After 25 Million Users at Tome, He Tore It Down to Build a CRM
Keith Perez walked away from Tome and its 25 million users because his own team didn't like the product; he judged that the CRM battle isn't about making salespeople do more work, but about whether you can fully model a company's customer reality.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
They killed a hit not because of the data, but because they didn't like it
Tome launched around the same time as ChatGPT and reached 2 million monthly users, so many that inference compute ran short and people had to queue. But Keith says what really made them stop wasn't a metric — it was that "none of us liked the product." Concretely: they couldn't see a path where a discerning, high-quality presentation maker would treat Tome as indispensable, nor one where investment banking or consulting people would use it to write memos. Their conclusion was that what the models lacked at the time wasn't general reasoning ability but context about who the presenter is, who the audience is, and what relationship holds between them — and that gap wasn't going to close just by waiting for models to get stronger.
— Keith PerezFollow the heat first, then discover the real problem is data
The Tome team dug through their own user base and found B2B users clustered in sales and marketing, so they ran free pilots with 12 large companies. The original idea was to help sales build new-business decks and proposals, but customers started asking: can you also do research, help us qualify leads, help us understand which companies are worth expanding into. They followed that heat, asking customers to open up access to CRM, call recordings and data warehouses, and found the hardest part wasn't doing that work — it was that the data in these systems conflicted with each other. The reality in call recordings and the reality in the CRM often didn't line up. Keith says reorganizing that data started to look like the most important work, so they concluded: if you can reorganize a company's reality into something both machines and people understand, that's more worth doing than the original path.
— Keith PerezNobody wants your four-month-old CRM, so pay them to use it
What they built first was a go-to-market assistant that AEs used every day, but they couldn't charge for it — the data wasn't theirs, and ten other companies were fighting over the same thing. So they cut headcount, rethought the CRM from scratch, worked in stealth for four months, and then discovered nobody wanted a four-month-old CRM. The only asset they had left was a huge office lease they couldn't get out of, so they posted on X and LinkedIn: use our CRM and we'll let you sit in our office for free. That's "negative pricing." The 10 startups they brought in were there every day, complaining about slowness and missing features, but giving feedback in Slack every two hours. Keith says this was the exact opposite of Tome: a barely working product that users were in every day, and that they cared about enough to give feedback by the hour.
— Keith PerezThe core of a CRM is relationships, not fields
Three of Lightfield's five founders came from Facebook. Their plain judgment was that the most important thing in a CRM is modeling relationships, so they borrowed directly from the Facebook timeline and built an activity log first — when you first made contact, what you said, what they replied, which meetings happened, which documents went back and forth, and eventually what the other party did inside your product and how they paid. All the traditional CRM field updates and stage progressions are triggered from that log, but underneath there's always a canonical relationship log. They also tried going fully unstructured and found queries too slow — a needle-in-a-haystack problem — and landed on semi-structured: stuff a lot of unstructured data into the activity log and let the system infer causality from the log.
— Keith PerezGet the data model wrong and sales will never go back and fill it in
They interviewed a lot of CRM consultants, asking what they actually do for clients, and found the highest-consequence thing is the data model: set your stages and fields wrong and you can't send salespeople back in time to fill in the data — it's over. So Lightfield chose to be "schema-less" — connect email, drop in a call recording, hook up the data warehouse, and the system assembles the relationship in real time; fields come later. Change your mind and it traverses the activity log and backfills. Alex Rampell summed this up as "intelligence is greater than schema," and contrasted it with the old era: back then, to save space, every column in a table had to have a predefined varchar length, like name varchar 25 — now they don't even want a schema.
— Alex RampellSomeone used it to find cutting-edge treatment for an Alzheimer's patient
Because it supported fully arbitrary schemas, custom objects and custom relationships from the start, Lightfield attracted a set of customers with very strange business models. Keith's example is Power: on one side it serves pharma companies, helping them find clinical trial participants; on the other it's a marketplace aggregating patients with hard-to-diagnose conditions and seeking cutting-edge treatments. They modeled both the B2C and B2B sides in Lightfield, wrote automations to scrape the FDA and clinicaltrials.gov, built a world model of "every ongoing clinical trial in the world," and then matched patients to pharma companies. The result: Lightfield genuinely helped an Alzheimer's patient find a cutting-edge treatment within days.
