Three-Phase Product Arc: Wedge → Platform → Agents

Harvey

L2

Harvey's three-phase arc is the clearest sequential example in the cohort. The wedge was deliberately narrow in two dimensions simultaneously: by workflow (legal research, drafting, and due diligence specifically) and by buyer (individual lawyers on personal pain, not law firms on institutional strategy). Founder Winston Weinberg's framing: "We had less friction actually in the beginning because we weren't pitching to firms. We were pitching to lawyers — individual lawyers. And so their pain was: I don't want to do this particular piece of my job. Like, I don't want to go through tens of thousands of documents and do this massive closing checklist." This B2C2B motion gave Harvey production usage and real feedback before the enterprise procurement cycle began. The Allen & Overy pilot — 3,500 lawyers, 40,000 queries over months — was not a sales inefficiency but a trust-building investment calibrated to the specific risk tolerance of the buyer. Phase 2 arrived only after the wedge was deeply proven at the most demanding law firms in the world. Vault, Knowledge, and Workflows were not built speculatively — they were built because Harvey had enough enterprise deployments to understand which workflow expansions were genuinely demanded. The platform expansion opened the corporate in-house market, which requires a different pitch (legal as a cost center, not a competitive differentiator) and unlocked a completely different buyer persona. By Q4 2025, 42% of Harvey's revenue came from Fortune 100 corporates — a segment that barely existed in early 2025. The law firm relationships actively seeded this expansion: "Hey, did you know this is how we can use AI to do XYZ?" became Harvey's most efficient enterprise sales motion. Phase 3 — 25,000+ custom agents, 400,000+ daily agentic queries, the "selling the work" revenue-sharing model — was made possible by Phase 1 and Phase 2 establishing the infrastructure and trust layer. Pereyra articulated the product logic: "At all times you have to basically expand the product and then collapse it back" — build specialized workflows, chain them into unified experiences, expand the surface area, then give customers the tools to build their own agents on top. The frontier technical problems Harvey is forced to solve — document-based persistent memory, multi-tenant collaboration across organizations — arise directly from the depth of the Phase 3 agentic vision. These are problems big labs have not solved because they have not been forced to.

Glean

L2

Glean's three-phase arc was explicitly and publicly articulated by founder Arvind Jain as a deliberate two-act strategy before Act Two was even visible to the market. Act One was enterprise search — a product with zero behavior change required on Day 1, that generated clean adoption metrics, and that secretly built a permission-aware knowledge infrastructure that competitors would take 3–4 years to replicate. Act Two was AI applications and agentic reasoning built on top of Act One's foundation. The sequence was not accidental: you cannot sell agents into an enterprise that does not have clean, permissioned, indexed knowledge. The three years Glean spent in stealth building 40 design partner deployments was infrastructure investment that made Act Two possible. The transition between phases was triggered by the market, not by Glean's product roadmap. When ChatGPT arrived in November 2022 and created board-level urgency for enterprise AI, Glean was the only company with production-grade permission-aware knowledge infrastructure already deployed inside hundreds of enterprise tenants. The product category label changed from "enterprise search" to "Work AI" in September 2024 — after, not before, the company had reached $100M ARR. The category expansion followed the proven wedge; it did not precede it. The agentic Phase 3 — Glean Agents launched February 2025 — was built on the same data infrastructure as the search wedge, with no new data integration required for existing customers. This is the structural advantage of the sequential arc: the knowledge graph that justified the initial $50–60K departmental pilot became the foundation for $1M+ agent platform deals. The same infrastructure, more surface area. Jain's vision for Phase 3 — "every employee should have a team of AI agents" — requires exactly the permission-aware, fully-indexed enterprise knowledge graph that Glean spent its first three years building. Companies that tried to start at Phase 3 found they were missing the Act One infrastructure that made agents safe to deploy.

Abridge

L2

Abridge's three-phase arc is distinguished by a Phase 1 wedge that was simultaneously the most emotionally compelling and the most technically demanding in the cohort. The ambient documentation problem had a national price tag — the AMA estimated physician burnout costs U.S. healthcare $4.6 billion annually in turnover alone — which meant Abridge did not need to create demand. It needed to demonstrate ROI faster and with higher clinical confidence than every competitor. Pilot outcomes were extraordinary and quantified: Seattle Children's achieved 79% documentation effort reduction; Lee Health had 86% of clinicians doing less after-hours work; UNC Health's CMIO reported a physician who had written a resignation letter and chose not to submit it after using Abridge. These are not 20% improvements — they are life-altering outcomes that generated the physician evangelism that carried the product across health systems without a sales motion. The Epic "first Pal" designation in August 2023 was the Phase 1-to-Phase 2 inflection event. It reduced implementation from months to two weeks and made Abridge the default path for the 38–42% of U.S. hospitals running Epic. Clinician count grew from approximately 8,000 pre-partnership to 60,000+ by late 2024 — a 7.5x increase in 18 months that is not achievable through direct sales alone. This growth funded and justified the Phase 2 platform expansion: from outpatient notes to emergency department, inpatient, nursing, and order generation. Phase 3 was marked by Rao's explicit repositioning of Abridge from a documentation product to a revenue cycle company — a reframe that changed the buyer persona from CMIO to CFO. "Providers are compensated for the care that they document, not the care that they deliver...we're a revenue cycle company, along with the other things we do." This is the structural Phase 3 expansion: the same conversational AI infrastructure that generates clinical notes can also generate coding, billing, and prior authorization recommendations — but these products are sold to a different executive, with a different ROI argument, at a different ACV. The Phase 3 vision — conversations as upstream asset for diagnostics, therapeutics, coding, risk adjustment, clinical trials, and ultimately outcomes — is the total addressable market expansion enabled by a proven Phase 1 wedge.

