What to Build at Unfair Weekend

A lot of work is still held together by screenshots, spreadsheets, forwarded emails, PDFs, voice notes, copy-paste, status calls, and “can you just check this manually?”

That is not because people love working this way. It is because, until recently, building software around messy real-world workflows was too expensive. Now the cost has changed.

A team can take one annoying workflow and rebuild it from scratch. Not as a dashboard that describes the problem. Not as a chatbot that gives advice. As a system that gets the work done.

The opportunity is not “AI for operations.” The opportunity is one specific workflow that people already hate.

The weekend proof

Show the old way and the new way side by side.

The old way: three tools, twenty clicks, ten minutes, a PDF, a spreadsheet, and two emails.

Your way: describe the intent, provide the needed input, and the task is done.

Directions worth exploring

• •invoice matching and reconciliation
• •university admin workflows
• •customs documents
• •supplier comparison
• •event sponsor tracking
• •clinic intake
• •referral routing
• •insurance documentation
• •legal document preparation
• •restaurant ordering
• •construction quotes
• •volunteer coordination
• •food planning for events
• •room booking
• •scholarship applications
• •internal approval flows

The phrase to remember is: one workflow. Not the whole company. Not the whole category. One painful workflow.

AI coding has collapsed the cost of building software by ten to a hundred times. The moat that protected legacy SaaS — millions of lines of code written over decades — is suddenly far less defensible.

That is bad news for incumbents. It is the biggest startup opportunity in a decade for everyone else.

A lot of expensive software is bloated, rigid, badly designed, and painful to use. It forces users to adapt to the software instead of the software adapting to the work.

Build the AI-native replacement. Not a chatbot bolted onto a 2010 interface, but a workflow rethought from scratch.

• •Replace form-filling with intent.
• •Replace manual reconciliation with completed work.
• •Replace software that asks the user to do everything with software that does the work.

And aim higher than project-management clones. Go after the products that look invulnerable.

The weekend proof

Pick one expensive or legacy product and show that its single most-used workflow can be done AI-native in a fraction of the steps, time, or cost. Do not rebuild the suite. Choose the one workflow people actually live in.

Directions worth exploring

• •ERPs
• •CRMs
• •procurement systems
• •accounting software
• •compliance tools
• •healthcare admin software
• •university systems
• •insurance workflows
• •legal operations
• •supply-chain systems
• •industrial control systems
• •chip design tools
• •HR software
• •learning management systems
• •internal enterprise tools

Demo Day moment

Side by side:

• •the legacy way: three tools, twenty clicks, ten minutes
• •your way: describe the intent, provide the needed input, and it is done

The proof is one real task that used to take twenty steps now taking two.

For the last 20 years, software companies sold tools.

A customer bought the CRM, the analytics dashboard, the project management tool, the email platform. Then the customer still had to do the work.

AI changes this. Now a startup can sell the completed job.

• •Not “software for accounting,” but accounting done.
• •Not “software for customer support,” but tickets resolved.
• •Not “software for outbound,” but qualified meetings booked.
• •Not “software for compliance,” but compliance handled.
• •Not “software for research,” but research delivered.

This is one of the biggest shifts in startups right now. The best AI-native companies may look less like SaaS and more like service businesses with software margins.

The weekend proof

The product should produce a finished result. Not advice. Not a dashboard. Not a recommendation. A completed useful outcome.

Directions worth exploring

• •AI grant application service
• •AI accounting operator
• •AI tax preparation assistant
• •AI customer support resolver
• •AI tender writer
• •AI procurement assistant
• •AI research analyst
• •AI compliance operator
• •AI insurance broker for a specific niche
• •AI back-office service for creators
• •AI reporting service for agencies
• •AI outbound operator for a narrow B2B market
• •AI document preparation service before a professional consultation

A good test: would a customer pay for the result even if they never saw the software?

A three-person team can now do work that used to require an agency, a consultancy, or a whole department. That does not mean every agency dies. It means the shape of service businesses changes.

The new agency might be one domain expert, one technical founder, and a swarm of AI workflows. The unfair advantage is not headcount. It is knowing the customer’s pain deeply enough to automate the right parts. This is especially interesting in boring, fragmented, high-friction markets.

The weekend proof

Show one small team doing one high-value service unusually fast. The service should feel specific. Not “AI marketing agency.” More like “AI review-response and local SEO service for independent hotels in Lithuania.” The narrower, the better.

Directions worth exploring

• •AI-native marketing agency for dentists
• •AI sales development agency for one B2B niche
• •AI documentation agency for dev tools
• •AI HR/recruitment service for small companies
• •AI legal ops service for startups
• •AI due diligence service for investors
• •AI tender/proposal-writing agency
• •AI SEO/content operations for old-school businesses
• •AI bookkeeping service for local companies
• •AI support team for small e-commerce brands
• •AI admin service for event organizers
• •AI localization service for online stores
• •AI operations service for restaurants or clinics

Do not build a generic agency. Pick a painful niche. Understand the workflow. Automate the ugly parts. Sell the outcome.