— Keith PerezWin the founders and engineers first, then wait for the sales VP to come around
Alex points out something often overlooked: the sales VP a greenfield company hires was trained by Salesforce and will say flat out, "I don't use that." Lightfield's design principle is to win first the founders, engineering leads and product leads — the people willing to use frontier technology to understand customers — while making every sales-led capability free for anyone in the company to use. That both helps the company understand what engineering, support and finance are doing and creates real intra-company network effects that make replacement harder. When that senior sales VP arrives and finds engineering, finance and support all using it for customer understanding and account scoring, the company itself pushes him to try it — and that's when they get their opening.
— Alex RampellPure seat-based and pure consumption pricing both crashed
They first tried pure seat-based pricing. Customers accepted it fine, but top users consumed 10,000 times more than the tail, which doesn't work long term. Then they tried pure consumption pricing, turning everything into credits — and within three weeks nobody touched any feature. Keith says those were the company's worst three weeks: lots of signups, no usage. Finally they split it into four kinds of work: everyday CRM work (logging meetings, filling fields, updating tasks) goes into the platform fee or seat fee, because customers don't want to budget annually for "the margin of error on core CRM"; pipeline generation is charged on consumption, because there's alpha there that turns into revenue; workflow automation is charged separately — say someone requests a demo on your website and Lightfield decides this is a deep tech company that should go to Henry and that one is health tech that should go to Matt; and finally intelligence and prediction, which Keith considers Lightfield's most under-explored part, where customers will pay for the alpha.
— Keith PerezNo swim lanes — whoever's free takes the problem
Keith says Tome's biggest lesson was that "a lot of people were playing house": the product lead, marketing lead and CS lead each had their own swim lane, and got offended if you commented on theirs. The result was a company that crawled, and that fundamentally couldn't pivot — a bunch of appendages not talking to the brain. Lightfield is the opposite: 40 people, no swim lanes, everyone owns both product and customer success, everyone joins the same standup every morning, they rank the most important problems — delivery, engineering, CS — and whoever's free takes it. They do continuous planning, with a list that can change daily and gets reassessed weekly. Because in this era anyone can get up to speed on a customer quickly through Lightfield, can ramp on a design system library quickly with an LLM, can auto-create tasks through Lightfield's Linear connection — so engineers, designers and CSMs all lead projects, and specialists just do more projects of the same kind.
— Keith PerezThe bar to start something is low; the bar to ship it is high
How do you stay aligned when everyone's a generalist? Keith's answer is weekly roadmap and GTM meetings, and constant editing: a customer asked for this — how does it relate to the mission? It doesn't, so we don't do it. He describes it as continuous editing to keep what the company puts out coherent. Paired with that are two bars: the bar to start something at Lightfield is very low, but the bar to ship is very high — they still hold a company-wide bug bash, something Keith learned at Instagram: before a product goes to customers, the company itself has to like it first.
— Keith PerezWhat makes him paranoid is speed, not competition
Keith says what worries him most is speed. He mentions reading a report about an a16z portfolio company that had been using some startup's CRM and, because that startup couldn't build the dashboard they wanted for four months, eventually moved to Salesforce. That's what makes him most paranoid: Lightfield has a great wedge — being the best CRM for new companies — but they have to build everything, so those people never get the idea of going back to the old world.
— Keith PerezIn a red ocean you can only be an expansion company
Facing a world where anyone can prompt a product into existence, Keith says one skill Lightfield developed is reading an account's expansion potential. As a CRM company in a red ocean, they have to be an expansion company: in a fiercely competitive market you often don't get the initial land you expected, but by year three or year five that doesn't matter — you become an important part of the customer's company. So they prioritize by an account's three-year expansion value, favoring building for the fastest-growing customers rather than the average one, while admitting that at this stage you have to be "a bit more extreme than usual" and build everything.
— Keith PerezSilicon Valley is for getting logos, not for making money
Keith says most systems-of-record companies' eventual revenue scale actually comes from "more than 50 miles away from here." He sees the Silicon Valley phase as a technique for getting reference customers: some customers raised $200 million but have only three go-to-market people, yet they'll be huge later, so you do the work to perfection now and then take their logos to the rest of the world — "we have healthcare, we have fintech" — and then say you also serve a lot of great manufacturing companies and can now do the same for them. Alex adds the absurdity of the phenomenon: you store exactly the same thing, but because I sell toothbrushes and you sell Coca-Cola, I need to see whether my peers use it.
