What you can build with AI right now

Over the past year, the question “what can you actually build with AI” has changed meaning. It used to be a technical question: is the model powerful enough, will it hold up. Today the models hold up almost everywhere, and the question has become a product one: what real problem are you solving, and how fast do you get it into people's hands. I look at AI from exactly that angle, not as a researcher, but as someone who assembles working products out of it.

What follows is an honest look at what you can realistically build right now, in May 2026. No rosy promises and no lists for the sake of lists: just a tour through the directions, with examples a small team or even one person can pull off.

The models matured: the bottleneck is no longer there

Foundation models caught up with almost every practical task this year. Claude Opus 4.7 comfortably holds a million tokens of context and leads on hard coding, GPT-5.5 is strong on agentic and terminal work, and Gemini 3.1 Pro, along with the dirt-cheap Gemini 3.5 Flash, covers reasoning and fast, high-volume operations for pennies. For a product that means one thing: access to intelligence has stopped being an advantage. Everyone has it, and it keeps getting cheaper.

In parallel, an integration standard emerged: MCP, the Model Context Protocol. In a year the number of servers grew from a thousand to tens of thousands, it was adopted by Anthropic, OpenAI, Google and Microsoft, and the protocol itself was handed to the Linux Foundation. A boring but important development: connecting a model to someone else's data and tools became a routine operation rather than a one-off integration every single time.

The bottleneck now isn't the model; it's the idea and the product around it. A great model inside a bad product helps no one.

Agents that actually work

The biggest shift of the year is that agents stopped being a demo. They learned to use tools, navigate a browser, run chains of tasks, and work in parallel as teams. The cost of a single browser-agent task dropped from nearly a dollar to a few cents, and that opens up a whole class of products that were economically pointless just a year ago.

The most obvious and at the same time most underrated direction is vertical agents for one narrow niche. Not yet another universal assistant, but a tool that knows the workflow of a specific profession: an agent for clinics, for photographers, for agencies. The same goes for voice agents that answer calls and book clients for small businesses, where response time has dropped to tens of milliseconds, and an agent like that runs a small company's front desk cheaper than a human operator. A separate story is agentic search over a company's private documents, and the automation of routine through browser agents: data entry, competitor monitoring, processing applications.

If you want something closer to engineering, autonomous bots that fix bugs and open pull requests in one specific repository are no longer science fiction: models pass more than three quarters of the real tasks in the SWE-bench benchmark. The key is to narrow the scope, not to try to cover everything at once.

Generative media: content became almost free

Images, video, voice and music turned from a toy into a production tool this year. Models like Flux 2 and GPT Image 2 are edited with plain text, no masks or layers; video models Veo 3.1 and Kling 3.0 generate clips with audio synced to the scene; ElevenLabs makes instant voice clones and dubs into dozens of languages while preserving the original timbre; and Suno has become a full studio with stem separation.

A lot of clear products grow out of this. Automated product photography that turns a single phone photo into catalog and lifestyle shots. Dubbing and localization services for video and courses, a thin wrapper over a ready API with a clear per-minute price. Short-form video pipelines where a script becomes a set of clips with auto-captions. Ad-creative generators that crank out dozens of variants per audience. And, something close to me, localization of educational content: one course can be re-voiced into many languages while keeping the instructor's voice.

Verticals where AI already makes money

The strongest stories right now aren't in horizontal tools but in specific industries. In healthcare, ambient scribes took off: systems that listen to a visit and write the medical note themselves; tens of thousands of doctors already use them, saving around fifteen minutes of paperwork a day. Next to that is the automation of prior authorizations with insurers, where a single person manages to run a product with tens of thousands of dollars in monthly revenue, precisely because the pain is narrow and sharp.

In law, assistants for reviewing contracts of one specific type are growing, always with a human in the loop who stays accountable. In accounting, automating reconciliations and month-end close is one of the fastest-growing segments of all. And from August 2026 the high-risk provisions of the European AI Act take effect, which by itself creates demand for compliance and documentation tooling. An important caveat: in all of these niches the winner isn't whoever generates text more cleverly, but whoever understands the workflow more deeply and takes trust seriously: signed data agreements, audit trails, the guarantee that the model doesn't train on someone else's sensitive information.

Education is a separate conversation, and one close to me. Studies show that AI tutors deliver a strong effect in a blended format, where there's a live structure to the lessons, and barely work if you leave a person alone with a chatbot. That's an important product lesson: here AI is an amplifier, not a replacement. Language practice with pronunciation feedback and roleplay dialogues is one of the most obvious and in-demand formats.

Business operations: services-as-software

A separate layer is the automation of office work, which analysts already call services-as-software: turning manual-labor budgets into a software subscription. Customer support now genuinely resolves more than half of routine tickets with agents, at a cost of a few cents per resolution versus several dollars for a human. BI copilots learned to reliably turn a plain-language question into SQL, as long as you ground them on a proper metrics layer. Meeting assistants pull out action items and drop them straight into the CRM. There's even a whole new market for monitoring how a brand is mentioned in answers from ChatGPT and Perplexity: search optimization moved into generative answers.

What to choose as an indie founder

If you're building solo or as a small team, you have a rare window right now. Open models like Llama 4, Gemma 4 and Qwen give you near-top quality with no per-token cost, and small models like Gemma 3 and Phi run right on a laptop or a phone. That opens up a whole class of private, offline apps for those who can't send data to the cloud: clinics, lawyers, government bodies.

New distribution channels appeared too. Marketplaces for skills and agents (Claude Skills, the GPT Store, platforms like Replit) pay creators a share and hand you an audience without marketing. You can sell a result rather than a subscription: “audit my store's SEO,” priced per task. And since the MCP ecosystem grew explosively and turned out to be full of holes security-wise (an audit found vulnerabilities in most servers), a separate, honest niche is tooling that fixes exactly that.

And the most underrated path to revenue is the productized agency: deliver a finished result at software margins, without the long race to acquire users. It doesn't sound trendy, but it works, and it's closest to how I started myself.

What to do with all of this

Strip away the noise and the picture is simple. Access to AI is no longer an advantage; everyone has it. The advantage is taste, speed, and distribution: the ability to pick a real problem, get a solution into people's hands fast, and improve it based on how they actually use it. For once, the technology isn't the constraint. The only constraint is whether you dare to build something on top of it and ship it.

So the best AI project right now isn't the smartest one; it's the one you'll actually ship.