The Forward Deployed / Customer Success track at SaaStr AI 2026 turned into something closer to an autopsy and a rebirth at the same time. Leaders from the fastest-growing companies in B2B + AI got on stage and said versions of the same thing: the post-sales playbook most of us spent the last decade building is now a liability.
The data backs it up. The CSM role grew over 700% through Q2 2022 and then flatlined for four years running. Over that same window, forward deployed engineering is up more than 1,000% and still climbing. That is not a rebrand. That is a structural shift in where the work lives and who does it.
Here’s what each speaker actually said, and the specific numbers behind it.
The Top 10 Unexpected Learnings:
- The CSM title is a tax on your customer experience. Ryan Seams (Assembly AI) watched technical buyers physically go defensive the moment he said “head of customer success.” Renaming the role to forward deployed engineer filled a recruiting pipeline that the “technical account manager” title couldn’t, two candidates in 2.5 months versus a full pipeline, with no change to the actual job.
- Saying “AI-powered” signals you’re behind, not ahead. Monica Perez (Lovable) stopped leading with AI entirely. Her logic: AI is becoming the baseline, the water the fish swims in, so making it the headline reads as not thinking ten steps ahead. The moment you stop saying AI in every sentence is the moment you’re actually AI-native.
- A traditional, on-site, “2001-looking” CS motion can still be the right answer. Tom Ronen (Harvey) runs EBRs, heavy on-site work, and old-school change management frameworks at an $11B company, and tells other CS leaders to do the same. Selling AI into a law firm isn’t selling software, it’s selling a change in how 30-year partners work.
- Adoption is now the starting line, not the finish. Ronen’s point: a firm can log into Harvey constantly without changing how it operates. The old SaaS shorthand of “good seat utilization equals a safe renewal” breaks with AI, where daily use of the wrong workflows still means churn.
- A single change-management document drove 2x revenue. Ronen cited a Harvard study of 1,515 startups: same AI tools, same training, but the group that got one document on how teams like theirs had succeeded showed 2x more revenue and was 18% more likely to acquire paying customers. The human layer, not the model, moved the number.
- NPS is dead, and the panel said so unanimously. Bobby Cooper’s panel with Ursula Llabres (Content Square), Ashvin Vaidyanathan (LinkedIn), and Daniel Silverstein (Carta) agreed: low response rates where non-responders are the churners, no clean correlation to GRR, and a heavy translation layer to action it. Llabres flagged a real gap, a great NPS score sitting next to weak retention.
- Over half of CSM activity has zero correlation with retention. Cooper’s (Retention Intelligence) platform data shows more than 50% of what CSMs do doesn’t tie to the goals on their plate. The fix from Vaidyanathan (LinkedIn): map every activity, human or agent, to whether it changes the product outcome in the intended window, and cut what doesn’t.
- Moving the “closed-won” line into implementation crushed churn. Cooper’s case from Weave: a deal only counted as booked once the customer hit a success threshold inside implementation, not at signature. That fixed the handoff, aligned sales to qualify better, and took churn from 4% per month to roughly half a percent while scaling from $8M to $200M ARR through IPO.
- The fastest companies to $100M ARR are all developer tools, and that’s the tell. John Gleason (SuccessVP): Cursor (12 months), Bolt (14), Lovable (8), Replit’s agent ($10M to $100M in six). Code inherits high context and verifiable correctness for free. Every other domain has to engineer those conditions, and that engineering is the new job of CS.
- You can rebuild your CS stack yourself, and version it weekly. Perez (Lovable) replaced Gainsight with a command center built on Lovable, maintained like a product with a feature backlog and weekly releases. She’s not technical. Her argument: the team closest to the customer should build the tool, and waiting on a vendor’s roadmap is now the constraint, not the budget.
Monica Perez, Lovable: 5 Counterintuitive Moves From the Fastest Company to $400M
Lovable is the fastest company in the world to $400M, and Perez leads customer success there. Her framing of CS is broader than CSMs: it’s every person whose job is to close the gap between what you sold and what the customer actually experiences. She walked through five moves that felt wrong but worked.
