Everybody's building. What aren't they asking?

The build wave is real and it's moving fast:

• A VP of Sales built his own Claude Code tooling and moved win rate 8% without buying a vendor.
• Al Chen at Galileo fed Claude Code his entire production codebase and customers noticed the quality lift.
• A weekend-coded CRM has three months of inbound interest nobody asked for.
• A practitioner open-sourced an AI job search system that scored 740+ offers and landed him a job.

I'm part of the build momentum. Every conference I've attended this year has someone declaring SaaS dead because of this trend. Databricks reports that only 19% of organizations have deployed AI agents — but those 19% are already creating 97% of new databases. Is some of it slop? Undoubtedly. But not all of it.

Here's a few key learnings from building:

First, your builds are only as good as the context you bring. Deal history, ICP definitions, competitive positioning, how their buyers actually buy, and so on. Tomasz Tunguz at Redpoint is one of many voices highlighting that model parity means the differentiation is your first-party data. Well that + "taste" + thinking about what to work on and what not to work on + still stitching it all together.

Second, people are underweighting maintenance. Who maintains the context layer once it's built? Across all business domains? And what if you do have that person but they leave? Zach Vidibor at GTM Engineer School calls this the "strategy compression problem." And Jason Lemkin's post-mortem on Agents #001 is the most honest assessment I've read of vibe-coded apps needing daily maintenance.

Third, the economics are shifting underneath the build wave. OpenAI lost roughly $5 billion on $3.7B in revenue last year. Some power users generated $35K in compute costs on $200/month plans — a 175x subsidy. Ben Thompson's compute economics piece calls out the inescapable fact that reasoning models reintroduce real marginal costs into a stack everyone assumed would trend toward zero. And the contracts already reflect it — sub-1-year SaaS contracts have tripled from 4% to 13% since 2023. Buyers are hedging as builders accelerate. Anthropic literally rewrites Claude Code's harness from scratch every few weeks. The models and primitives shift constantly.

I don't think SaaS dies. But I think the maintenance line item that replaces the SaaS line item is going to surprise a lot of finance teams who haven't modeled it yet.

Stated preference vs. revealed behavior. Gartner says 67% of B2B buyers prefer a rep-free buying experience (646 buyers surveyed Aug-Sep 2025 — treat as directional, not gospel). That stat will show up in every Q3 budget deck as justification for headcount cuts, and most of them will get it wrong. The gap between "I prefer rep-free" and "I bought without a rep" is enormous. I've been a power buyer of enterprise software for over a decade, and I'd check the "prefer rep-free" box in a heartbeat. I'd also tell you that every complex deal I've closed involved a human at the table when it mattered. The transactional mid-market may genuinely go rep-free. But for complex enterprise deals, the human isn't the bottleneck — they're the trust layer the buyer needs before they can say yes internally. Or the person they can sue if things go sideways?

The field advantage compounds. Brett Queener's data makes the point from the supply side: 75%+ of early-stage pipeline through events and field moments (his Bonfire Ventures portfolio, not universal — but the pattern holds across the lean GTM teams I've worked with). Not digital-first pipeline. Human-surface pipeline. When execution is commoditized — when everyone can automate outbound and scale content at roughly similar quality — the field tilts toward whoever built the network AI can't replicate. Pick your hill carefully and show up there, physically, repeatedly, until you own the terrain.

The apprenticeship is disappearing. Bill Binch has carried a quota for 116 consecutive quarters. Twenty-nine years of reading rooms, navigating seven-figure deals, knowing when to push and when to listen. That's judgment, and judgment doesn't automate. But if AI handles the commodity work, you need fewer total reps. Which raises the question of how the next generation gets to where Bill is. Cold calls teach them to read tone. Data entry forced them to learn the CRM. Basic qualification built instinct for what a real deal looks like. The busywork was part of an apprenticeship. What happens now?

On April 16, Anthropic shipped Opus 4.7 and deliberately broke backwards compatibility. Existing code threw errors. The new tokenizer consumed up to 1.35x more tokens. The model pushed back on instructions it used to follow.

My long-running agents stopped following their plans overnight. The teams that adapted quickly had three things: version-controlled prompts, eval suites, and a human who could tell them whether the output actually changed. Most teams have none of these. The ones who built the thickest automation layers were, ironically, the most exposed.

It gets worse and more meta if we extract this out a few years. For example, an eleven-year software engineer posted that he'd caught himself unable to debug without AI, something he called the scariest thing he'd seen in the industry. The muscle atrophy is already happening. I'm not saying I panic when Claude goes down, but I'm not saying I don't either.