A fast-growing share of software’s users are no longer human. Agents are already browsing the web, doing research, making purchases, monitoring information, writing code, calling tools, and operating legacy systems on someone’s behalf.

But most software was built for humans clicking buttons. So agents are forced to use slow, inconsistent, brittle interfaces: websites, forms, dashboards, dropdowns, popups, unclear docs, and onboarding flows full of assumptions meant for people. Agents need a different foundation.

• •machine-readable interfaces
• •structured documentation
• •clear authentication
• •predictable workflows
• •ways to discover a tool, understand it, sign up, and complete a task with no human in the loop

Every major category of software people use today can be rebuilt for agents as first-class users. The next user of software may not have eyes, hands, patience, or a browser tab open.

The weekend proof

Pick one real tool or service and show that an agent can discover it, onboard itself, and complete a genuine task entirely on its own. The proof is not that you exposed an endpoint. The proof is that an agent used it.

Directions worth exploring

• •agent-readable booking
• •agent-readable ordering
• •agent-readable invoicing
• •agent-readable quoting
• •agent-readable scheduling
• •CRM updates for agents
• •form submission for agents
• •local service discovery for agents
• •checkout flows for agents
• •supplier comparison for agents
• •support-ticket workflows for agents
• •restaurant reservations for agents
• •venue booking for agents
• •gym, clinic, dentist, tutor, or repair-shop workflows for agents

Demo Day moment

Side by side:

• •a human clicking through the original interface
• •an agent using yours to finish the same task with zero clicks

It discovers it, authenticates, understands the available actions, and completes the job on its own. Make something agents want.

Building for agents as users is one opportunity. Building the missing infrastructure around agents is another.

Agents need memory. They need permissions. They need identity. They need payments. They need debugging. They need logs. They need context. They need monitoring. They need ways to recover when they fail. They need interfaces for asking humans for help. They need ways to prove what they did.

A lot of the next great AI companies may not be agents themselves, but infrastructure for agents.

The weekend proof

Build one missing layer around one agentic workflow. Do not build a universal agent platform. Take one agent doing one real task and make it more useful, reliable, observable, safe, or trusted.

Directions worth exploring

• •agent permission system
• •agent activity recorder
• •agent payment layer
• •agent debugging tool
• •agent test suite
• •agent memory layer
• •human approval interface
• •agent marketplace
• •tool registry for agents
• •agent identity layer
• •agent cost monitor
• •agent failure recovery system
• •agent handoff system
• •agent “black box recorder”
• •agent sandbox for risky actions

The question is not “can you make an agent?” The better question is: what does the world need once agents are everywhere?

Most companies are badly documented. The real knowledge lives in Slack messages, old Notion pages, random Google Docs, people’s heads, customer calls, messy spreadsheets, and decisions nobody remembers making.

This is a huge problem for humans. It is an even bigger problem for AI.

If AI is going to do real work inside companies, it needs to understand how the company works. Not in a generic way. In a specific way.

• •Who decides what?
• •Which customers matter?
• •Which processes are real and which ones are fake?
• •What has already been tried?
• •Why did the team choose one path over another?
• •What does “good” look like here?

The company brain is not just a search box. It is a living system that understands context, decisions, workflows, people, and taste.

The weekend proof

Turn one messy company context into a living, usable system. Not “chat with company docs.” That is the boring version. A better version detects repeated workflows, maps decision points, remembers exceptions, and gives agents or humans the context they need to complete the next instance correctly.

Directions worth exploring

• •decision log that writes itself
• •Slack-to-workflow mapper
• •onboarding assistant that explains how a company actually works
• •repeated-question detector
• •product decision memory
• •“why did we do this?” search
• •agent context layer for one team
• •automatic process documentation
• •customer-call memory system
• •internal knowledge graph
• •company-specific AI chief of staff
• •tool that finds contradictions across internal docs
• •system that maps unofficial workflows inside a company

The best version of this is not another internal wiki. It is a company becoming legible to itself.

Most software still assumes every user should see the same interface. That is weird.

• •A beginner and an expert should not have the same UI.
• •A founder and an accountant should not have the same dashboard.
• •A human and an AI agent should not have the same workflow.
• •A user in a hurry and a user exploring deeply should not be forced through the same path.

AI makes software interfaces more fluid. The interface can change based on the user, the task, the context, the data, or the goal. Software becomes less like a fixed product and more like a set of primitives that rearrange themselves.

The weekend proof

Show one interface changing meaningfully based on the user, goal, context, or workflow. Not cosmetic personalization. Real workflow adaptation.

Directions worth exploring

• •onboarding that adapts to what the user wants
• •admin dashboard that rewrites itself for each role
• •form that only asks what it needs
• •website that changes based on visitor intent
• •enterprise tool with human mode and agent mode
• •browser extension that makes ugly software usable
• •product dashboard that redesigns itself based on the user’s goal
• •AI UI layer for complex enterprise products
• •personal operating system for work
• •no-code tool that lets users generate their own internal software
• •interface that becomes simpler as it learns the user

This is not about making prettier chatbots. It is about making software feel alive.

