For 5 years he struggled to get traction. Then with AI he grew 12x— in 1 year | Scott Stevenson, co-founder of Spellbook

Five lean years of experiments set up a legal AI breakout

Spellbook found product-market fit after about five years of struggle, when the team applied new AI models to a problem it already knew well. Staying small and running constant cheap experiments kept the company alive and ready, and a disciplined feature filter made the AI launch land inside the Word workflow lawyers already used.

Takeaways

• Years of struggle ended with an AI pivot. Rally spent years in a valley of despair selling templates and software to law firms. Lawyers rejected rigid templates, which showed they needed bespoke drafting. After GPT-3 and Copilot, a first autocomplete prototype took about two weeks of evenings and weekends.
• Spellbook grew from 100 firms to nearly 2000. Rally had around 100 law firms. The AI product reached close to 2000 firms in roughly a year to a year and a half, with one to two million contracts a year moving through it. Funding reportedly rose from about $10m to $20m within six months.
• Shipping inside Microsoft Word was crucial. Spellbook was built as a Word add-in rather than a web app, despite messy Microsoft APIs. Lawyers already work in Word. Rules for products are explainability in a five second GIF and zero time to value, with no habit change.
• Staying lean let the team keep experimenting. With five to eight people, Rally refused to scale or raise burn until customers pulled the product from its hands. Over a hundred landing pages and an experiment every two weeks produced the learning. Board skepticism and pressure to blitzscale were resisted.
• Small, nimble law firms made fit measurable. Large firms can take one to two years to roll out software, which obscures product-market fit. Rally shifted toward small firms and used a CAC payback well under 12 months as its test. A menu of services pitch cut demo cost sharply.
• Features are ranked by pattern matching, not stories. Spellbook scores about 50 candidate features against 16 characteristics, such as zero habit change and already being in the workflow. Stevenson says this beats verbal reasoning. He distrusts compelling stories and wants board meetings to be story-free.
• Most features show wiggles of death. Stevenson sees a pattern across launched features of brief usage followed by silence. One customer portal, after months of hype, was used only by internal staff. Early usage is therefore not evidence of real pull.

Stevenson's advice to his 2018 self

• Stay lean until product-market fit, even when stakeholders push to scale.
• Launch new apps and brands rather than bolting features onto one foundation. The AI product launched under its own brand and website, separate from Rally.