Building with AI
A practical guide for founders on building AI-first companies and products that leverage computational labour from day one
AI STARTUP LANDSCAPE 2026
Source: Crunchbase, Qubit Capital
We are living through the most significant shift in how companies are built since the internet. This is the moment for thinkers, builders, and dreamers to create something new. But there is a crucial distinction: companies that are AI-first — where AI capabilities are the foundation — versus AI-added — where AI features are bolted onto existing products. The difference determines whether you are building the future or optimizing the past.
The AI-First Mindset
The mental shift required to build AI-first is profound. Instead of asking "how can AI help our existing process?" ask "what becomes possible with AI that was not possible before?" This is the difference between incremental improvement and transformative products.
Consider that WhatsApp had just 55 employees when acquired for $19 billion, and Instagram had 13 when it sold for $1 billion. AI amplifies this leverage exponentially. Today, companies like HOLYWATER reach 55 million users with only 10% of their 200-person team being engineers — generating over 200,000 creative concepts monthly across 12-15 languages using AI tools.
AI-Added Approach
A traditional CRM that adds a "AI-powered email suggestions" feature. The core product remains the same — AI is a checkbox on a feature list, not a fundamental shift in value delivery.
- • Limited upside potential
- • AI is a feature, not the foundation
- • Easily replicated by competitors
AI-First Approach
An AI agent that autonomously handles customer relationships — researching prospects, personalizing outreach, scheduling meetings, and following up. The entire value proposition depends on AI capabilities.
- • Transformative value proposition
- • Deep moat through AI expertise
- • Scales without proportional headcount
Finding Your AI-First Opportunity
The best AI-first opportunities share a common trait: they unlock value that was previously impossible or economically unfeasible. According to Foundation Capital, enterprise AI is no longer experimental — it is mission-critical infrastructure. Companies are betting billions that AI will automate 30-50% of knowledge work by 2027.
The 100 People Test
Ask: "What service would require 100 people to deliver, but AI could do with 2?" Klarna's AI assistant handled 2.3 million conversations in its first month, cutting resolution time from 11 minutes to under 2 minutes — a $40M profit improvement. These are the high-leverage opportunities.
The Expertise Gap
Where is expert knowledge scarce and expensive? AI can democratize access to expertise in legal, medical, financial, and technical domains. Vertical AI startups building in these spaces report 50-70% cost reductions for their customers.
The Scale Blocker
What valuable work does not get done because it does not scale? Personalized education, 1:1 coaching, detailed analysis. AI-powered chatbots can now handle up to 80% of routine tasks according to IBM, freeing humans for high-value work.
Vertical vs Horizontal AI
Vertical AI goes deep in one industry. Horizontal AI provides broad capabilities across domains. For most founders, vertical is the better bet.
Vertical AI Advantages:
- • Less competition from tech giants
- • Clearer value proposition
- • Faster sales cycles
- • Domain expertise as moat
Examples of success:
- • Harvey (legal AI) - $2B valuation
- • Decagon (customer service)
- • Physical Intelligence (robotics)
- • Anduril (defense)
Technical Architecture
Building AI products requires key architectural decisions early. The good news: you do not need to train your own models. According to Menlo Ventures, 76% of enterprise AI use cases are now purchased rather than built internally, up from 53% in 2024.
Build vs Buy Decision
| Approach | Time to Market | Control | Best For |
|---|---|---|---|
| API-First (OpenAI, Anthropic) | Days-Weeks | Limited | MVPs, most startups, rapid iteration |
| Fine-Tuned Models | Weeks-Months | Moderate | Domain-specific quality, differentiation |
| Custom Models | Months-Years | Maximum | Large orgs, unique requirements, deep moat |
API Pricing Comparison (2026)
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Best For |
|---|---|---|---|
| Claude 4.5 Haiku | $1 | $5 | High-volume, speed-critical |
| Claude 4.5 Sonnet | $3 | $15 | Balanced cost/quality |
| Claude 4.5 Opus | $5 | $25 | Complex reasoning |
| GPT-4.1 | $3-12 | $12-48 | General purpose |
| GPT-4 | $5 | $15 | Established, reliable |
Source: CloudIDR LLM Pricing, IntuitionLabs
Common Patterns
Agent Framework Comparison
| Framework | Best For | Complexity |
|---|---|---|
| LangChain/LangGraph | Complex workflows, broad integrations | Medium-High |
| CrewAI | Multi-agent teams, role-based tasks | Medium |
| OpenAI Agents SDK | Production-ready, function calling | Low-Medium |
| AutoGen | Research, autonomous code generation | High |
Source: Softcery AI Framework Guide
Cost Management
AI costs can spiral quickly. Gross margins for AI companies average 50-60% compared to 80-90% for traditional SaaS. Build cost tracking from day one.
Cost Reduction Tactics:
- • Prompt caching (90% savings on repeated context)
- • Batch API (50% discount on async workloads)
- • Model selection (use Haiku for simple tasks)
- • Response length limits
Monitoring Essentials:
- • Cost per request/user/feature
- • Token usage patterns
- • Cache hit rates
- • Model performance vs cost tradeoffs
Team Structure
You do not need a team of ML PhDs — but you do need the right combination of skills. According to Dover, AI engineer openings have tripled since 2022, with big tech paying $300k+. The competition for talent is fierce, but startups can compete on culture, speed, and impact.
Essential Early
Full-Stack Engineer + AI
Understands LLM APIs, prompt engineering, and can build end-to-end. Look for ML lifecycle skills and comfort with ambiguity.
Product Designer/Manager
Can design AI interactions, handle uncertainty in outputs, and communicate with non-technical stakeholders.