Move 1: Stop talking about AI as the headline. Every deck in the room said “AI-powered,” including hers. Her point: leading with AI signals you’re not thinking ten steps ahead. AI is becoming the baseline, the ambient layer that powers everything, not the story. If you’re the fish, AI is the water. Lovable’s onboarding doesn’t start with what AI can do. It starts with what the customer will unlock. The moment you stop saying AI in every sentence is the moment you actually become AI-native.
Move 2: Rewire the clock. Quarterly cycles, 90-day plans, QBRs, annual renewals. All of it was built for products that shipped a couple times a year. Lovable ships multiple times a day, so CS has to move at product speed. Her own hiring proved the point: a one-week work trial where she onboarded a real Fortune 100 customer before she was even hired, then met every single customer in her first 30 days. No 90-day plan survives a world that changes completely in 90 days. Every customer gets a Slack channel with an AI bot scoped to their specific instance, their projects, their users, even their contract terms. Hundreds of individualized bots, no harder to build than one.
Move 3: Design for experiences, not activities. CS has been organized around who runs onboarding, who owns the renewal, who handles implementation. That ships your org chart to the customer. Customers don’t want an org chart, they want an outcome. Lovable has a full-time team running hackathons at customer offices. At T-Mobile, teams shipped over 100 working applications in a single day. They offer product management and engineering as a service. They run founder masterminds on Discord 24/7. And it changes who you hire: builders, not playbook-runners. “Show me your projects” beats “tell me your process.”
Move 4: Control your stack. Tools used to amplify humans. Then SaaS flipped it and humans started managing tools, 15 logins and 12 dashboards. Rigid tools also move on 12-month cycles, so if a tool moves slower than your team, the tool is the constraint. Lovable rebuilt its entire CS platform on Lovable, replacing Gainsight with a command center they version weekly. It tracks the portfolio in real time, surfaces risks and expansion before they become fires, and generates individualized living customer hubs. Her message: you don’t need to be technical to own your stack.
Move 5: Value and revenue are the same thing. Consumption and credit-based billing flips the incentive structure. The more a customer uses the product, the more they pay, so revenue is driven by adoption, and adoption is driven by everything CS does. No awkward upsell, just “let me help you do more of what’s working.” Her proof point: EXP Realty ran a company-wide hackathon, built 27 real products including a transaction management system and a microsite generator, replaced two major tools, saved over $1M, and then launched an internal app store so anyone could keep building. In an AI world, possibility is no longer impressive. Proof is.
Tom Ronen, Harvey: Change Management Is the Moat, Not the Cost
Harvey is the leading AI platform for legal, crossing $190M ARR in January, used by over 100,000 attorneys across 60+ countries and 1,500+ organizations, at an $11B valuation. The number that doesn’t make headlines: nine major releases in the last 30 days. Not bug fixes, new surface areas.
And Ronen’s confession was the opposite of what the room expected. Harvey runs a traditional CS motion. Executive business reviews, heavy on-site work, change management frameworks. When CS leaders ask him how to run an AI-native playbook, he tells them this, and they look at him sideways.
His argument: selling AI into a 200-person law firm isn’t selling software, it’s selling a fundamental change in how partners who’ve practiced the same way for 30 years do their work. No automated QBR or fancy health score gets you there.
The supporting data from a Thompson Reuters survey of 1,700+ legal professionals: 26% of legal teams use GenAI, double the prior year. 78% believe AI will be central to their work. And the number that matters for CS: firms with a visible AI strategy are 4x more likely to see ROI than those with ad hoc adoption.
Three things shift when AI is the core product. The busy work goes (Harvey automates EBR prep and next-steps tracking across books of 40-50 accounts). The assumption that adoption equals success goes (a firm can log in constantly without changing how it operates). And the human layer stays.
