From Tome to Lightfield: A CRM That AI Can Understand

Original source: Why the Next Generation of Enterprise Software Looks Nothing Like Salesforce

Program: a16z Podcast. Guest: Lightfield cofounder and CEO Keith Peiris. Hosts: Joe Schmidt and Alex Rampell. YouTube publication date: 2026-09-16. Duration: 00:52:04.

On 2026-09-09, a16z announced that it was leading Lightfield's forty-seven-million-dollar Series A. It also publishes this interview. This article draws on the conversation and official materials; product results, customer examples, and internal business figures remain attributed to the speaker.

Tome had reached roughly twenty-five million users. Keith Peiris and his team still struggled to identify the people who would find the product indispensable.

The AI presentation tool had, at one point, been adding about two million users a month. Yet Peiris recalls being unable to see a convincing path from impressive generation to something demanding professionals would depend on every day.

The team eventually shrank the company and rebuilt around customer relationship management software. The new product was Lightfield.

An old problem followed them into the new market: how much background does the AI actually have? A presentation depends on the speaker, the audience, and their relationship. Sales, renewals, and product decisions depend on understanding customers and the history behind their behavior.

Lightfield puts a detailed record of customer relationships at the center of the product, then gives AI that record to work from. The choice also shapes its interfaces, pricing, and organization.

Growth did not settle the question of value

In recounting the pivot from Tome, Peiris treats model capability as only part of the problem.

The team considered getting smaller and waiting for better models. But professional presentations rely on context that rarely appears in a prompt: what the presenter knows, what the audience cares about, and what has already passed between them. Stronger general reasoning would not automatically supply that missing information.

They looked for enterprise uses within their existing audience and recruited 12 sales and marketing pilots. The initial offer was help with presentations and proposals. Customers soon asked for research, lead qualification, and opportunities to expand existing accounts.

Connecting the CRM, call recordings, and data warehouse exposed the harder task: reconciling their contents. A conversation might tell one story while the CRM told another. Some crucial background had never been recorded at all.

The first response was a sales assistant. Peiris says people used it daily, but the team struggled to charge enough. Other systems controlled the underlying records, competing assistants were plentiful, and pricing power was weak.

The team then reduced its size and spent roughly 4 months rebuilding a CRM. Finding willing users was difficult. Empty office space became an incentive: startups could work there for free if they used the software.

The first group comprised 10 companies. Features were missing and the product was slow, yet users kept returning and sent Slack feedback roughly every 2 hours. Peiris saw a kind of engagement that felt different from Tome.

Growth had demonstrated willingness to try the product. Repeated dependence gave the team a clearer reason to keep building.

Relationship history comes before fields

Lightfield starts with a timeline of a customer relationship.

Peiris says 3 of its 5 founding members came from Facebook. Drawing on the timeline idea, they began with the sequence of interactions between businesses: initial contact, messages, meetings, exchanged documents, and eventually product usage and payments.

That chronological activity record became the foundation. Familiar CRM fields, deal stages, and tasks could then be updated from it.

If a team later adds a field or changes its customer categories, the history remains available to revisit and extract the relevant information. Capturing that information no longer depends entirely on a salesperson knowing which box to fill in at the time.

Peiris calls this a business world model: a changing representation of customers, events, and the reasons the business changed state. People and agents can work from a shared account of what happened.

There is a practical limit to the promise of configuration-free setup. The team tried fully unstructured storage and found queries too slow. It settled on a semistructured approach, retaining extensive interaction history while organizing customers, contacts, and relationships for retrieval and analysis.

Users can defer decisions about fields; the system still has to organize the data carefully. Peiris describes an initial experience of connecting an inbox and waiting about 5 minutes for the record to assemble. The interview provides no measured guarantee across different data volumes.

Account expansion illustrates the purpose. Deciding where to deepen a relationship requires conversations, support tickets, and product usage to be considered together, followed by comparison across accounts. A customer name and a sales stage supply little of that background.

A relationship model can also connect patients and trials

The interview's Power example shows why custom relationships matter.

As Peiris describes it, Power helps pharmaceutical companies find clinical trial participants while also serving people looking for treatment opportunities. It models both sides in Lightfield and combines the records with information from the FDA and ClinicalTrials.gov to support matching.

The objects extend beyond the familiar company, contact, and sales opportunity. Trials, participants, pharmaceutical companies, and their requirements and connections all need representation.

Peiris says the process helped a person with Alzheimer's find a frontier treatment opportunity within days. The interview supplies no enrollment record, treatment outcome, or evidence of efficacy. The example describes finding and matching information; it does not establish whether a treatment worked.

The product-design lesson is more general: a business model needs room for the relationships that actually exist in the customer's work, without forcing every organization into the same small set of tables.

New companies still inherit old purchasing habits

Lightfield initially pursued new companies, where years of installed software and accumulated processes created less friction.

Peiris acknowledges that the team did not begin with a clear strategy for persuading established CRM customers to switch. He and a colleague contacted startups and YC companies directly, aiming to become a new business's customer system and learn from its use.

He mentions customers that grew from 0 sales representatives to 100. After roughly 6 months of observing fast-growing companies, the team saw a possible opening beyond automated emails and lead scoring: helping management understand customers and steer the business. These are his observations; the customers' growth cannot be attributed to Lightfield.

However, new companies hire sales leaders accustomed to existing software. A founder might embrace a new system, only for an incoming sales executive to demand Salesforce.

One response was company-wide access within plans negotiated through Lightfield's sales team, without requiring an extra paid seat for every colleague. Engineers could understand customers, finance could access information relevant to revenue recognition, and customer success could assess accounts.

