Keith Peiris built the AI presentation tool Tome to 25 million users in the first year of ChatGPT, then shut the product down and started again with five people and no product.
The usual move when a product outruns the technology is to shrink the team and wait for the next model. Peiris ran that argument and rejected it, because what the model was missing was not reasoning.
"None of us like the product which is kind of a funny thing to say."
He now runs Lightfield, a CRM company that raised a $47 million Series A led by a16z, and he found his second product by giving away free office space to ten startups who agreed to use an unfinished one.
The full interview is covered here so you can skip it. 53 minutes of audio, 18 minutes of reading.
Here are the 15 lessons that matter.
👤 Guest: Keith Peiris, Co-founder and CEO of Lightfield, who previously built the AI presentation tool Tome to 25 million users and then shut it down
🎙️ Hosts: Alex Rampell and Joe Schmidt, Partners at a16z, which led Lightfield's $47 million Series A
📰 Published: 16 September 2026 on YouTube and the a16z podcast feed
🔴 YouTube | 🔗 Episode page | ⏱️ 53 min | ✅ Time saved: 35 min
Key Takeaways
A product with 25 million users was shut down because its founders could not imagine a discerning professional using it
The blocker was context about the presenter and the audience, not model quality, so waiting for a better model would not have fixed it
The hardest part of doing sales work for customers turned out to be reconciling their data, and that became the company
The call recorder and the CRM routinely disagreed about what had happened
A free CRM with people who like it and will not pay for it has no pricing power, because the data is not yours
Ten startups were recruited with free desk space rather than a discount, and they filed feedback every two hours
Consumption-only pricing froze the product for three weeks — signups arrived and nobody touched anything
Under pure seat pricing the heaviest customers consumed 10,000 times what the lightest ones did
Giving every employee a free seat is a retention strategy aimed at the Salesforce-trained sales VP who has not been hired yet
A CRM is harder to switch than a bank, which is why reference customers in the same industry decide the deal
Forty people, one standup, no swim lanes, and a list of problems that anyone free can pick up
1. 2M Users, No Path to Great
Peiris started Tome because he and his co-founders were consumer people who thought large language models would change how people communicate. They had been working on selfie design at Instagram and Messenger, and they moved to the storytelling of ideas.
The product was a presentation generator, and it launched into the ChatGPT moment with no capacity to serve the demand. "We got 2 million users a month. people were lining up when we didn't have enough inference to support them."
The founding premise was narrower than the product became. "We built this company because we wanted to help professionals tell expert stories hard stories."
What they could not picture was the professional use case. Peiris said they could never see a high-quality, discerning presentation maker using the tool in an indispensable way — not for memos, not in investment banking or consulting.
"We just thought that the technology constrained us to being a tool for individuals and students."
2. Context, Not Reasoning
A host pressed on the obvious alternative: models were improving fast, so why not shrink the team and wait for the curve to carry the product. Peiris said they considered exactly that.
They had watched the jump from GPT-3.5 to GPT-4 and knew the leaps were real. The reason waiting would not work was specific rather than general.
"But I think the biggest issue was that the model just didn't have enough context to really understand the presenter, the audience, the relationship between the presenter and the audience and like no amount of general reasoning gets you past that."
The best case, on his reading, was a great one-shot tool for a speculative presentation, which was not a business the founders wanted.
The host named the trap he was avoiding: the assumption that if the technology keeps advancing it will eventually be good enough, which is comfortable to believe and dangerous when a company is burning money.
3. Following the B2B Heat
The next product came out of the old product's user base rather than a strategy document.
"We had 25 million users or something" — and inside that base, the business users were in sales and marketing.
They ran twelve free pilots at large companies, offering to make decks and proposals, and the sales teams immediately asked for other things: research, lead qualification, understanding accounts for expansion.
"I think what makes early stage founders good at this is that you sort of have no priors." Peiris said early-stage founders are willing to ignore the thing they built and follow the heat.
The first request they made of customers was access to what he called the context box — the CRM, the call recorder, the data warehouse. He assumed at the time that meant Salesforce.
The discovery that became the company was that the data did not agree with itself. "Then we realized actually the hardest part of doing this work was actually making sense of all of the data across all of these disparate systems"
He described the call recorder as often holding a different view of reality than the CRM, which made reorganizing the record the most important work rather than a prerequisite to it.
4. Liked It, Wouldn't Pay
Before the CRM there was a go-to-market assistant, and it failed in a specific and instructive way.
"We got people to like it and then we couldn't get anyone to pay for it."
The reason was ownership of the data. Account executives used it every day, but the underlying record belonged to somebody else, and ten other companies were selling the same thing.
So they shrank the team again and rebuilt the CRM from first principles, working for about four months with no users.