His proof was a story. An Amlaw 50 firm, deeply skeptical, CIO benchmarking the application layer against foundation models directly. Over repeated on-sites, they identified an influential funds partner, got lunch, and a legal engineer asked one direct question: what’s slowing you down today? That led to building real workflows on the partner’s live matters: a DDQ review agent, a debt review summary tool, an LPA consolidation workflow. At the partner retreat, that same partner grabbed his laptop and demoed the workflows himself. Requests flooded in. The account expanded 16x.
The research backs the human layer. A Harvard study of 1,515 startups gave both groups the same AI tools and training; the group that got a single change management document showed 2x more revenue and was 18% more likely to acquire paying customers. A Microsoft survey of 500 enterprise decision-makers found AI vision and change management confidence correlated most with success, above the technical factors. Box’s Aaron Levie put it directly: the winners in AI will be the ones who deliver the change management that drives workflow change, not the best model or UI.
Harvey’s three pillars: build trust progressively (rooted in ADKAR, starting with a session that demystifies LLMs before anyone logs in), invest in domain expertise (legal engineers who are ex-practicing lawyers, with roughly 40% of Harvey’s GTM team carrying a legal background, and books of business segmented down to the smallest economic unit), and go deep not wide (daily use as the bar, customers uploading their golden precedents and playbooks until Harvey becomes infrastructure you can’t rip out).
Ryan Seams, Assembly AI: Your Customers Don’t Want a CSM, They Want an Engineer
Assembly AI builds voice AI infrastructure for developers. Seams opened with a story about his own title. When he introduced himself as head of customer success on a listening tour, customers physically changed their body language and went defensive. Not because of him, because of what the title now signals to technical buyers: a commercially minded message-relayer there to run the QBR and get the renewal.
What technical customers actually want is someone who can build. His example: a customer asked for additional metadata out of Assembly’s transcription API and said if Assembly couldn’t do it, they’d build it themselves or buy a competitor. One forward deployed engineer sat down, integrated three open-source models, deployed a custom API endpoint on Railway, and shipped it as a working product over a weekend. The customer stayed, and the work became a candidate for the product roadmap.
The title journey is the most useful part for founders. Assembly went CSM to TAM (technical account manager) to FDE. The TAM experiment was a disaster in recruiting: two good candidates in two and a half months. Switched to forward deployed engineer and the pipeline filled with the right people. The actual job barely changed. It was the brand and the stigma. He shared a customer quote from a call literally begging for an FDE, and the person on the call had updated their Slack title to FDE the day before.
The metrics are where it lands. Assembly’s customer support agent (internally “CubClaw,” the baby version of their internal FDE agent “BearClaw”) moved end-to-end resolution from 10-15% with a vendor to 75% in 45 days. His point on honest metrics: deflection rate alone is vanity. The real number is pickup rate times end-to-end resolution rate, so an 88% deflection rate on 20% of issues isn’t what it sounds like. Net result, Assembly scaled customer support across 3x year-over-year customer count with the same headcount.
His three takeaways: hire FDEs not CSMs, retool the CSMs you have into FDEs (anyone can learn to code with Claude Code and similar tools, but you do have to train them), and treat your customer experience like a product that FDEs build and agents scale.
The Panel: Bobby Cooper, Ursula Llabres, Ashvin Vaidyanathan, Daniel Silverstein
Cooper (Retention Intelligence) moderated Ursula Llabres (Content Square), Ashvin Vaidyanathan (LinkedIn), and Daniel Silverstein (Carta). A few threads stood out.
Treat AI like new hires, not a SaaS deployment. Llabres’s framing: ask what tasks AI can do like a human would, rather than bolting agents onto broken processes. Carta deployed Claude to all 2,000 employees two months ago and Silverstein’s guidance to his teams is to stop thinking linearly. Tell Claude the outcome and let it do the thinking on process, because CS practices have been step-by-step-by-step for years and that’s not how these systems work.
Does AI improve CX or just save money? Jury’s out. Efficiency is short-term and easy to see, so companies are mining it first. The harder, unsolved problem is how human and machine journeys hand off to each other. Strong hypothesis that it improves experience, too early for a clean data point.