Gong

L2

Gong's three-phase arc spans nine years and is notable for the stall that occurred at the Phase 2-to-Phase 3 transition — a fragility that every company in this cohort should study before designing their own arc. The Phase 1 wedge — call recording and behavioral analytics — was genuinely differentiated: it captured 6,000 words per hour of sales conversation that CRM would never record, creating a data asset that did not exist before. The 11/12 trial-close conversion rate in the alpha cohort was a PMF signal, not a sales skill result. Eilon Reshef's test: "9 out of 10 complaints were how come you didn't even record this call?" — users were angry when a call was not captured, which is the highest possible retention signal. The Phase 2 category creation move — naming "Revenue Intelligence" in October 2019 — was the right move at the right time. CROs could not ignore an outbound email about "Revenue Intelligence" because the word "revenue" was literally in their job title. The Gong Labs content flywheel, the Series C positioning as "the next evolution after CRM," and the ~140% NDR from seat-based expansion produced the hypergrowth phase (2019–2022) that most people associate with Gong's success. The Phase 2 stall came in 2023 when SaaS companies froze sales hiring. Gong's seat- based expansion model had a structural dependency on headcount growth in sales organizations. When that headcount contracted, NRR degraded from ~140% to numbers consistent with 16% YoY growth — not because customers churned, but because the expansion mechanic stopped. Amit Bendov's response was Phase 3: 18 purpose-built AI agents by October 2025 that generate value independent of seat count growth. The AI Briefer, for example, produced a 19% win rate increase and 40%+ on deals above $10,000 — these are outcome-based value metrics that can be priced independent of headcount. The stall and recovery sequence is the most important lesson in Gong's arc: Phase 2 platform expansion with a structural dependency on one economic variable (headcount) is fragile. Phase 3 agentic value creation can diversify that dependency if executed before the fragility manifests.

Decagon

L2

Decagon's three-phase arc is the fastest-executing in the cohort — $0 to approximately $50M ARR in 15 months, $4.5B valuation by January 2026 — built on the most precisely targeted Phase 1 wedge of any company studied. The wedge selection process itself is instructive: Zhang and Sreenivas ran 100+ structured discovery interviews before writing a line of product code, asking directly "How much would you pay for this?" across support, sales, and operations departments. The WTP signal from support was immediate and specific: "People were like, yes, if you can deploy this thing, I will sign a $150,000 check immediately, right? And this happened repeatedly." Other departments gave conditional, delayed, low-dollar WTP. The wedge was chosen by WTP discovery, not by market size analysis. The Phase 1 mechanism was a fixed 4-week pilot with pre-agreed pricing and success metrics defined at kickoff. This structure compressed the proof timeline while eliminating post-pilot renegotiation. The Duolingo English Test case illustrates the speed: a prior vendor had worked for a full year and failed to get chat live; Decagon went live in one month and achieved 80% chat deflection immediately. The 4-week pilot structure worked specifically because enterprise support has three properties that enable it: high interaction volume, near-real-time measurable metrics (deflection rate, CSAT), and a clear cost baseline (labor cost per interaction). Phase 2 — the Agent Operating Procedure (AOP) platform — was built from the pattern extracted across the first three bespoke implementations. AOPs are human-readable, versioned, overridable routing and decision rules that make the AI's behavior transparent and controllable. Once a customer has built 75+ AOPs (as Rippling did), the switching cost is not technical — it is the accumulated operational knowledge encoded in those procedures that would be lost in a migration. This is Phase 2 platform as moat, not platform as feature. Phase 3 — proactive agents that initiate customer contact, launched March 2026 with Hertz as the reference customer — represents the expansion from reactive support automation into customer relationship management. Voice 2.0 (sub-400ms latency, 6x cost reduction) opened the phone-based support volume that chatbots had never competed for. Each phase was built after the prior phase was proven in production, not in parallel with it. The three-phase sequence took Decagon from zero to a $4.5B valuation in approximately 30 months.