"And it turns out no one wants to use your four-month-old CRM."
5. Negative Pricing
The only asset left over from the first company was an office lease they could not get out of, and it became the acquisition channel.
Peiris posted on X and LinkedIn with the offer: "You can sit in our office space if you use our CRM." One of the hosts named the model for what it was — negative pricing — and Peiris agreed.
"So we found 10 startups to come use the product and for some reason they were in it every day."
The signal was the complaining, not the satisfaction. They were angry about everything missing and about how slow it was, and they were sending Slack feedback roughly every two hours.
That is the contrast Peiris draws with Tome: a barely working product people cared enough about to complain about hourly, against a polished one nobody at the company loved.
6. Model First, Work Later
Asked what was broken about the CRM as it exists, Peiris said the team's advantage was not knowing the category well enough to accept its assumptions.
He describes three jobs a CRM does at a growth-stage company: it stops reps forgetting things, it powers low-level automation, and it powers the forecast.
The naive view was that the third job is the only one that matters. "If we can help you completely model your business and your customer reality, then the rest will be easy."
Everything after that, on his framing, is prompts and tool calls rather than a separate product surface.
This put them against the rest of the category. Peiris said the AI CRM billboards around Silicon Valley have all been about doing the work — lead scoring, outbound emails — while his team went at high-fidelity business modeling instead.
The obstacles that followed from that choice were the rep's manual entry and the quality of other systems' interfaces, not the messaging.
It took roughly six months of watching customers before the wedge became clear, and he places it at a higher level than the category expected: better understanding of your own company so you can steer it.
7. The Activity Log Primitive
The architecture came from the founders' previous employer rather than from any CRM.
"Three out of our five founding members came from Facebook." The view they brought was that the most important object in a CRM is the relationship.
So they modeled the Facebook timeline: when you first reached out, what each side said, which meetings happened, which documents went back and forth, then what the customer does in the product and how they pay.
Traditional CRM behavior sits on top of that log rather than beside it. The system updates fields and stages as a consequence of the log. "But you always have this like canonical log of relationship that everything's built on top of."
They tried the purely unstructured version first and it did not work. "So we actually tried going fully unstructured. And we found that the queries just took too long." A host named the failure as the needle-in-a-haystack problem.
The compromise is semi-structured: large volumes of unstructured data stored in the activity log, with the log itself used to infer cause and effect.
The worked example is a customer success question. Asked whether an account is ready for expansion, the system reads what the customer said, what their support tickets look like and how they have been using the product — a month without a login is an answer — then compares accounts across the schema before going deep on any one.
8. Schemaless by Design
Peiris said his team interviewed CRM consultants and wrote down what they actually do for a living.
The answer was the data model — the single most consequential decision a consultant helps a company make.
The cost of getting it wrong is unrecoverable, which is the argument for not making the decision at all. "If you get the wrong stages, the wrong fields, you can't get the reps to go back in time and fill it out."
So Lightfield asks for connections rather than a schema: email, a call recorder, the data warehouse, built-in enrichment, and the relationships assemble in real time. Fields get filled in later, and changing your mind means traversing the log again to refill them.
A host compared the setup to a consumer product — press sync, wait five minutes, and it is there.
Rampell's digression on where schemas came from is the sharpest version of the point. Old relational databases made you predefine how many characters a column could hold, to save space, so a name field might be capped at 25 characters and a longer name would not fit.
The phrase the room settled on was "Intelligence is greater than schema."
9. A Marketplace for Trials
Because the schema is arbitrary, the customers Peiris is proudest of are the ones whose businesses do not fit a sales pipeline.
A company called Power works with pharmaceutical companies to find clinical trial participants, and runs a marketplace on the other side for patients looking for frontier treatment.
Both sides are modeled in the same system, with custom objects and custom relationships. Peiris said the company built automations that scrape the FDA and clinicaltrials.gov to assemble a picture of every trial running anywhere.
The matching then runs between the two. "And Lightfield actually helped someone with Alzheimer's find frontier treatment within days."
10. Hostages, Not Customers
Rampell used the segment to set out the framework a16z applies to startups selling into established categories.
"So, if you think about like startups normally, selling into the brownfield is hard just because it's brown." A brownfield market is one an incumbent has already trampled.
His compressed version of what that means for a buyer: the saying he uses is that "the best companies have hostages not customers", and SAP is the example he gives.
The cloud transition is his example of breaking into a brownfield anyway, by redefining the problem — the CRM running on the mainframe in your office, the person who maintained it has left, so move to a cloud-based vendor instead.
"I'm just going to build the best product in the world and then brand new companies that are untethered from any existing software solution." That is the greenfield alternative, and it is where Lightfield started.
