a16zWhy the Next Generation of Enterprise Software Looks Nothing Like Salesforce
At a glance
WHAT IT’S REALLY ABOUT
Lightfield’s AI-native CRM rebuild: timeline-based world models over rigid fields
- Keith Peiris describes pivoting from Tome (AI presentations) to Lightfield after realizing the presentation product couldn’t become indispensable without deeper user-and-audience context.
- Lightfield repositioned CRM from a rep-maintained repository into a “business world model” that unifies emails, calls, meetings, and product/warehouse data into a machine-usable record for agents and humans.
- The product’s core architectural choice is a canonical, semi-structured activity log of the customer relationship, from which traditional CRM fields, stages, and insights can be inferred and backfilled.
- Go-to-market started greenfield with fast-growing startups to iterate quickly and find a brownfield wedge, while driving company-wide adoption to overcome Salesforce-trained buyer inertia.
- Lightfield evolved pricing from failed extremes (seat-only vs consumption-only) into a hybrid model: fixed pricing for core CRM reliability and consumption for pipeline, automations, and forecasting intelligence.
IDEAS WORTH REMEMBERING
5 ideasIf the missing ingredient is context, wait-less: rebuild around context capture, not model improvements.
Lightfield’s team concluded that no amount of better general reasoning would make AI presentations reliably “indispensable” because the model lacked key context (presenter, audience, relationship). That pushed them toward a domain where capturing and structuring context is the product, not a nice-to-have.
The AI wedge in CRM isn’t “do more tasks,” it’s “make a coherent business world model.”
In pilots with sales and marketing teams, the biggest bottleneck wasn’t generating content or automations—it was reconciling incomplete, conflicting data across CRM, call recorders, email, and warehouses. Lightfield reframed the core problem as reorganizing business reality into a machine-usable model.
Make the relationship timeline the primitive; derive the CRM schema from it.
Instead of starting with objects/fields, Lightfield built a canonical chronological activity log (inspired by the Facebook timeline) capturing every interaction and artifact in the relationship. Traditional CRM fields/stages are then inferred/derived from that log rather than being the source of truth.
Pure unstructured data fails at speed; semi-structured logs balance fidelity with performance.
They tested fully unstructured storage but found querying too slow (“needle in a haystack”). The resulting semi-structured approach stores rich unstructured events in the log while enabling fast traversal and structured comparisons across accounts.
Schema should be revisable; backfill should be a first-class capability.
Lightfield reduced the “data model is destiny” setup risk by allowing customers to connect systems (email, call recorder, warehouse) and assemble relationships first, then define/modify fields later—and even refill fields retrospectively by replaying the activity log. This turns CRM setup from a fragile upfront decision into an iterative process.
WORDS WORTH SAVING
5 quotesDeep down, I think at an instinctual level, um, none of us liked the product.
— Keith Peiris
I think at the end of the day, if you're a, a, a founder, you have to, like, love the product that you're building, and you have to be excited for your customers to use it.
— Keith Peiris
If you can reorganize reality for a company in a way that machines can understand, and also so for humans to understand, um, that feels like a way more interesting and enduring company than the one that we're on right now.
— Keith Peiris
We stack rank the most important problems. Some of them are delivery, some of them are engineering, some of them are CS. And then whoever's free just takes them.
— Keith Peiris
I think the m- the most important thing to, to remember is that, uh, almost none of the noise around you matters when you're in a pivot.
— Keith Peiris
High quality AI-generated summary created from speaker-labeled transcript.