That approach supplied more context and made the product useful across the organization. A replacement decision then had to account for how those teams worked. Peiris says this gave Lightfield a better opportunity to demonstrate its value to the incoming sales leader.

Keep the tables and rethink the configuration

Peiris takes a pragmatic approach to how people should use an AI product.

He still runs sales meetings from a spreadsheet view and recognizes that many people want the same dashboard every morning. Lightfield keeps tables and dashboards alongside natural-language interaction and a command-line interface.

Sales automation changes more substantially. Configuring a follow-up sequence traditionally means connecting triggers, variables, and branches in a flowchart. In the workflow he describes, a user discusses the task with an agent, which writes an execution recipe informed by the business record and then runs it.

Familiar ways to inspect the business can remain while the work of configuring automation gets shorter. They do not have to change at the same pace.

Customers can also use MCP or the command-line interface to bring Lightfield data into agents they build themselves. Peiris emphasizes that the data belongs to the customer and that the product must continue earning its place.

He says some customers build their own agent runtime and later return, citing entity recognition, retrieval precision and recall, and response speed. He also describes larger companies attempting an internal knowledge system and discovering that modeling customer relationships is difficult in its own right.

These are vendor accounts of customer experience, without comparative testing in the interview. They nevertheless identify a concrete operating cost: someone has to maintain synchronization, data quality, relationship resolution, and execution workflows. A working demonstration covers only part of that responsibility.

Pricing improved after separating four kinds of work

Lightfield first tried charging exclusively by seat. Customers understood it because familiar software was sold that way.

But Peiris says heavy users could consume about 10,000 times as much as light users. The figure concerns consumption, not revenue or productivity. A headcount-based bill could not comfortably accommodate the difference.

The team then tried credits for everything. Customers became reluctant to use the product. Peiris describes roughly 3 difficult weeks with registrations but little activity.

Conversations with customers led to four categories of work:

• Everyday CRM. Capture meetings, update fields, and maintain tasks. Customers want a predictable budget without calculating the cost of basic recordkeeping.
• Pipeline generation. Research targets, enrich information, and pursue meetings. Paying for additional work is easier to justify.
• Workflow automation. Research an incoming demo request and route it to the appropriate salesperson, for example. The work performed is relatively clear.
• Intelligence and forecasting. Use the business record to examine hiring, sales processes, and possible directions. What customers will value here is still being explored.

At the time of the interview, Lightfield combined a platform fee and seats for core CRM with consumption charges for other work. This describes the pricing logic the team had reached; the product's current materials determine the actual plans.

Charging for completed sales presents another problem. Outcomes depend on the customer's product and market demand. The same prospecting effort can produce very different results for an established product and a startup still searching for a clear position.

Peiris therefore says Lightfield was charging for work performed, without yet being able to consistently tie its fees to the final sales outcome.

Forty people share priorities, with a high bar to ship

Peiris also thinks Tome organized itself like a mature company too early.

Product, marketing, and customer success defended separate areas of responsibility. Giving feedback across those boundaries became difficult, while long-range plans made changing direction harder.

The Lightfield routine he describes is more direct. The company had 40 people, all attending the same daily standup. They ranked the most important problems together, whether those involved engineering, delivery, or customer success. Someone available and able to help would take the next problem.

Planning continued throughout the week, with a weekly reassessment of priorities. Engineers, designers, and customer success staff could each lead projects. Specialisms remained, but job titles were less restrictive boundaries.

Accessible context helped make that possible. Lightfield supplied customer background, model tools could work with design resources in Figma, and its connection to Linear helped create tasks.

Freedom to start work required continued editing. The team reviewed the roadmap each week and checked customer requests against the company's direction. Before shipping, it still held company-wide sessions to try the product and find problems. Peiris wanted the team to approve of the experience before customers received it.

A low threshold to start a project, combined with a high threshold to release it, was his way of keeping speed and coherence together.

The product has to grow with its customers

Asked what worries him most, Peiris answers speed.

A CRM that works for a newly formed company may not support its later management needs. If reports, workflows, or collaboration lag for too long, the customer has a reason to return to an established system.

Lightfield consequently considers an account's expansion potential over roughly 3 years when setting priorities, leaning toward the needs of its fastest-growing customers. The first contract may be small; keeping pace with the business determines the longer relationship.

Silicon Valley customers also serve as references for expansion into other industries. Peiris mentions companies that had raised two hundred million dollars but employed only 3 people in sales and go-to-market work. Their immediate revenue contribution could be limited, while their future value as a reference could be substantial.

Healthcare is one example. He says deals can involve around 50 stakeholders, and credits complex-context handling, early security investment, and customer references with helping sales. Those are his explanations, not an independent certification of security capabilities.

Buyers of a core business system also want to know whether comparable teams can rely on it over time. Alongside product capability, relevant customer references help establish trust.

From recording the past to examining the next decision

Peiris is most excited about scenario planning: how many salespeople to hire, which product to build, and which market to enter.

Near the end of the conversation, he describes an enterprise-focused customer that examined its business record, identified an opportunity for a midmarket product, and started building it. The customer is unnamed and no subsequent business results are disclosed.

That ambition depends on the preceding work. Completeness, conflicting records, and correctly understood relationships all affect the quality of a decision. The interview does not supply a measure of forecast accuracy.

The move from Tome to Lightfield brings a persistent task into view: helping a machine understand a real business requires ongoing collection and organization of context. Generation is one part of that effort. Lightfield's product positioning rests on that judgment.

Peiris's closing advice to founders considering a pivot is concrete: identify customer pain, believe in solving it, and concentrate on those customers. The startups using the free office space, opening the product daily and asking for missing features, had once given him a clearer direction than a large user total.