Summary

Mukherjee expects AI-enabled sales organizations could shrink to 10–20% of today’s size, though he concedes the same-revenue outcome still “has to be proven.” The surviving sellers will manage two to three times more work, while management layers compress and sales-and-marketing OPEX could fall from roughly 40% to 20%. Strategic humans remain because enterprise buyers still value someone who will “get on planes” and build non-transactional relationships.
AI will intensify an existing sales power law in which the top 10% of sellers generate 90% of revenue. The best reps will use LLMs as “the new computer” to research faster, support larger books, and widen their lead; everyone else must retrain or risk irrelevance. Mukherjee’s blunt diagnosis is that easy spending created order takers “shooting fish in a barrel,” while winning difficult deals remains “hard, hard work.”
Consumption pricing turns signed contract value into “just words” unless customers actually use the product. Mukherjee favors the Frank Slootman-style model in which quota carriers own both commitment and consumption. In the alternative pod model, account managers are explicitly paid on renewals and expansion, unlike traditional customer-success teams optimized around onboarding and CSAT. The investor implication is tighter labor accountability, but also a harder incentive-design problem.
Rox deliberately avoids experimental AI budgets in favor of measurable operating—or people—budgets. It runs a 30-day, paid, white-glove evaluation against dollar-normalized outcomes, adoption, and activity; it prefers pipeline progressed and reports 75% weekly active usage and more than 50% daily actives. “If not, you can fire Rox” is the commercial posture, because probabilistic products that lose user trust may face retention and churn problems.
Focus, rather than channel proliferation, powered New Relic’s self-serve revival. Mukherjee concentrated on a latent audience of 100,000-plus developers, treated free usage as consideration rather than conversion, and cut channels where New Relic could not be world-class. His founder version is equally sharp: “We don’t do any free design partners”—customers need financial skin in the game, while the product should remain open enough to exploit incumbents’ three-to-nine-month time to value.
Rox’s product thesis is that the scarce asset is complete customer context, not generic text generation. Its agents combine CRM, warehouse, and public data to research and prioritize accounts, compose outreach, and prepare meetings—but the company is deliberately “the anti-AI SDR”: humans review and send. Infinite automated supply raises the noise floor, while enterprise selling still takes “13 to 15 touches” across email, LinkedIn, meetings, and travel.
The startup-versus-incumbent race comes down to whether incumbents gain speed before startups gain distribution. Mukherjee’s answer is compressed time to value: target a low-six-figure upper-mid-market enterprise land within 45 days of the first meeting, prove ROI during a 30-day paid evaluation, then expand toward high six or seven figures. That durable land-and-expand trajectory matters more to him than matching AI companies racing to $40 million, $50 million, or $100 million in a year.

Deep dive

1. PLG works when one audience and one operator control the system

Mukherjee’s central lesson from New Relic’s zero-to-$100 million self-serve journey is “focus.” Roughly 100,000-plus developers who entered the workforce from 2008 to 2014 still regarded New Relic as the default brand; concentrating investment on winning them back felt like “drilling into an oil reserve.”
His operating model separates bits from atoms: let the product deliver the “product yes,” then use humans for the “commercial yes.” Running PLG and enterprise simultaneously is “very hard” without one C-suite owner who understands engineering, product, sales, services, and support well enough to prevent siloed funnels.
Sales therefore does not disappear at the top end. Large buyers “don’t give a F about products or services”; they want pain removed and a trusted vendor relationship. The highest-value sellers jointly shape product roadmaps and multi-year commercial structures instead of merely getting an order form signed.

2. Distribution economics reward concentrated channels, not activity

Mukherjee decomposed New Relic’s funnel into awareness, consideration, and conversion, then found organic content and performance marketing were the strongest awareness channels. Events and DevRel consumed resources without being areas where New Relic could outperform, so he shut event marketing down.
The host’s counterexample—Ramp surrounding conferences with hotel key cards, billboards, and multiple coordinated touchpoints—wins a concession: events can work as a system. Mukherjee’s preferred version is still narrower: skip being one of 100 sponsors and host “an amazing baller dinner” that creates two or three deep relationships.
A free user is not a converted customer; “free-tier product is just a better website.” Product use belongs in consideration, where prospects tinker instead of reading marketing pages, and the company must still invest deliberately in moving them to paid conversion.
For SEO, the top 10–15% of pages can generate 90% of value, so produce fewer, higher-quality pieces and distribute them well. SEM is “a complete cartel” except where 10–20 tightly defined keywords create a stitched path from intent to ad, landing page, product use, and an aha moment—with returns visible within a month or quarter.

