AI in practice

Skills are plain instruction files anyone can make, and the fastest way to create them is to ask the LLM itself: hand it your ICP and tell it to write the scoring skill it will need later.

The folder structure is the operating structure for both humans and AI; the agent file states purpose, the index file catalogs contents. Rule of thumb: if the agent is optimized, you are leveraged.

Every output should feed its performance data back into the layer, so message tests, conversion rates and social analytics keep improving what the skills produce next.

Metro's CEO ran exactly this process. Her read from three weeks at Baden-Baden with Mercedes, BMW, Bosch, BASF and Allianz leadership: the sense of urgency exists, what is missing is orientation and the how-to.

His fiancee, a technical accounting manager, went from terminal-shy to a fully green GitHub commit wall within weeks. Karpathy built a custom Sonos controller web app in about 30 minutes after asking Claude if it could see the speakers on his network.

Vibe coding barely worked before the late November model releases; different search providers win different query types (Exa strong on people and coding docs), so wrap multiple providers and let the model choose.

The border shifts almost daily: Goeldi hand-coded a market-analysis agent 18 months ago that Codex now does on request, and Claude's Excel and PowerPoint plugins effectively ended a cohort of AI-for-finance startups. Generic knowledge work like NDA review collapsed from two days of lawyer time to two minutes for free, while deep-context work such as M&A remains defensible.

A chatbot is now the base case and earns no pricing premium. On the internal front, Goeldi's portfolio companies have AI resolving 40 to 70 percent of customer service requests end to end, in some cases with the agent reading the code base and filing a pull request that fixes the reported bug before a human looks at it.

Modiano says the best startups she works with treat OpenAI as an extension of their team, are always on an A-B test, and use feedback loops in both directions: to the provider on model behavior and from their own customers via always-on listening.

Her 90-day sequence: align stakeholders on a minimal KPI set (less is more), derive required capabilities from strategy, list which system covers each (Excel or nothing = automation flag), assign end-to-end accountability for data like the invoicing address, clean up, then automate something with immediately visible impact in hours saved or leads won. Without end-to-end accountability, everyone automates in silos.

Boards to build: collections (collected and outstanding), payments calendar, and an operations board with today's cash, recent changes, risks and runway-impact actions. Chasing overdue invoices is trivially automatable with trigger-based emails and Slack alerts; one client's investor asked how they collected so much cash in two weeks after a simple open-invoice trigger to the back office.

Find real use cases through customers of automation freelancers, not LinkedIn posts; the good agencies are fully booked and grow via word of mouth. Keep industry-specific ERPs people know and layer an integration platform as a service on top for orchestration.

First connect CRM to billing; do not build complexity you do not need. The warehouse becomes the single source of truth for revenue and KPIs, built together with the BI or data team, which Simone calls finance's best friend.

Simone implemented three ERPs and calls all of them a nightmare: nothing works out of the box and customization takes time. Reality check: an ERP is mostly accounting plus consolidation, and you still bolt on spend management, FP&A, and treasury tools. Her advice for greenfield stacks: try the new AI-ERP wave (Rillet, Campfire) rather than rushing into legacy NetSuite or SAP.

B2C teams naturally test on themselves; for B2B Vivid routes all new products through the several companies in its own group, which pulled the whole product organization into managing their own entities.

Goeldi: a lot of hidden efficiency potential is buried in your data, and AI can help you find it.

Any daily task that follows fixed rules (copy A to B, rename files, run the same analysis) is a pattern software can capture; now non-developers can have the LLM write those little automations themselves.

The enabling shift: you give the tool an intention and context rather than instructions on how to code, and it reads API documentation, tests cases and iterates on its own.

His conclusion is a posture, not a forecast: stay open-minded and adapt as fast as possible rather than architect the product team of 2029 in advance.

Coinbase is working on embedding payments into web infrastructure so crawling AI agents automatically pay content creators, and Ethereum is building an identity and reputation layer for agent-to-agent transactions. Subscriptions persist because they kill mental transaction costs, so change starts at the margin.

Transparency on risk profiles enables more billable tests and better claims toward insurers. The blocker in Europe is mindset, not technology: willingness to move is scarce outside the DAX.

Microsoft closed its public Bing API in August 2025 to force Azure grounding. Google charges about 35 dollars per 1,000 grounding queries versus roughly 7 at Exa, pricing high to protect a nearly 400 billion dollar consumer search business from being wrapped.

Exa therefore sells latency tiers, from sub-200ms instant search with less reranking to slow deep research with LLM verification in the loop; async coding agents happily wait an extra hour. Getting version-correct API documentation is a niche that matters for agents but not humans.

Underspecification used to cost a week between check-ins; now it costs the three minutes you spent writing the paragraph, so rapid iteration replaces upfront perfection.

Enhanced SaaS is the most common and least original; the newer categories are where the structurally different opportunities emerge. The matrix comes from mapping what is actually appearing in the market.

Complex systems of record like SAP or Salesforce survive, but specific-need companion products can now be built without writing code manually. Meanwhile AI lets established SaaS vendors expand features into adjacent categories in both directions, sharply increasing competitive intensity in niches.

Useful proprietary data relevant to a large customer group is uncommon. Portfolio company Vestigas digitizes large construction sites with a modern agentic stack and mentions AI exactly zero times on its customer-facing website, because customers care that it works, not what powers it.

Wehmeier points to ElevenLabs and Black Forest Labs horizontally, and Wayve (autonomous driving), Corti (healthcare), and PhysicsX (manufacturing) vertically. He frames AI as a once-in-a-century shift with a runway measured in decades, not months.

