Founders ditched 2 million monthly users because they disliked the product

1. Killed a hit product — ▶ 1:36

   Tome reached about 2 million monthly users on GPT-generated presentations, with people lining up when inference ran short. The founders still stopped. None of them liked the product, and they couldn't picture discerning users in memos, investment banking or consulting. They considered shrinking the team and waiting for model leaps like 3.5 to 4. The real gap was context: the model didn't understand the presenter, the audience or their relationship, and "no amount of general reasoning gets you past that." Waiting was not a fix.

2. Follow the heat — ▶ 5:20

   Mining a user base of roughly 25 million, the team found the B2B users were in sales and marketing. They signed 12 free pilots. Customers kept asking for more: research, lead qualification, expansion analysis. Lightfield asked for their CRM, call recorder and data warehouse. The hard part turned out to be reconciling incomplete, conflicting data, since the call recorder often disagreed with the CRM. Reorganizing that reality for machines and humans felt like a more enduring company than the one they were running.

3. Negative pricing — ▶ 7:33

   The first product, a go-to-market assistant, was used daily by account executives but nobody would pay. It wasn't Lightfield's data, and 10 other companies competed for the same job. So they shrank the team and built a CRM from first principles in the dark for about 4 months. Nobody wanted a four-month-old CRM. The only asset left was a giant office, so they offered desk space to anyone who used it. Ten startups came, and they were in it daily, furious about missing features and speed, sending Slack feedback every 2 hours.

4. Model first, schema later — ▶ 10:11

   The naive thesis was that modeling the customer relationship matters most, while "the world ran towards CRM for work." Three of the five founders came from Facebook, so they built a chronological activity log: outreach, meetings, documents, product usage, payments. Fully unstructured storage made queries too slow, so they went semi-structured and let the system infer causality. Consultants said the data model is the make-or-break decision, so Lightfield is effectively schemaless. Connect email, calls and warehouse, wait 5 minutes, and refill fields later by traversing the log.

5. Wedge is understanding — ▶ 20:26

   Lightfield had no sharp thesis for breaking into brownfield. It started by winning new companies, with the chief of staff prospecting on LinkedIn and emailing YC startups. Some customers grew from zero reps to 100. After about 6 months of staring at them, the wedge emerged. It sat higher than the "do the work" promise on AI CRM billboards, such as lead scoring and outbound email, which Salesforce will also do. The wedge is helping a company understand itself so it can steer through chaotic company building.

6. Free seats for everyone — ▶ 24:01

   Convincing founders and engineering leaders was expected to be easy. Convincing a newly hired VP of sales was not. So on sales-led plans Lightfield gives the product free to anyone in the company. It shows Lightfield what engineering, support and finance are doing, and it builds network effects that make rip-and-replace harder. When a seasoned VP says they only know Salesforce, others push back: engineers use it to understand customers, finance for revenue recognition, customer success for account scoring. Peiris stays pragmatic on UI, still running meetings from the spreadsheet view.

7. DIY harnesses fail — ▶ 29:42

   Some customers use Lightfield as the system of record and build their own harness over MCP or CLI. Within weeks they usually find Lightfield's harness did more: better entity recognition, better precision and recall, faster. Seed-stage founders who planned to build it over four weekends came back reporting hallucinations and bad emails. Larger companies say they'll build their own company brain, and many return because modeling customers is the hardest part. Peiris's stance is that it's their data, and Lightfield must earn the right to their productivity daily.

8. Charge for work — ▶ 32:28

   Pure seat pricing failed because the heaviest users consumed 10,000x more than the lightest. Pure credits failed too: for 3 weeks nobody touched anything, the "worst three weeks of the company's life." Lightfield sorted its work into four buckets: everyday CRM, pipeline generation, workflow automations and forecasting. It landed on a platform fee plus seats for core CRM and consumption for everything else. Outcome pricing is off the table because outcomes depend on product-market fit. Prospecting for OpenAI would be easy, and prospecting for a seed startup with no website would not.

9. No swim lanes — ▶ 36:53

   At Tome, Peiris says, people were "playing house": product, marketing and CS leaders guarded their lanes, plans ran too far ahead, and pivoting was impossible. Lightfield's 40 people share one daily standup, stack-rank the problems, and whoever is free takes the next one. Planning is continuous. Connections to Linear and Figma let anyone ramp on a customer or the design system, so everyone is a generalist. The bar to start a project is low and the bar to ship is high, enforced by company bug bashes borrowed from Instagram.

10. Paranoid about speed — ▶ 40:02

    The thing that makes Peiris "the most paranoid" is a report on ElevenLabs leaving a startup CRM for Salesforce after about 4 months of waiting on dashboards. Lightfield's wedge is being the best CRM for new companies, so it has to build everything or they will go back to the old world. As a CRM in a red ocean it treats itself as an expansion company, valuing accounts over a 3-year horizon and building for its fastest-growing customers rather than the average one.