— Keith PerezBig companies won't build their own system, but they want to build their own company brain
On DIY, Keith says it's slightly overstated: when the ICP was still seed-stage founders, they'd often hear "I could pay you, or I could build it myself over four weekends," and their response was "good luck, call us in five weekends" — and the person would come back saying the hallucinations, the mis-sent emails, everything was wrong. But at bigger companies they don't hear "I want to build my own system of record"; they hear "I want to build my own company brain," and they hear it a lot. Many of those people come back later saying building a company brain or a business world model is genuinely hard, and the hardest part is probably modeling customers — "we tried it, we didn't like our own results, so now we're coming to you."
— Keith PerezWhat excites him most is being the company's crystal ball
Keith says what excites him most is Lightfield becoming the crystal ball companies use for scenario planning: how many salespeople to hire, what product to build next, which way to go. He gives a customer example — a company selling to enterprises discovered in Lightfield that it needed a mid-market product, and built an entire new product line. Joe connects this back to Alex's point about "extremism": in an era when you can build anything, the answer is instead to have a tool like Lightfield and throw frontier intelligence at a genuinely complex decision — something that used to require very smart people, a lot of ops, SQL and a pile of other things, and now takes an afternoon or a weekend of talking to Lightfield.
— Keith PerezWhen you pivot, almost all the noise around you doesn't matter
Asked what advice he'd give his past self, Keith says the most important thing is: when you pivot, almost all the noise around you doesn't matter. You just need to find the pain point, be motivated to build a product or service that solves it, and then focus on customers almost obsessively — everything else is noise. He recalls the things he heard at the time — the office reminds people of the good old days, the food isn't inspiring enough, how will my options be repriced — and says none of it really mattered; put on blinders and stay on the core.
— Keith PerezIn their own words · checked verbatim
there are a lot of metrics behind it, but deep down, I think at an instinctual level, none of us like the product.
Keith Perez3:01
sometimes you get to this point and you're like, oh, well, if the technology kind of continues advancing, then it will be good enough. Yeah. But that's also a little bit of a danger. It's called hopium.
Keith Perez4:01
Screw that up. It's over. It's over. If you get the wrong stages, the wrong fields, you can't get the reps to go back in time and fill it out. It's over.
Alex Rampell15:26
Basically, intelligence is greater than schema.
Alex Rampell15:26
You know the expression, man plans and God laughs. Now it's man plans and open AI laughs.
Keith Perez37:13
I think in many ways, your CRM is maybe harder to move off of than your bank.
Keith Perez45:55
almost none of the noise around you matters when you're in a pivot. You just need to find pain. You need to be inspired to build a product or service that solves that pain. And you need to be like maniacally focused on your customers.
Keith Perez49:57
Figures
| Tome monthly active users | 2 million | 3:01 |
| Tome cumulative users | 25 million | 6:04 |
| Lightfield Series A | $47 million, led by a16z | 2:01 |
| Early customers acquired via the free office | 10 startups | 8:11 |
| Lightfield team size | 40 people | 37:13 |
| Consumption gap under pure seat-based pricing | top users were 10,000x the tail | 32:00 |
| Length of the pure consumption pricing experiment | three weeks | 32:00 |
| Length of stealth development rebuilding the CRM from scratch | about four months | 8:11 |
| Old-school SQL field length example | name varchar 25 | 15:26 |
Glossary
- business world model
- A model of a company's reality, built from customer emails, calls and meetings, that AI agents can use directly.
- greenfield / brownfield
- Greenfield means serving brand-new companies not yet locked into existing software; brownfield means an incumbent-held market where customers are already "captive."
- schema
- A predefined set of fields and table structures; Lightfield argues for no schema, inferring from the activity log after the fact.
- hopium
- Keith's term for the self-soothing judgment that "things will be fine once the next generation of models gets stronger."
- ICP
- Ideal customer profile — the type of customer a company should prioritize serving.
How to listen
Founders considering a pivot or a product rebuild, product and pricing leads at AI-native SaaS companies, and investors watching the moats of systems of record.
The show intro and guest introduction from 0:00-2:01 can be skipped; go straight to the main segment at 2:01.