How Gong builds products: pods, design partners, and radical autonomy

Executive overview

Most product teams build features, then find out if anyone uses them. Gong inverts this: every pod works with 6–24 design partners throughout development, converging on validated demand before launch.

The result is near-100% feature adoption. The same pod structure — cross-functional, autonomous, job-to-be-done-oriented — also lets Gong move fast on AI without losing rigor.

Working directly with customers at every stage is not a nice-to-have; it's the mechanism that eliminates wasted building.

The pod model

• Each pod: one PM, one UX designer, a tech lead, 5–7 engineers, fractional writer and analyst
• Pods are organised around a job to be done (e.g. conversation intelligence, forecasting), not metrics
• Pods have full autonomy to decide how to solve the problem and which design partners to work with
• ~25–30 pods currently; structure has stayed consistent from Gong's earliest days
• A dedicated research coordinator (borrowed from recruiting-coordinator concept) handles all outreach and scheduling — PMs arrive to pre-booked meetings
• Sales, product marketing, and customer success form a "virtual pod" around the core team

How design partner programs work

• Every pod runs 6–24 design partners per feature or product; as few as 5 for niche capabilities
• Partners are nearly always existing customers who have expressed interest in the capability
• For new products: weekly or bi-weekly sessions showing in-progress work, iterating on feedback
• For enhancements: looser cadence, a few meetings, then launch
• The PM's core skill is distinguishing must-have feedback from outliers — ask "how happy are you 0–10?" and aim to move the score from 6 to 8–9
• One-customer requests are still acted on in enterprise deals, but design partner programs target cross-customer patterns
• At ~7–9 design partners, requests start to converge — that convergence signals you've learned enough
• Gong claims close to 100% of features built this way end up used by a significant number of customers

Autonomy and trust

• Autonomy is framed as a selfish leadership choice: people produce more when they bring themselves to the work
• Teams decide independently whether to act on customer feedback or escalate — they are not punished for either
• PMs are expected to proactively solicit review (weekly sessions available), not wait for top-down checkpoints
• Leaders must also align peers (sales, CFO) to accept less direct visibility in exchange for higher velocity and engagement
• Features and even product lines have emerged bottom-up from hackathons, then received resources once they proved themselves

Making fast decisions

• For most decisions, both options are roughly equal — spending weeks deliberating does not improve outcome quality
• The 51–49 rule: if it is not a clear 70–30 call, just decide and move
• Sleep-level intuition is often as good as two weeks of analysis for domain-familiar decisions
• Applies to most day-to-day product decisions; one-way door decisions (major acquisitions, new market entry) still warrant more rigor
• Deep domain expertise is a prerequisite for trusting fast instincts on complex calls

AI product development: lessons from building early

• Gong avoided the term "AI" at launch because customers associated it with mistakes; now the field has overcorrected the other way
• Do not assume LLMs solve everything. Specialised models (e.g. deal prediction) still outperform LLMs for structured, domain-specific tasks
• Without measurement frameworks you cannot improve: Gong uses an Elo-style ranking system to evaluate AI output quality
• Key roles needed: data scientist (guides what's feasible, designs measurement), prompt engineer (optimises LLM inputs and edge cases)
• Even if you outsource model building, internal expertise is required to specify inputs, evaluate quality, and conceptualise product workflows
• Figma's framing "first draft" is a useful model for setting user expectations around AI accuracy
• Pods with embedded AI specialists can iterate quickly on both LLM and non-LLM (SLM) approaches

The spiral method for rapid learning

• When entering an unfamiliar domain, find one person and ask: what is this, and who else should I speak with?
• Talk to each referral; keep going until new conversations add less than 5–10% new information
• Works for technical domains (Reshef used it to learn deep learning) and customer discovery (e.g. understanding account manager personas)
• The signal that you are done: you start hearing the same things repeatedly across independent sources

Lessons from early failure: ICP focus

• Previous company hired 20 salespeople before having a focused ICP — all failed because no repeatable product-market fit
• At Gong, initial ICP was hyper-specific: US companies, selling in English, over video (WebEx), software priced $1K–$100K — roughly 5,000 companies
• A narrow pond creates word-of-mouth: one candidate told an interviewer they would only join companies using Gong, which directly converted that company to a customer
• Viral B2B effects only work when customers talk to each other — wide, unfocused markets prevent this