Peiris was blunt about why. They did not have a sharp enough thesis for attacking the incumbent, so they went to win new companies — LinkedIn prospecting and cold emails to Y Combinator companies, run by Peiris and his chief of staff.
The bet is that a fast-growing customer becomes a large customer inside the same contract. He said Lightfield has customers who joined with zero reps and now have a hundred.
11. Free Seats as a Moat
Rampell put the objection that kills greenfield software companies: the greenfield company eventually hires a brownfield VP of sales who makes the purchasing decision.
He told it as his own story. "I actually remember when I started one of my first companies, I was so adamant about not paying $85 a month for Salesforce that I used this thing called Sugar CRM, which was free." He gave up not on product grounds but because the VP of sales he hired refused to use it.
Peiris said the design principle was written for exactly that hire. Founders and engineering leaders would adopt frontier technology willingly; the incoming sales VP would not.
So the sales-led plans give the product away to everyone inside the customer's company, not only to the sales team.
The first effect is coverage — the system sees what engineering, support and finance are doing, not only what sales is doing.
The second is the intended one. "The other thing is it'll create sort of real company network effects that make it harder to rip and replace us."
He says it has already played out. A seasoned sales VP arrives, says the product is good but "I only know how to use Salesforce" because of how they were trained, and the rest of the company pushes back, because it is how engineering understands customers, how finance does revenue recognition and how customer success scores accounts.
12. Knobs vs Plain Language
A host asked whether the right move is to reproduce the tables and dashboards people already know or to push them to natural language.
Peiris came down on pragmatism. "I think we had this view of if we really want to be like a real enterprise CRM, we can't be religious about the way people work." He is the acting sales manager at his own company and runs meetings from the spreadsheet view.
He concedes the dashboard argument outright. Lightfield, he said, has not "won the war on deterministic dashboards", and "I think everyone wants to look at the same dashboard every morning"
Where he has reimagined a workflow, it is the sales sequence. The old version made you express five emails, their timing and their triggers as arrows, conditions and variables.
The new version is a conversation. "The agent writes out a recipe for you. The recipe takes into account what's in your world model and it sort of runs with it."
Sales leaders resisted it in the same words each time — they wanted their knobs and switches — and then came around on the grounds that it was more efficient and there was less to learn.
The host's analogy was his new car. He had come from a BMW with nine million buttons; the newest Teslas have no gear stalk at all, and the result is intuitive enough for a five-year-old, which is the same trade as schema against intelligence.
13. Seats, Then Consumption
Lightfield arrived at its pricing by running both extremes and finding neither worked.
Pure seat pricing matched the category and was well received, and it was unsustainable. The heaviest users consumed 10,000 times what the lightest did.
Pure consumption pricing failed in the opposite direction. Everything became a credit and nobody spent them. "You could imagine it was the worst three weeks of the company's life"
The resolution was to split the work into four buckets. Everyday CRM work — capturing a meeting, filling out records, updating tasks — customers expect covered by a fixed fee, because nobody wants error bounds on the core system in next year's budget.
Pipeline generation is where customers accept consumption pricing, because enrichment produces meetings that convert to revenue.
Workflow automation is the third, and Peiris's example is inbound routing: someone requests a demo, the system researches them and sends the deep-tech company to one rep and the health-tech company to another.
The fourth is forecasting and scenario planning, which he calls the most undiscovered part of the product and the place customers will pay for the analytical edge.
The landing point is a platform fee plus a seat charge for the core CRM, with consumption pricing on everything else, and Peiris said it has worked.
On outcome-based pricing he says the category cannot support it. "I think the hardest thing about being a sales company is that our outcomes are sort of dependent on the strength of your product market fit" — prospecting for OpenAI would be efficient, prospecting for a seed-stage startup with no website would not. "we have to charge for the work."
14. No Swim Lanes, 40 People
Asked about velocity, Peiris named the thing he thinks slowed the first company down.
His co-founder Henry's line for it is that Tome had "a lot of people playing house" — a product leader, a marketing leader and a customer success leader, each with a swim lane and each annoyed by feedback from outside it.
His diagnosis is that the structure made the company impossible to pivot, because the appendages were not talking to the brain. Planning too far ahead compounded it.
The replacement is a single list. "Everyone shows up to the same standup every morning. We stack rank the most important problems." Whoever is free takes the next one, the list can change daily, and it is reassessed weekly.
Everyone owns product and everyone owns customer success, and engineers, designers and customer success managers all run projects.
What makes generalists possible is tooling rather than hiring. Anyone can get up to speed on a customer through Lightfield, on the design system because the model can reach Figma, and on tasks because Lightfield connects to Linear.