11. Valley as reference logos — ▶ 44:12

    Most system-of-record revenue comes from outside Silicon Valley. Peiris treats Valley customers, such as ones that raised $200 million with three go-to-market people, as reference logos for healthcare, fintech or manufacturing. Buyers fear picking the wrong system, since a CRM may be harder to leave than a bank. Health tech worked because the deals have around 50 stakeholders, the context engineering handles that, and Lightfield signed BAAs and ran penetration testing for years. Referenceability makes the next deal easier.

The work is not the wedge. Every AI CRM billboard in Silicon Valley promises some version of do the work: score the leads, send the outbound, run the sequence. Keith Peiris thinks that is the wrong place to attack an entrenched system of record. Salesforce will send emails too. The opening is the layer underneath, a record of the customer relationship accurate enough that once it exists, "the rest should just be prompts and tool calls." It took Lightfield about 6 months of watching customers to see this clearly.

The idea did not start in CRM. It started with a failure at Tome. The presentation product reached roughly 2 million monthly users, and the founders still shut it down. They did not wait for better models, even though they could see the leaps from 3.5 to 4. Their reasoning was that the model lacked context about the presenter, the audience and the relationship between them, and "no amount of general reasoning gets you past that." The bottleneck was information, not intelligence. Everything Lightfield did afterwards follows from that premise.

The pilots confirmed it. Tome's business users turned out to be in sales and marketing, so the team signed 12 free pilots. Customers kept asking for more work: research, lead qualification, expansion analysis. To do it, Lightfield asked for access to the CRM, the call recorder and the data warehouse. The hard part was reconciling them. The data was incomplete and conflicting, and the call recorder often told a different story than the CRM. Then came the sharper lesson. Lightfield built a go-to-market assistant that account executives used every day, and nobody paid for it. It had no pricing power because it wasn't their data, and 10 other companies were chasing the same job. Whoever owns the record owns the leverage. So the team shrank and built the CRM.

The mechanism is deliberately simple. Three of the five founders came from Facebook, so they modeled the relationship the way a timeline does: an activity log of outreach, meetings, documents, product usage and payments. Fully unstructured storage made queries too slow, a needle-in-a-haystack problem. They moved to a semi-structured log that the system uses to infer causality. Ask whether an account is ready for expansion and it reads the conversations, the tickets and the logins, then compares accounts. The same log removes the most consequential decision CRM consultants sell, the data model. Connect email, calls and the warehouse, wait 5 minutes, and fill in the fields later by traversing the log.

Getting customers was harder than building. Lightfield had no clean thesis for breaking into brownfield accounts, so it began with new companies. It offered free desk space to anyone who used the product, and 10 startups came. They were angry about everything missing and sent feedback every 2 hours, which Peiris read as engagement. Some grew from zero reps to 100. Expansion then came by giving the product free to everyone at a company on sales-led plans. When a seasoned VP of sales arrives wanting Salesforce, engineering, finance and customer success have already built their own reliance on it. The Valley is a source of reference logos more than revenue. Buyers fear choosing the wrong system, and a logo in health tech or fintech makes the next deal easier. Lightfield also signed BAAs and ran penetration testing for years before those deals arrived.

Pricing followed the same logic of separating what is predictable from what is not. Pure seats failed because the heaviest users consumed 10,000x more than the lightest. Pure credits failed because nobody touched anything for 3 weeks. Lightfield landed on a platform fee plus seats for core CRM and consumption for pipeline generation, workflow automations and forecasting. Outcome pricing stays off the table, because outcomes depend on the customer's product-market fit. Peiris says Lightfield has to "charge for the work."

The obvious objection is that all of this should erode as models improve. That is the argument Peiris used against waiting at Tome, and it cuts the other way here. His answer is that customers who build their own harness on top of Lightfield's record come back within weeks, because entity recognition, precision and recall are harder than they look. But that evidence is anecdotal and comes from the vendor, and it is drawn largely from seed-stage founders and early enterprise experiments. If better models narrow that gap, the moat shrinks to the data model and the customer relationships.

Peiris knows where the risk sits. He fears speed, not a rival model. A young company like ElevenLabs left a startup CRM for Salesforce after about 4 months of waiting on dashboards. Lightfield's wedge is the best CRM for new companies, and those companies grow into the old world unless Lightfield builds everything first. That is why it runs 40 people with no swim lanes, one daily standup and low-bar, high-bar project rules. It is a bet that organizational velocity keeps pace with customers who scale faster than any CRM has had to serve. Whether that holds when those customers hit 100 reps is still unproven.