YouTube Summary · 11 Jul 2026
VIDEOMay Habib, CEO of Writer (valued $1.9B), breaks down why Fortune 500 companies are drowning in AI tools but starving for ROI. She argues the enterprise doesn't need another chatbot — it needs a preferences stack: brand voice, compliance, audit trails, and shared workflows that non-technical power users can operate. Writer trains its own Palmyra models for 90%+ of LLM calls, claiming 30-50% token efficiency vs competitors. The real bottleneck isn't technology — it's organizational rewiring, politics, and trust.
Source: YouTube — Upstarts Podcast: May Habib · Duration: 45:01 · Guest: May Habib, CEO & Co-founder, Writer · Host: Alex Conrad, Upstarts Media
Writer's core insight: model quality is commoditized, but the 'preferences stack' (brand voice, compliance, scaffolding, audit trails) is the real moat. Hermes already does this partially — memory layers, skills, and GBrain encode user preferences. But we could go further: a formal 'preferences layer' that enforces output style, tone, citation requirements, and per-project conventions automatically, not just as memory notes but as active guardrails on every generation.
✅ Actionable: Audit current memory → skill → GBrain pipeline for gaps where preferences are noted but not enforced
May's anonymous CFO benchmarking program across enterprises is brilliant. Hermes could track token usage per task type, per skill, per model — and surface a FinOps dashboard showing where tokens are being burned efficiently vs wastefully. This aligns with M~'s 'never waste 1-3 hours/day' philosophy: track where AI compute is actually delivering value vs just churning.
✅ Actionable: Build a token-usage tracking skill that logs model calls, tokens, and task outcomes
Writer can demo agentic capabilities in 30 days but real ROI takes months because of org/people factors. Same for Hermes: we can build a skill or automation quickly (production), but embedding it into M~'s actual daily workflow (scale) is different. We should distinguish 'demo works' from 'habit formed' in our own rollout tracking.
✅ Actionable: Add a 'Scale Check' to skill deployment — verify actual repeated use, not just initial success
Writer's 'enterprise brain' creates proactive shared context across agents and sessions — if you're about to talk to Blue Shield's CRO, it surfaces all related sessions and agent actions. Hermes has GBrain and session_search but lacks a 'pre-meeting context assembly' skill that auto-surfaces everything about a topic/person before you engage.
✅ Actionable: Build a 'briefing' skill that pre-assembles GBrain + session context for any meeting/topic
Writer's Champions program identifies power users hitting 500M-1B tokens/month and aggregates their patterns to help executives justify investment. Hermes could track which skills/automations are 'champions' (most-used, highest time-saved) and periodically surface a usage report to M~ showing ROI of the system.
⚠️ Worth exploring: Track skill usage frequency and surface a monthly 'Hermes Champions' report
May argues the 'preferences stack' (brand voice, compliance, scaffolding) is Writer's moat because model improvements increase the need for preferences. But this cuts both ways: as models get better at understanding context and instructions, they'll also get better at inferring preferences from fewer examples. What requires a full scaffolding layer today may be a simple system prompt in 2 years. OpenAI and Anthropic are actively building memory, custom instructions, and enterprise features that could absorb Writer's differentiation.
Writer trains Palmyra because 'we never felt the capital gap meant a capabilities gap.' But the labs are spending $10B+ on training runs. May says 'labs may be weeks ahead' — but that gap is widening, not narrowing. Synthetic data helped close the gap in 2023, but the frontier is moving faster now. At some point, maintaining a competitive model becomes a capital expense that a $1.9B company can't sustain against multi-trillion-dollar labs. The 'we're weeks behind' claim may be optimistic.
May dismisses lab consulting arms ('Deploy Co') because 'the enterprise hates service companies.' But Palantir built a $100B+ company on exactly this model. The reality is enterprises hate BAD services, not services per se. Writer's own onboarding involves significant hand-holding — she admits 30-60 days to show value, and the 'rewiring' takes much longer. That IS a services business dressed up as product. If labs offer free consulting (Anthropic's $10M credits), enterprises may take the free option over Writer's paid platform + implicit services cost.
May's advice — 'survive, don't sell, don't quit' — is classic survivorship bias. She survived AND succeeded, so the advice sounds wise. But for every Writer, there are 100 companies that 'survived' right into oblivion. The real question isn't whether to quit, but whether to pivot. May herself pivoted from Cordova to Writer — she didn't 'survive' with the original product. The advice would be more honest as: 'survive by being willing to burn it all down and rebuild,' which is what she actually did.
May shares powerful stories about Sequoia's 'inspiring' rejection and never socializing with male VCs. But she then says she should have had her male co-founder pitch instead — which, if followed, would have reinforced the exact bias she's describing. The tension is real, but the solution isn't to capitulate to bias. Also, her claim that she works 80-90 hours/week 'every week' and isn't 'that good of a mom' reveals the toxic expectation structure, not a strategy. The real insight is that the system penalizes women for having families while rewarding men for the same — and no amount of individual heroism fixes a structural problem.