Domain Expert
Deep knowledge in your vertical. They define what "good" looks like and catch errors that engineers would miss.
Add Later
ML Engineer
For optimization, fine-tuning, and building custom pipelines when API-first is not enough.
Data Engineer
Infrastructure for training data, evaluation sets, and production monitoring at scale.
AI Quality/Safety
Evaluation frameworks, safety guardrails, and ongoing quality assurance as you scale.
Hiring Reality Check
71%
of founders now prefer AI-fluent juniors over "2022-style" seniors
+40%
demand increase for ML engineers — the sharpest shortage
90 days
structured onboarding period recommended for AI roles
Source: 0to1 Advisor
Go-to-Market Strategy
Selling AI products is fundamentally different from traditional software. According to BCG, 40% of enterprise buyers now cite seat reduction as their primary way to decrease software spending — AI enables this shift.
Show, Do Not Tell
AI products need demos. Let prospects try it with their own data and problems. The "magic moment" sells better than any pitch deck. Perplexity AI reached 10 million active users and an $18B valuation by letting the product speak for itself.
Address Accuracy Head-On
Every buyer asks "but is it accurate?" Be honest: 47% of enterprise AI users have made decisions based on hallucinated content. Show your quality controls. Position as augmentation, not replacement. The best models now achieve sub-1% hallucination rates for general knowledge.
Pricing Models
The industry is in flux: 73% of AI companies are still experimenting with pricing, testing an average of 3.2 approaches in their first 18 months. Seat-based pricing dropped from 21% to 15% of companies in 2025, while hybrid models surged to 41%.
AI Pricing Models
| Model | Pros | Cons | Example |
|---|---|---|---|
| Usage/Credits | Aligns cost with value, protects margins | Unpredictable for buyers | OpenAI, HubSpot |
| Outcome-Based | Pay for results, low risk for buyer | Hard to measure/attribute | Zendesk AI ($0.99/ticket resolved) |
| Hybrid (Seat + Usage) | Predictable base, flexible scale | Complex to explain | Zendesk, Salesforce |
| Flat/Seat | Simple, predictable | Misaligns value in AI world | Traditional SaaS |
Source: Growth Unhinged
Common Pitfalls to Avoid
The technology should solve a real problem. "We use AI" is not a value proposition. Focus on the outcome, not the technology.
Getting AI to work reliably is hard. Microsoft research shows it takes approximately 11 weeks for developers to fully realize productivity gains from AI tools. Budget significant time for iteration.
AI fails in unexpected ways. 120+ cases of AI legal hallucinations have been documented since 2023, with one $31,100 penalty. Build monitoring, logging, and graceful degradation from day one.
Set realistic expectations. Even top models have 6.4% hallucination rates on legal information vs 0.8% for general knowledge. A product that works 90% of the time can still be valuable if users understand its limitations.
Without proper evals, you are flying blind. Knowledge workers spend 4.3 hours per week fact-checking AI outputs. Build evaluation into your product from the start.
The Founder's AI Stack
Beyond building AI products — use AI to run your company more efficiently. Developers using AI tools report 55% productivity improvements, with Copilot contributing 46% of all code written by active users.
Building
AI-first IDE with multi-file context. 40-50% faster coding workflows for complex projects.
Tight GitHub integration, 90% of Fortune 100 companies use it. 15M+ active users.
Terminal-based agent for complex, multi-file tasks and codebase exploration.
Writing & Content
Long-form writing, analysis, nuanced content. 1M token context window.
Versatile, great for brainstorming and quick tasks.
Integrated into your workspace for docs and wikis.
Research
AI-powered search with citations. Market research, competitor analysis.
Upload documents for deep analysis and Q&A over your data.
Operations & Design
AI-powered automation between apps. Natural language workflow building.
AI-generated React components and UI from descriptions.
Visual assets, mockups, marketing imagery.
Productivity Reality Check
A rigorous METR study (July 2025) found experienced open-source developers using AI tools took 19% longer to complete tasks — despite believing they were 20% faster. However, newer developers and specific task types show clear gains. The key is knowing when AI helps vs when it adds friction.
AI helps most with:
- • Boilerplate and repetitive code
- • Unfamiliar languages/frameworks
- • Documentation and tests
- • Initial drafts and scaffolding
AI helps less with:
- • Deep domain expertise tasks
- • Novel architecture decisions
- • Debugging complex systems
- • Code you know well
Key Takeaways
Start AI-First
Build around what AI makes possible, not what existed before. The best opportunities unlock value that was previously impossible.
Go Vertical
Deep domain expertise beats horizontal plays for most founders. Less competition, clearer value, faster sales.
Use APIs First
76% of enterprise AI is purchased, not built. Start with APIs, optimize later. Speed to market matters more than custom models.
Manage Costs Early
AI margins are 50-60% vs 80-90% for SaaS. Build cost tracking, use caching, and align pricing with your cost structure from day one.
Join the Builders
Planetary Labour is for founders who see AI not as a feature to add, but as the foundation of a new kind of company. We're building the infrastructure for computational labour — join us.
Explore Planetary Labour →Related Articles
How to Build Agentic AI →
Complete technical guide to building agentic AI systems with architecture, frameworks, and deployment patterns.
How to Build AI Agents →
Step-by-step guide covering frameworks, architecture, memory, tools, testing and deployment.
Best Agentic AI Frameworks →
Compare top frameworks including LangGraph, CrewAI, Microsoft Agent Framework, and more.
AI Coding Agents →
Compare GitHub Copilot, Cursor, Claude Code, and more AI coding assistants with pricing and benchmarks.