How AI is reshaping enterprise software and why the TAM is expanding

Executive overview

Enterprise buyers don't want models — they want outcomes. The model is an abstracted component; the software stack, workflow logic, and proprietary data around it are the value. As token costs converge toward zero, gross margins for AI software companies rise, mirroring what happened in cloud.

AI is not a zero-sum displacement of labor budgets. It expands the total addressable market by enabling work that was never done before. The analogy to SaaS is direct: TAMs grew 10x when deployment friction dropped, and AI will do the same.

The cost of intelligence is going to zero — the moat is the software built on top of it.

The "wrapper" debate

• The wrapper critique has ~2% truth: avoid building something ChatGPT will fold directly into its own UI.
• The real question is how much software surrounds the model output — workflow logic, proprietary data, system integrations.
• In B2B, the customer buys an outcome (e.g. automated contract review, customer support resolution), not a model.
• Abstracting the model away lets you swap in better models instantly while the customer experience stays stable.
• Box example: 81% gross margins — nearly the same as file storage once appeared impossible to achieve.

Model companies and commoditisation

• Very few pure model companies exist; most are software companies that happen to have models.
• OpenAI and Anthropic revenue is effectively software/API revenue — security, compliance, SLAs, account management.
• Meta's open-source commitment guarantees a pricing floor: any frontier model must eventually match the cheapest acceptable alternative.
• DeepSeek confirms this dynamic — Meta is structurally forced to match any open-source reasoning advance.
• Token pricing across hyperscalers will converge the same way storage pricing did.

What enterprise actually looks like right now

• Fortune 500 banks are ~10% of the way into general AI assistants, ~1% into anything resembling agents.
• Line-of-business executives don't care about the underlying model; CTOs and AI leads do.
• Goldman Sachs CEO citing AI-generated S1 drafts in 10 minutes signals a vibe shift — equivalent to the moment banks said they'd never go cloud.
• AI creates competitive pressure cloud never did: cloud was a back-end efficiency story; AI visibly changes the output customers receive.
• AI-native new hires won't accept archaic tooling, creating a talent acquisition problem for laggards.

Context vs. core: what enterprises will buy vs. build

• Context: necessary functions that don't differentiate you — HR, ERP, CRM. Buy these from ISVs.
• Core: proprietary value proposition — drug discovery algorithms, wealth personalisation, recommendation engines. Build or heavily customise these.
• Misclassifying core as context (outsourcing your IP) or context as core (reinventing your HR system) wastes resources or creates competitive risk.
• Most AI exposure for knowledge workers by 2030 will come from ISVs, not internal builds.
• Internal chat bots are a temporary phenomenon; GUI and workflow software will reassert once the UX matures.

Business model shifts

• Usage-based pricing is displacing annual contracts in AI — revenue scales automatically with value delivered.
• Outcome-based pricing (per lead, per resolved ticket, per qualified outcome) is emerging as a new category.
• AI is elastic: tasks that once required months of staffing (e.g. 10,000 leads) can be spun up in a week.
• A company that scaled to $12M ARR switched underlying models multiple times; enterprise customers never noticed.
• Margins improve automatically as token costs drop — one company went from 30% to 80% gross margin across model release cycles.

Why the TAM expands rather than transfers

• SaaS expanded the CRM market from ~10,000 addressable customers to millions; AI follows the same logic.
• AI automates work that was simply not being done, not just work that humans were paid to do.
• Contract data extraction, cross-language content, coding acceleration — nearly zero existing spend to displace.
• Competitive markets force reinvestment of AI efficiency gains into growth, not just margin capture.
• The consumer always wins: better products, lower costs, more services — with abundant AI, Jevons paradox applies broadly.

On open source and trust

• Open source is a net positive: startups get speed without licensing costs; enterprises get supported commercial versions.
• Enterprise security comfort is increasing as model providers mature their compliance, privacy, and regulatory controls.
• A small permanent segment (~10%) will always run on-prem or in private enclaves — mirrors the legacy of on-prem software.
• Cloud normalised the idea of trusting infrastructure you don't own; AI inherits that trust foundation.

Why AI couldn't have happened without cloud first

• Moving AI into enterprises in 2005 — even with equivalent model capability — would have been DOA: on-prem data, no modern APIs, no trust in hosted infrastructure.
• Consumer adoption of AI (via ChatGPT, Perplexity, Grok) creates internal pull: workers notice the gap between consumer tools and enterprise systems.
• Enterprise IT stacks now span hundreds of vendors vs. a dozen 15 years ago — they can absorb a new AI vendor without structural disruption.
• Cloud, SaaS, and mobile consumerisation were prerequisites; AI is the next compounding layer.