The check on the resulting sprawl is a high bar at the other end. Starting a project is cheap; shipping one is not. "We still do company bug bashes. It was something I learned at Instagram when I was there. The company needs to like this before it goes to customers."
Prioritization is settled on expansion value rather than revenue today. Peiris said Lightfield weighs an account's likely growth over a three-year horizon and has leaned toward building for its fastest-growing customers rather than its average one.
15. Logos, Not Revenue
Peiris does not expect Silicon Valley to be where the money is, and said so plainly.
"As a CRM company in a red ocean space, we have to be an expansion company." In a crowded category you rarely win the land you wanted, so the case has to be made on what the account is worth in year three or year five.
The startup customers are a marketing asset. "I've always thought about this moment in company building as a trick to get reference logos."
The logos are worth having because the customers are real. Lightfield has "customers that have raised like $200 million" with three people on their go-to-market team, and Peiris expects them to be large companies.
A host contrasted Silicon Valley's overhype with the rest of the world's underhype — people who tried ChatGPT once in November 2022, watched it hallucinate, and put it away, and who now expect it to take their jobs.
Peiris's answer to why industry reference customers matter so much is switching cost. "I think in many ways your CRM is maybe harder to move off of than your bank."
Health technology is where that has compounded for Lightfield. He credits complex deals with many stakeholders, context engineering, and security work done early — penetration testing and signed agreements — and says winning a health-technology deal has become easy on the strength of the customers he can now point to.
On customers building it themselves, he says the threat is overstated. "I think it's a little overtalked about." Seed-stage founders said they could do it over four weekends and came back when it hallucinated and sent bad emails.
The larger-company version is different and more serious. "We hear I'm going to build my own company brain." Those customers come back too, he said, having found that building a company brain or a business world model is harder than it looks: "And I think maybe the hardest part of it is modeling the customers."
Where customers build a harness on top of Lightfield through its command line or MCP interface, Peiris says they underestimate what the existing one does — entity recognition, precision and recall, speed. "But after a few weeks, they almost always realize, wait a minute, your harness was actually doing quite a bit."
Bonus Insights
What worries Peiris is not a competitor but his own speed. He described reading accounts of companies leaving a newer CRM for Salesforce, including an a16z portfolio company that waited about four months for dashboards and then moved. That is "the thing that makes me the most paranoid", and the conclusion he draws is that Lightfield has to build everything: "So that those folks never feel the desire to go to the old world."
The feature he is most excited about is not a sales feature. He wants Lightfield used to decide how many reps to hire and what to build next, and cited a customer selling to enterprises who worked out through Lightfield that it needed a mid-market product and built a new product line on the strength of it.
Schmidt's framing of that is the investor's case for the category: the data and context are already assembled, so this is the moment to "throw frontier intelligence at a really complicated decision" — work that used to require analysts, operations staff and SQL.
The exchange updated an old saying for the era. "Now it's man plans and OpenAI lives."
Peiris's advice to a founder mid-pivot is to stop listening to almost everything. "I think the most important thing to remember is that almost none of the noise around you matters when you're in a pivot. You just need to find pain." He listed what the noise sounded like at Tome — the office reminding people of the good old days, uninspiring food, questions about how options would be repriced — and said none of it mattered. "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."
Peiris's bottom line is that the defensible thing in an AI-era application company is not the work the agent does but the record it works from: build a model of the customer's business that is accurate enough to answer open questions, and the automation, the dashboards and the pricing arrange themselves around it.
Products, Companies & Tools Mentioned
Lightfield (Peiris's CRM, built on an activity log rather than a schema, and the subject of the interview)
Tome (His previous company: an AI presentation tool that reached 25 million users and was shut down because the founders could not see professionals using it)
Salesforce and HubSpot (The pricing model Lightfield copied first, and the incumbent a newly hired sales VP insists on)
Sugar CRM (Rampell's own attempt to avoid paying $85 a month for Salesforce, abandoned when his sales VP refused to use it)
SAP (Rampell's example of a company with hostages rather than customers)
Power (The customer that models a clinical-trial marketplace on both sides in Lightfield and matched an Alzheimer's patient to frontier treatment within days)
ElevenLabs (The a16z portfolio company that left a newer CRM for Salesforce after waiting months for dashboards — the story Peiris says makes him most paranoid)
Facebook and Instagram (Where three of Lightfield's five founding members came from; the timeline is the model for the activity log and the bug bash is a practice Peiris took from Instagram)
OpenAI (Named both as the outbound-prospecting customer any sales tool would want and in the line about plans)
Figma and Linear (The tools that let a generalist pick up a design or a task without a specialist)
Tesla and BMW (A host's contrast between nine million buttons and a car with no gear stalk, used as the consumer version of schema against intelligence)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: