ExpertsFMEric Glyman, Co-founder of Ramp
David Senra · July 12, 2026
5 Selected Lessons:
framework · 3 reasoning_pattern · 1 heuristic · 1
Elon's algorithm applied to internal processes
Source
David Senra — Eric Glyman, Co-founder of Ramp (July 12, 2026) · at 6:32 ▸
When this applies
when building or reviewing any internal process, feature, or workflow in any company
The takeaway
Before optimizing or automating a process, first ask whose requirement it is, whether it's needed at all, and simplify before you speed up or automate.
Why it works
Sequentially questioning requirements, necessity of steps, simplification, speed, and automation prevents unnecessary complexity from calcifying into permanent process.
One level up
Automation applied to an unexamined process just makes the waste move faster; simplification must precede acceleration.
Where it stops holding
This stops working when the process's necessity and origin are already well understood and verified — repeating the questioning sequence then just wastes time.
Detail
procedure:
- Ask whose requirement this is in the first place.
- Ask what you're actually trying to achieve.
- Ask if every step is needed.
- Simplify the process.
- Cut unnecessary parts.
- Accelerate development time.
- Automate the simplified process.
benchmarks: {}
rationale: Applying this to every dollar/hour decision keeps products as 'scaffolding' for a service outcome rather than feature bloat.
In their words
whose requirement is this in the first place? What is it that we're trying to achieve? Do we need every single step? Is there a way to simplify the process? Can we cut out unnecessary parts?
Build and broadcast a single customer-outcome scoreboard
Source
David Senra — Eric Glyman, Co-founder of Ramp (July 12, 2026) · at 31:46 ▸ · at 35:18 ▸
When this applies
when any company wants to prevent internal incentives from drifting away from customer value as it scales
The takeaway
Reduce your guiding values to a small number that are actually measurable, instrument the business to track them continuously, and display them visibly and repeatedly company-wide so they can't be forgotten or gamed in private.
Why it works
Explicitly measuring and continuously displaying a small number of customer-outcome metrics (dollars saved, hours saved) forces every new feature decision to trace back to a verifiable value claim, and public visibility prevents the metric from being quietly ignored.
One level up
A value only functions as a value if it is instrumented, visible, and inescapable — otherwise it is merely a poster.
Where it stops holding
This stops working when the chosen metric doesn't actually capture customer value (metric gaming) or when the company has too many competing top-level metrics, diluting focus (the '20 elementary school values' problem).
Detail
procedure:
- Pick only the handful of metrics everyone genuinely agrees matter (not an exhaustive values list).
- Build systems (e.g., accounting software integration) to measure them as a source of truth.
- Review the scoreboard on a fixed cadence (monthly at Ramp).
- Make the scoreboard visually inescapable (dashboards, Slack channels, physical wall displays, first slide of sales pitches).
benchmarks: {"ramp_example":"median customer cuts expenses 5%+/year; median customer revenue growth 16% vs US average 3-4%."}
rationale: Constant, visible exposure to the real scoreboard — as with Citadel's risk dashboard — keeps an organization from drifting off its stated purpose.
In their words
Did we save, in aggregate, customers more money? Did we save them more time? And for individual customers, did the experience get better or worse?
Identify and manage emerging expense categories before they become line items
Source
David Senra — Eric Glyman, Co-founder of Ramp (July 12, 2026) · at 43:49 ▸ · at 47:27 ▸
When this applies
when any new category of spend (new technology, new vendor type) emerges faster than existing budgeting/governance processes can track it
The takeaway
When a new cost category with real marginal costs emerges, build attribution (who spent it, on what, with what return) early, rather than waiting for it to become a conventional budget line.
Why it works
A spend category with real marginal cost (unlike flat-fee software) and no established budget process is invisible until it's already overrun; building tracking/attribution (by purpose, owner, ROI) before the category matures prevents waste and informs allocation decisions.
One level up
Financial control systems must be built ahead of the spend curve for genuinely new cost categories, because by the time a category is 'normal,' the waste has already compounded.
