escalation path.

Second, AI tutors at scale. Khan Academy's Khanmigo (built on GPT-4) is the current proof of concept: not a teacher replacement, but a 24/7 patient explainer that meets a student at their level and never sighs. Duolingo, Quizlet, and the entire Chinese after-school market are shipping variants. The bottleneck is no longer capability. It is content licensing, evaluation, and the slow trust-building required with parents and ministries of education.

Third, embodied robotics in warehouses. Amazon's Sparrow and Sequoia systems are already picking, sorting, and stowing in active fulfillment centers. Figure, Agility Robotics, and 1X are running humanoids in BMW and GXO sites under real production conditions. This is not a household robot folding your laundry. It is a structured environment with predictable lighting, known objects, and forgiving failure modes — and that is exactly why it ships first.

The pattern is consistent. Where the world is messy, AI loiters in pilot purgatory. Where the world is constrained, it deploys.

3-5 years: science, voice, and code

By the late 2020s, three more capabilities cross the line from demo to default.

AI-accelerated scientific discovery, in narrow domains. DeepMind's AlphaFold 2 (2021) predicted the structure of essentially every known protein — around 200 million — and AlphaFold 3 (2024) extended it to interactions with DNA, RNA, and small molecules. Google's GNoME (2023) identified roughly 380,000 new stable inorganic crystals, more than the cumulative human total. Expect similar leaps in catalysts, battery chemistries, and antibody design. Not "AI cures cancer." AI compresses the search space; human scientists still run the experiments and write the IND filings.

Real-time multimodal interfaces. GPT-4o and Gemini Live already process voice, video, and screen content in near-real time. Within five years, a phone that watches what you watch and answers in your accent is not a wow demo — it is the default user interface for anyone under 25 doing anything that isn't work. The interesting consequence is the slow death of typing as the primary input for ordinary tasks: search, navigation, shopping, scheduling.

Autonomous coding for routine tasks. GitHub claims Copilot writes around 40% of the code in repositories where it is enabled (their telemetry, take with salt). Cursor, Cognition's Devin, and Anthropic's Claude Code have shifted the pitch from autocomplete to "give me a working pull request." The work that survives is judgment-heavy: architecture, debugging strange production failures, deciding what not to build. The tedious 80% — boilerplate, bug fixes, glue code, migrations — gets eaten. Junior developers do not disappear; the path to becoming a senior one gets steeper, because the floor moves up.

What most people miss: the bottleneck has flipped. Three years ago the question was "can the model do it?" Now it is "can your organization absorb it?" Compliance, integration, change management, and middle managers are the friction layer. Whoever solves diffusion captures the value.

5-10 years: the deep professions

Ten-year forecasts are mostly wrong. Treat this section as the longest plausible bet, not a prediction.

General AI assistants worth using. By 2033, a serious "agent that knows your life" — calendar, email, files, finances, health, documents — is plausible. Apple, Google, and Anthropic are visibly building toward it. The blocker is not the model. It is privacy law, liability, and the unanswered question of who gets sued when the assistant books the wrong oncologist or sends the wrong wire.

AI doctors at population scale. Not replacing physicians — extending them. Triage, radiology pre-reads, dermatology screens, mental-health first contact, chronic-condition coaching, medication review. The UK's NHS is already piloting AI dermatology and breast-cancer screens. The genuinely transformative market is not the wealthy world. It is the billions of people without reliable access to a primary-care doctor. There, an "okay" AI clinician on a phone is a vast improvement over the status quo, which is no clinician at all.

AI-driven legal services. Routine drafting, discovery review, contract analysis, due diligence — already happening at firms like Allen & Overy (using Harvey) and inside corporate legal departments. Within a decade, "ask your lawyer" becomes "ask your lawyer's AI, then maybe your lawyer." Bar associations will fight it on unauthorized-practice grounds. They will lose, slowly, the same way travel agents and stockbrokers lost.

The honest position: in each of these, the capability lands well before the diffusion. AlphaFold has been available for five years and most large pharma companies still can't fully operationalize it. The same lag will define the 2030s for medicine and law.

Still science fiction

Four things you should treat with skepticism whenever a CEO mentions them.

AGI on demand. The phrase has no agreed definition. OpenAI's charter calls it "highly autonomous systems that outperform humans at most economically valuable work." DeepMind has its own definition, Anthropic another, Meta a fourth. That is a goalpost on wheels. Whether you call GPT-7 or Claude 5 "AGI" will depend mostly on the stock options of the person making the claim.

An AI-discovered theory of everything. Physics is not a search problem the way protein folding is. Constructing a unifying theory of quantum gravity requires inventing concepts no model has been trained on, validated by experiments that don't yet exist. Models can help — surface anomalies in data, suggest symmetries, accelerate simulations — but the headline "AI solved fundamental physics" will keep being wrong, in interestingly different ways.

Fully autonomous AI corporations. The fantasy of a one-prompt company that markets, sells, hires, banks, and litigates itself ignores the fact that companies exist inside legal systems, payment rails, and trust networks built for humans. You can already run a small Shopify store with agents. You cannot run Pfizer, JP Morgan, or Boeing with them. Not in a decade. Probably not ever in the form the pitch decks describe.

Mass technological unemployment within five years. Every previous wave — looms, tractors, electricity, computers — destroyed jobs at the task level and created them at the role level, on a multi-decade lag. AI will do the same. Faster than tractors, slower than the hot-takes assume. The disruption is real; the timeline is not three years.

What this means for you

If you build, the play is not "use AI." Everyone uses AI now. The play is to find a workflow where the capability already exists but diffusion has not — a regulated industry, a non-English market, a profession with terrible software, a country with a doctor shortage — and ship the integration. The frontier model is, increasingly, a commodity. The deployment is the moat.

If you invest, watch diffusion metrics, not benchmarks. Active enterprise seats, net retention, time-to-value, gross margin after inference cost. NVIDIA crossing $100B in annual data-center revenue tells you the capex is happening. The harder question is which application-layer companies survive long enough to actually see the revenue downstream of that capex.

If you are a citizen — the more common case — the right intuition is patience without complacency. Most confident predictions about 2028 are wrong in both directions: the doomers overshoot the speed, the dismissers undershoot the eventual scale. The capabilities are arriving. The institutions, jobs, and norms around them will move slower than the demos suggest and faster than the regulators would like. Plan, vote, and reskill on that assumption.

Ask Gradient

If you want to stress-test any of these forecasts against current research, try Gradient's Ask Gradient. A good starter prompt: "Which 2-3 year AI deployment predictions made in 2023 actually came true by 2026, and which ones quietly slipped?" The answer is considerably more sobering than the average AI newsletter is willing to admit, which is precisely why it's worth reading.