Modiano compares her startup customers now versus one to three years ago: with models plus tools like Codex, going from idea to working prototype has never been cheaper or faster, so teams can fail and iterate at much lower risk. Her advice for the year: just try, build in public, and ship.

Modiano's two-by-two (workforce versus customers, efficiency versus new use cases) shows enterprises default to internal productivity. Her gaming example: giving every NPC a voice creates a whole new universe inside an existing game, a pie whose size is not even known yet.

The two moves were switching from a zoo of tools to one procure-to-pay flow and moving the outsourced accountant to month-end close and tax only, with AP/AR automated. Zumbuhl notes the biggest impact windows are moments of pain: before an exit, a business model change, or buy-and-build.

Her Swiss accounting system could only connect one bank account while she invoices in EUR and CHF. The weekend build: Supabase as single source of truth, a Make scenario that generates and emails the PDF invoice automatically. Her conclusion: once a pain point is clear, small automation teams deliver value fast for the SMEs that carry the economy.

Clients get excited about every new MCP, OpenAI feature, or AI sales tool. Patrick's rule: humans set the strategy and the value proposition; AI is an enhancement for good ideas, not a replacement for critical thought.

Data centres are among the top polluters by building type. Bill Gates argues AI's energy-saving benefits will outweigh its consumption, a claim still unproven.

Their I-XRAY demo streamed video to Instagram, matched faces, and let them win strangers' trust by pretending to know them, demonstrating that anonymity in public is technically dead.

A Claude-powered expense receipt bot built in 15 minutes handled 1,200 invoices in a month; agent architectures that let an AI orchestrate an AI fundamentally question the right to exist of many SaaS solutions.

A documentary traced the invisible human supply chain behind Google, Facebook, Amazon and Uber products; the polish of AI assistants hides this labor.

This leads to fundamentally different architectures where latency, ranking, and verification trade-offs shift.

Builders must design systems that can evolve quickly because capability rankings change with every release cycle.

While many startups build AI to replace entire jobs, the study suggests demand is inverted: people want expense reports automated before marketing campaigns.

One of the most progressive national approaches to AI in education, setting a precedent for treating AI fluency as a baseline skill.

Steven Bartlett's visualization; paying for one AI tool and using it daily for a month already puts you in the top 1% of AI knowledge on the planet.

Performance marketing, content marketing, outbound, support and admin tasks will all be delegable to AI, so solo founders can build companies that previously required whole teams.

a16z analyzed startup spending on AI tools; the resulting list of state-of-the-art workplace tools shows adoption is no longer confined to engineering.

Sub-Saharan Africa sees AI as the biggest opportunity lever while Western Europe remains the most cautious region.

The record revenue-per-employee numbers of AI-native applications mask a payroll-to-token cost swap rather than genuine efficiency.

Companies that open entirely new customer capabilities create expanding markets instead of shrinking margins.

Building AI infrastructure and actually adopting it are two very different things, a distinction that matters for judging where Europe sits.

His own direct tool usage is collapsing while agent-mediated usage skyrockets. For Miro that layer is the canvas: text chat is a poor medium for iterating, so agents get a visual thought-process surface humans can point at and correct.

The search API market grows around 10x year on year but is still single digit billions. The startup window is to build the best product before Google announces a subsidized search API; a raw Google search API would let anyone wrap and replace Google.com.

Consequence: corporate software approval processes built around scarce engineering effort break when software can be implemented on demand that morning, live while useful, and be discarded. GitHub goes mainstream as the place non-engineers store their ideas.

Goeldi compares the moment to the 2000s birth of SaaS itself: an entirely new commercial model emerged, and nobody yet knows what the AI-era equivalent looks like. Founders need obsessive optimism, but not belief in something obviously wrong.

Quick low-capital proof points offset Europe's harder financing environment, and AI may automate the cross-country selling, regulation, and admin friction that has always dragged on European companies. Goeldi doubts AI will change European risk appetite itself, but it lets young founders throw things at the wall efficiently.

Modiano is explicit that OpenAI will keep improving models and horizontal capabilities for everyone; durable startup value lives in taking what is in the model and creating something specific for a cohort, industry, or user type.

Against the agents-orchestrating-agents vision she saw presented at conferences, Zumbuhl argues real finance teams still work completely differently, that no ERP does end-to-end AI-driven accounting yet, and that someone must monitor accuracy, joking that the former auditor in her wants an auditor agent for the other agents.

OCR invoice reading existed 10 years ago at SoundCloud; AI removes the training burden. But automation fails when invoices scatter across employee inboxes instead of one dedicated mailbox. She sees agents taking over repetitive work next, yet insists finance needs 100 percent accuracy, so human-in-the-loop review of hallucination-prone output remains mandatory. Claude, she notes, has become genuinely good at Excel.

A dedicated tool adds change management, logins and support burden before the framework is even customized. Adopt software only when the sheer number of OKR owners makes spreadsheets unmanageable. A side benefit of shared sheets: teams see each other's draft OKRs and can comment.

Owning the tech stack and regulatory licenses eliminated dependencies for Vivid and allowed faster iteration on customer experience.

Startups need deep domain knowledge or highly specialized workflows that large model providers will not prioritize.

Andreas Goeldi notes OpenAI's Agent Builder killed many business models overnight; it is unprecedented that an incumbent innovates at this pace, so startups must retreat to niches and positions spanning multiple platforms.

Roi Krakovski's analogy: building commodity software yourself is like growing tomatoes instead of buying them.

AI products create expectations: the more intelligent a product appears, the less users tolerate its limitations. A mediocre core AI feature means users try once, get disappointed, and leave. Ship the rest rough, but make sure the magic works.