3. Paid pilots and open products expose genuine demand

“We don’t do any free design partners” is Mukherjee’s categorical position. Without even modest cash at risk, customers can agree cheaply while founders spend weeks or quarters serving them; paid alpha or beta partners create enough skin in the game to sustain real collaboration.
In parallel, the product should remain open, particularly against incumbents whose demo and implementation gates produce three-, six-, or nine-month time to value. Shipping publicly while embarrassed expresses confidence that the team can “ship your way out of holes” and removes friction from the sales motion.
Mukherjee initially thought Datadog beat New Relic through product, but changed his mind after competing directly: the differentiator was six to seven years of consistent go-to-market execution. Stable segmentation and compensation let talent progress from inside sales to elite enterprise roles while institutional knowledge compounded.

4. Consumption revenue makes ownership and incentives unavoidable

On the “backside of COVID,” Mukherjee concluded that the game had shifted from committed revenue to usage or consumption. Contract values became “just words”; realizing revenue required adoption, forcing companies to change leadership, retrain frontline teams, and rebuild underlying systems.
The Slootman school says invest in bag carriers and make AEs own both committed and consumed revenue: “You eat what you hunt.” Mukherjee regards that intensity as directionally right because commission creates skin in the game and an owner mentality closer to how founders and CEOs think.
The host’s pushback—one AE cannot chase megadeals while building champions and wall-to-wall adoption—surfaces the account-manager role. Unlike traditional CS roles centered on onboarding checklists and CSAT, AMs hold explicit numbers for renewals and expansion and are compensated on upsell.
The alternative is a five- or six-person pod containing an AE, AM, CS, sales engineers, and others, with only one true hunter. It can work under exceptional leadership, but Mukherjee invokes Charlie Munger’s incentive lesson: aligning that many fixed-salary contributors around customer outcomes and consumption is extremely difficult.

5. AI magnifies the power law—and exposes a softened workforce

“Sales is a power law. The top 10% bring in 90% of the revenue.” Mukherjee expects AI to make that distribution more acute as president’s-club sellers, elite sales leaders, and strong CEOs exploit information arbitrage and improve faster than peers.
The host worries this creates a two-tier culture inside the same team. Mukherjee agrees the distribution will become more bimodal; management must retrain people for a fundamentally changed job, help them evolve, and accept that failure to adapt can become “an existential decision.”
The past four to five years of abundant spending made too many sellers order takers “shooting fish in a barrel.” Selling something new or large still demands deep customer understanding, and Mukherjee says it took him two to three months longer than expected to find the outlier reps Rox needed.

6. AI can recover a quarter or more of sales-ramp time

A rep traditionally needed two to two-and-a-half quarters to ramp. With ChatGPT, Perplexity, or Rox compressing research on products, customers, 10-Ks, 10-Qs, SEC filings, and executive transcripts, Mukherjee believes a capable seller can gain a quarter to a quarter-and-a-half.
The immediate management move is mundane but consequential: let employees use the tools they already use privately and solve professional data-access barriers. Then replace two or three months of stale classroom training with AI-enabled learning, live deals, and supervised customer exposure.
The host objects that customers are precious and unprepared reps can damage the company. Mukherjee’s early-stage answer is mentorship: after a few weeks, new hires should shadow a ramped AE or the founder because frontline expectations change too quickly for simulated training.
His example happened within weeks: Rox’s deep research initially wowed customers, but after Deep Research launched, buyers began expecting “PhD-level research.” Strong hires detect such changes by asking about buyer psychology, expectations, pain, and propensity to purchase—not merely reciting standard discovery questions.

7. Complete context is the moat; agents are the leverage

Rox runs alongside Salesforce, HubSpot, or Dynamics and inside the customer’s warehouse, indexing internal context plus public information. The aim is to give each seller a “bat suit”: real-time knowledge of usage, support tickets, buying intent, growth, and prior interactions that no CRM record alone contains.
Its current jobs are customer research and prioritization, allowing companies to double or triple books until sellers manage tens or hundreds of accounts. It then composes outreach and handles pre-meeting research, in-meeting support, and follow-ups—the equivalent of giving every knowledge worker a presidential briefing team.
Mukherjee calls Rox “the anti-AI SDR.” Agents decide whom to contact and draft sequences, but humans review and send because autonomously filling the pipe could dilute the company’s brand. Authenticity requires the rep’s address, voice, judgment, and willingness to build a real connection.
Infinite outbound supply does make outbound less effective. One email will not close an enterprise deal; Mukherjee says it commonly takes 13–15 touches through email, LinkedIn, meetings, and travel. The agents produce the material, but “you have to really put in the work.”

