AI, ads, and agents: Hot Takes on B2B SaaS's future

Executive overview

Public SaaS stocks are down 30%+ and headlines declare B2B SaaS dead, but the hosts argue this misreads the situation. Software doesn't disappear — it changes names. The real question is whether incumbents can adapt fast enough, not whether software businesses survive.

Three panelists (Rob Walling, Tracy Osborne, Einar Volset) stress-test four AI-adjacent topics: whether AI is truly disrupting SaaS, whether models are meaningfully improving, what ChatGPT ads mean for founders, and whether OpenClaw is a real productivity shift or noise.

AI is software by another name — SaaS dies only if we build God; until then, it's just the next ASP-to-cloud transition.

Is AI killing B2B SaaS

• "AI killing SaaS" requires assuming companies that made billions deploying software will suddenly fail at deploying software.
• Vibe-coding lowers the barrier to building small, niche tools — but enterprise complexity (e.g. HubSpot, Klaviyo) cannot be vibe-coded away.
• The real threat is incumbent complacency, not AI itself — HubSpot that ships AI is fine; HubSpot that doesn't is vulnerable.
• Most non-technical users will keep paying for software; the founder bubble overstates how many people will roll their own.
• SaaS was downloadable software, then ASPs, then cloud — "AI agents" are the next rename, not a death sentence.
• Einar's framing: betting SaaS is dead = betting established software companies will fail at deploying new software capabilities.

Are AI models actually improving

• Tracy: moved from ChatGPT to Claude after repeated failures on Zapier-related queries; for communication and summarisation use cases, Claude is better — but neither feels dramatically different than 12 months ago.
• Einar: biggest personal productivity unlock was Claude Code, a harness around the model, not a raw model capability jump — structural tooling mattered more than any version bump.
• Rob: uses Claude to repurpose spoken audio into prose that matches his writing style; iterated training improved results, but unsure if the same was possible a year ago.
• General consensus: incremental gains per model version, but no step-change inflection; most panelists don't notice upgrades from one minor version to the next.
• Einar's constraint theory: models trained on essentially all available internet text; meaningful further improvement requires either a new data source (Tesla's real-world driving data is cited as an analogy) or a genuine architectural breakthrough — just adding compute and synthetic data hits diminishing returns fast.
• No panelist believes we are at, or near, AGI-level capability that would make SaaS irrelevant.

ChatGPT ads and the opportunity for founders

• Sam Altman pushed back on Anthropic's Super Bowl ad, calling it dishonest; OpenAI's stated principle is to never run ads the way the ad depicted.
• Einar: Claude will eventually run ads too — a free product at consumer scale has almost no other viable monetisation model; radio, TV, and Google all followed the same path.
• Tracy: personally averse, but acknowledges average users are less resistant; Amazon Prime's in-show interactive ads work because enough users convert, even if power users ignore them.
• Early-mover opportunity: ChatGPT ads could be the next cheap ad channel, similar to Google AdWords circa 2005–2006 (5–10 cent clicks) or early Facebook ads (15–25 cent clicks).
• Rob built HitTail from $1,500 to ~$30K MRR largely on early cheap Facebook clicks — the same dynamic may replay on AI-native ad inventory.
• For B2B SaaS founders squeezed by expensive Google and Meta CPCs, AI-platform ads are worth monitoring and testing early.

OpenClaw: useful agent or overhyped

• OpenClaw (formerly ClaudeBot) gives a local AI agent persistent access to your entire computer — files, email, calendar, browser.
• Einar: bought a dedicated Mac mini for it; biggest limitation is that it's only genuinely useful when it has full access to your data, which creates real security exposure.
• Tracy: security concerns outweigh current utility — worried about an agent auto-filtering things she actually needs, analogous to Superhuman suppressing email she cared about.
• Rob: would install it but is waiting for a polished, one-click consumer version rather than investing setup time before Apple or someone ships a better alternative.
• Apple is the natural winner here: already holds the secure context (device, iCloud, iMessage) that makes an always-on agent actually useful — but Siri's current inability to even deduplicate calendar events shows how far behind they are.
• Consensus: real category, not clubhouse-level hype — but current implementation is too rough and too risky for non-technical users; the polished version hasn't shipped yet.

M&A context: asset vs stock purchases

• Roughly 90% of transactions Discretion Capital handles are stock purchases; asset purchases are more common below ~$1M ARR or ~$5M exit price.
• 70% of B2B SaaS acquisitions in the $2M–$20M ARR range involve private equity or PE-owned buyers — most founders don't know this.
• Buyers prefer asset purchases (simpler, re-depreciable); sellers with C Corp status and QSBS eligibility generally prefer stock sales to preserve capital gains tax advantages.
• Rule of thumb: stock transactions become standard around $1M ARR / $5M purchase price.
• Discretion Capital's free M&A guide (discretioncapital.com/guide) covers buyer universe, deal structure, and market pricing for this size range.