Video, voice, images, and screenshots are no longer just content. They are inputs.

• •A phone recording can become documentation.
• •A screen recording can become a tutorial.
• •A warehouse photo can become an inventory issue.
• •A customer call can become a CRM update.
• •A clinic visit can become a follow-up checklist.
• •A construction site photo can become a report.
• •A lecture can become a personalized study path.

The world produces messy media constantly. Most of it disappears into camera rolls, inboxes, Slack channels, and forgotten folders. AI can turn it into work.

The weekend proof

Take one messy input and turn it into a finished useful output. Not a summary for the sake of a summary. A structured output someone can act on.

Directions worth exploring

• •voice note to project plan
• •screen recording to SOP
• •construction site photos to issue report
• •clinic visit recording to follow-up checklist
• •lecture video to study path
• •customer call to CRM update
• •warehouse photo to inventory exception
• •event footage to sponsor report
• •product demo video to documentation
• •Zoom call to decision log
• •repair-shop photos to quote
• •restaurant kitchen photos to hygiene checklist
• •classroom recording to teaching material
• •field inspection video to compliance report

A good demo here should feel almost magical. Messy input in. Useful work out.

The most obvious AI ideas are already crowded. The less obvious ones live in hard verticals: healthcare, agriculture, defense, manufacturing, construction, energy, logistics, government, education, legal, science.

These industries are messy. They have regulation, legacy systems, offline workflows, weird stakeholders, old software, and domain-specific details. That is exactly why they are interesting.

AI is most powerful when paired with domain knowledge. If you understand a specific industry better than everyone else in the room, that is an unfair advantage.

The weekend proof

Solve one real task for one specific vertical user. Not “AI for healthcare.” Not “AI for agriculture.” Not “AI for education.” One person. One workflow. One painful task.

Directions worth exploring

• •farmers reducing pesticide use
• •nurses documenting care
• •labs managing protocols
• •factories handling maintenance
• •construction companies preparing quotes
• •logistics teams managing exceptions
• •schools handling admin
• •municipalities processing forms
• •clinics preparing referrals
• •researchers tracking literature in one niche
• •lawyers organizing evidence
• •insurance teams checking documentation
• •energy teams monitoring equipment
• •field workers getting AI guidance on physical tasks

The harder the domain, the less likely someone can copy you with a weekend prompt.

As AI moves from answering to acting, people will ask harder questions.

• •Can I trust this?
• •Who approved this?
• •Where did this answer come from?
• •What data did it use?
• •What did the agent do? Can I undo it?
• •Did it hallucinate? Is it safe to deploy? Is it compliant?
• •Is it manipulating the user? Is it leaking private information?

These questions are not side features. They are startup opportunities. The more useful AI becomes, the more trust infrastructure matters.

The weekend proof

Take one AI workflow and make it safer, testable, auditable, permissioned, or reversible. Show the risky version. Then show your version catching errors, asking for approval, logging decisions, limiting permissions, or preventing a dangerous action.

Directions worth exploring

• •evals for customer-support agents
• •approval flow for risky actions
• •audit trail for agent work
• •hallucination monitor for legal or financial outputs
• •compliance checker for AI workflows
• •permission layer for company tools
• •red-team tool for small AI teams
• •“undo” layer for agent actions
• •source-checking system for high-stakes answers
• •AI policy enforcement for teams
• •human-in-the-loop infrastructure
• •safety tests for medical or mental-health products
• •model behavior monitoring
• •confidence and uncertainty display for AI outputs

Some categories carry a much higher burden than an ordinary hackathon app. A therapist, symptom checker, legal adviser, or financial adviser can cause real harm while still looking convincing in a demo. If you work near a high-stakes decision, make uncertainty visible, cite sources, limit what the product claims to do, and design a clear handoff to a qualified human.

Consumer AI does not have to mean another companion app. There are still huge opportunities in personal software that understands taste, context, goals, identity, and timing.

People want help making decisions. They want better recommendations. They want curation. They want learning. They want taste. They want tools that make them more capable without making them feel automated into nothing.

The best consumer AI products will not just answer questions. They will help people choose, plan, express, learn, buy, remember, or improve in a way that feels personal and non-generic.

The weekend proof

Show one product that helps a person do something better in a way that feels specific, useful, and alive. The output should not feel like generic AI slop. It should feel like taste.

Directions worth exploring

• •AI taste assistant for outfits
• •AI interior-design companion
• •AI travel planner with real constraints
• •AI music discovery guide
• •AI personal shopper
• •AI social planner
• •AI learning companion for one skill
• •AI curator for internet rabbit holes
• •AI memory tool for families
• •AI creative coach
• •AI habit tool with social stakes
• •AI assistant for people moving cities
• •AI tool for couples planning life admin
• •AI gift finder that actually understands the person

This is where fun matters. A consumer product does not only need to work. It needs to make people want to come back.