8. Workflow interfaces beat blank chat boxes for daily work

Rox’s first three to four months produced an agent that knew everything about every customer behind a chat interface. Mukherjee’s verdict is blunt: “This just does not work.” Users wanted completed work laid out for them, not another blank box requiring them to formulate every request.
The product moved from chat to suggested prompts, output panels, and ultimately job-specific workflows. Account reports and org charts arrive pre-created; emails are queued as tasks; meetings, outreach, and account research occupy separate surfaces, while chat remains available for deeper inspection.
Adoption changed “insanely” once Rox decomposed the interface around jobs to be done: usage expanded from a few people to the whole company quickly. Mukherjee does not claim this is the terminal UI state; the backend context architecture is opinionated, while the interface must keep tracking changing consumer expectations.
Relevance remains difficult. A company earnings release may be interesting without creating a compelling reason to contact anyone; the host argues a buyer’s personal LinkedIn post can be more actionable. Rox is consequently moving from company-level signals toward contact- and human-level insight.

9. Retention will separate operating systems from AI experiments

Rox’s 30-day, paid, white-glove evaluation tracks three outcomes with the CFO, CRO, and C-suite: a dollar-normalized result, adoption, and activity. Pipeline progressed is preferred over pipeline generated because agents should let one human manage a larger book while maintaining engagement quality.
Mukherjee reports 75% weekly active usage and more than 50% daily actives, with an iOS app and Slack supporting the ambition to become a daily driver. If people do not use it or the metrics fail, “you can remove us.”
Churn is the issue that “keeps us up at night” because agent outputs are probabilistic. If two users receive different Siemens research and one result is poor, each failure shaves away trust until the user stops returning; Mukherjee expects many enterprise AI tools to suffer exactly that retention problem.
Although experimental AI budgets remain massive, Rox deliberately avoids them and sells against operating and increasingly people budgets. Only a small set of operationally aggressive CEOs, COOs, and CFOs currently sees an existential need to transform this year; most enterprises remain early in the adoption curve.

10. Sales headcount could collapse before relationship selling does

Mukherjee says sales organizations could become 10–20% of their current size, calling it a probable—not proven—productivity path. The humans retained will handle strategic, high-ACV relationships and two to three times more work: a five-account book might become seven this year and 12 next year.
Quotas and organizational design must then change together. Fewer sellers holding larger books imply compressed management layers, while sales and marketing’s roughly 40% share of OPEX could potentially fall to 20%.
Roughly half of Rox’s customers are private-equity-owned, while 50% of inbound is portfolio-wide rollouts, so efficiency is plainly part of the demand. Mukherjee hopes savings also raise compensation for supercharged frontline sellers and fund R&D, which he considers underinvested across many scaled technology companies.

11. Data moats survive where the internet is not truly open

The host asks whether agent swarms kill ZoomInfo-style providers. Mukherjee’s distinction is between public company information, where agents are approaching parity, and contact data accumulated and verified over 10–15 years through software plus operational rigor.
Rox currently enriches and verifies information customers already possess rather than recreating every proprietary dataset. LinkedIn remains a valuable walled garden, so providers that sell such contact data occupy a market Rox does not intend to enter.
Incumbents possess enormous distribution and data, but implementations can take quarters and require systems integrators. Rox’s counter is a product that gets started today and offers compressed proof of ROI—an answer to Alex Rampell’s question of whether “the incumbent acquire[s] speed before the startup acquires distribution.”
As former point solutions bundle and every incumbent repositions around AI or agents, Mukherjee wants Rox to become the default AI-native bundled provider. The right to expand must first be earned through faster time to value, not merely a broader product story.

12. Repeatability, durable revenue, and human judgment remain the tests

A repeatable motion has two gates: a rep can turn a target list into a high-quality first meeting without the founder, then convert that meeting into a customer without extreme founder assistance. Mukherjee plans to manage five reps directly before hiring a leader to double the team.
His target hire has carried the relevant ACV for two to five years, demonstrated creative deal progression, and genuinely wants performance-linked earnings. Ninety-minute working sessions combine radical context-sharing with a live deal scenario; by the 30-minute mark, the quality of a candidate’s questions usually reveals whether “they have it.”
Rox targets a low-six-figure land—about $100,000—within 45 days of the first meeting, followed by a 30-day paid evaluation and a path toward high-six- or seven-figure expansion. Mukherjee prioritizes “enduring revenue” over hype-driven growth comparisons and says compensation should remain rationally explainable if disclosed.
He says the warning about preemptive rounds is not wrong, despite taking them himself: Rox raised $50 million and retained roughly $47 million in the bank. His largest mistake was undervaluing talent despite spending 40–60% of his time on it; he would have moved faster toward applied AI engineers rather than research hires, and later endorsed bringing product designers in early.
The quick-fire conclusion preserves the episode’s human core: the unchanged tactic is “get on planes”; the dead one is “automated emails.” Mukherjee once spent two months helping a buyer map a promotion path unrelated to the product—the buyer got promoted, completed the deal, and now runs the broader organization.

Nabeel Hyatt, GP @ Spark Capital: To Win in AI, Investors Need to Change Their Approach | E1255 Liberation Day, Tariffs, US v China Open Source, OpenAI, $CRWV, TikTok | BG2