Where it stops holding
This stops working when the new spend category is small enough that instrumenting it costs more than the waste it would catch.
Detail
procedure:
- Recognize the new spend category (AI token spend) has real marginal costs unlike typical software.
- Track who spent, on what (OpEx vs R&D), and what the return was.
- Monitor model substitution opportunities — about six months after a frontier model launches, an open-weight model often matches it at roughly 1/100th cost.
- Build routing/wrapping systems to send tasks to the cheapest adequate model rather than always using the frontier model.
benchmarks: {"anthropic_run_rate":"over $50B/year run rate, cited as rumor","labs_combined_projection":"labs projected to pass $300B/year, ~1% of US GDP","model_cost_decay":"~6 months after a frontier model, an open-weight model of similar quality may cost ~1/100th"}
rationale: No company budgeted for this type of spend five years ago; building the control layer now captures the gap before it becomes standardized.
In their words
I think there's going to be a third mega category created around how do we manage knowledge work that will be done on behalf of organizations and done through these tokens.
AI turns determined generalists into cross-domain specialists
Source
David Senra — Eric Glyman, Co-founder of Ramp (July 12, 2026) · at 10:12 ▸
When this applies
when evaluating whether to keep narrow functional specialists or empower generalists with AI tools across any knowledge-work organization
The takeaway
When a shared intelligence layer exceeds any individual's domain knowledge, drive and willingness to push through unfamiliar territory becomes more valuable than possessing the credential/skill itself.
Why it works
LLMs have ingested more domain-specific text (code, medical cases, financial filings) than any single human expert, so a sufficiently determined non-expert can use them to approximate expert-level output in an adjacent domain, removing the hard boundary between crafts.
One level up
Tools that democratize access to expertise shift the scarce resource from skill to determination.
Where it stops holding
This stops working when tasks require physical skill, regulatory licensure, or tacit judgment that LLMs cannot substitute for, or when the generalist lacks the stubbornness to iterate past initial failures.
Detail
approach: Treat AI as an extension of your own capability boundary rather than a tool you hand off to specialists.
signals:
- Someone keeps pushing past the edge of their formal expertise using AI assistance.
- Organizational silos form when people identify by craft rather than mission.
- Cross-functional tools (AI-assisted sales pitches, tracking, outreach) blur traditional role boundaries.
rationale: If the organization's shape assumes specialization is required, AI's generalist-empowerment effect invalidates that assumption and should trigger a redesign.
In their words
And now if you stubborn enough and pushing and willing to start to go build systems, you can extend your boundaries.
Hire for proof of work and proof of drive, not resume
Source
David Senra — Eric Glyman, Co-founder of Ramp (July 12, 2026) · at 14:44 ▸
When this applies
when hiring for any role where raw capability matters more than formal credentials
The takeaway
Look for independently generated evidence of exceptional drive/spike in ANY domain (even unrelated ones like gaming) rather than filtering by pedigree.
Why it works
Credentials are a lagging, noisy proxy for ability; a demonstrated body of self-directed work (built something, won something, led a niche community) is direct evidence of the exact trait — obsessive focus and initiative — that predicts future high performance.
One level up
The signal you actually need (capacity for obsessive execution) can be observed outside the context you need it for, so don't limit your search to people who exhibited it in your industry.
Where it stops holding
This stops working when a role genuinely requires certified/regulated expertise (e.g., licensed professions) where proof-of-work can't substitute for credentials.
Detail
rule: Prioritize candidates with visible proof of obsessive, self-directed achievement (any domain) and genuine motivational alignment with the mission, over traditional resume signals.
threshold: Evidence should show top ~0.1-1% level focus/output in some field, sustained over significant time (the host's Founders podcast example: years of consistent work).
rationale: Two business days of actually working with someone reveals more than 15 hours of interviews; referrals and alternative communities (GitHub, niche hobbyist scenes) surface asymmetric information about true capability.
In their words
I'm less interested in what is the resume. I'm far more interested in, like, proof of work.