The activity problem is real. Cooper’s platform data shows over 50% of CSM activities have no correlation with retention or the goals on their plate. Vaidyanathan’s response: map every activity to whether it changes the product outcome for the customer in the intended timeline, and if it doesn’t, it has no place in the playbook, human or agent.
The one KPI. Time to value, near-unanimous, with Llabres adding “handover success” as the place most failure and data loss happens. Cooper’s case from Weave: they moved closed-won from a sales/RevOps milestone to a threshold inside implementation, which fixed the handoff problem (a signed deal that never kicked off was never counted), drove alignment, and took churn from 4% per month to about half a percent while scaling from $8M to $200M ARR through IPO. LinkedIn ties comp to percent of successful seats and customers, with clawbacks if you miss, which made sales qualify better and pulled product into the same metric.
The KPI to kill. NPS, unanimous. Low response rates where the non-responders are often the churners, time-based delivery that catches people at the wrong moment, no clean correlation to GRR, and it requires a heavy translation engine to action. Silverstein can’t kill it because the board cares, so he supplements with CSAT and CES and aggregates. Cooper also flagged time-to-live as arbitrary when it’s based on a finance milestone with no correlation to value.
John Gleason, SuccessVP: Two Forces Are Rewriting the Function
Gleason runs SuccessVP, an applied-AI fund backed by CS leaders at Anthropic, OpenAI, Harvey, Lovable, Toast, GitHub, and GitLab, and previously took Motive’s CS from $1M to over $300M ARR. He closed the day with the cleanest framework of the event.
His starting point: CS was never one right thing, it was a response to SaaS. Different archetypes mapped to different business models, enterprise deployment (Salesforce, HubSpot), product-led (Slack, Atlassian), consumption (Stripe, Snowflake). The question is whether AI is a new archetype or a new operating environment. He argues it’s the latter, driven by two forces.
Force one: non-deterministic systems. The same prompt with the same data can yield different answers. That’s the architecture, not a bug, and it changes how software has to be supported. AI succeeds when two conditions are met: high context and verifiable correctness. Code has both baked in (centralized, versioned context; tests that pass or fail), which is why code was first. In GTM, tier-one support and lead qualification have both, which is why AI took hold there. But 56% of CEOs see financial gains from AI and 95% of enterprise AI pilots generate no financial return at all. The problem isn’t the models, it’s structural: most business problems lack context or can’t verify correctness, and someone has to engineer both. That someone is now the vendor. He pulled 100 FDE job descriptions and they split cleanly into two camps: context (Hebbia, Harvey, Rogo embedding to capture business context) and correctness (Glean, Cursor, Hippocratic owning evals and quality bars). The title is incidental. The work is the new requirement of CS.
Force two: units of work. The valuable thing is shifting from the software to the outcome the software produces. He framed every business model as the distance between when a customer pays and when they pay again for value. On-prem stretched it so far the renewal was a new sale. SaaS compressed it to an annual contract, which created CS. Consumption shrank it further. Outcome-based pricing eliminates it: value and the work become the same thing, a resolved ticket isn’t evidence of value, it is the value.
Put the two forces together and you get high context plus verifiable correctness equals a unit of work, and the unit of work is what you get paid on.
The growth data: in SaaS, $100M ARR in under a year was nearly unheard of. In applied AI, Cursor hit it in 12 months, Bolt in 14, Lovable in 8, and Replit’s agent went from $10M to $100M ARR in six months. Every one of those is a developer tool, because code inherits the conditions for free. Most domains don’t, and engineering those conditions is the new job of CS. Sequoia’s framing: services are the new software, and there’s roughly $6 of services spend for every $1 of software. Applied AI is going after the work, not the software market, which is why CS moves from protecting an annuity to operating a toll booth.
His four metrics for the new world: eval pass rate (the new churn-or-renew signal), time to first value (now a real number because the unit of value is discrete), velocity (units of work completed per day on a cohort basis, not just usage), and a reset NRR (structurally uncapped, because the ceiling isn’t seats anymore). And the practical starting move: a time-and-motion study to define your unit of work, map the context and correctness inputs, pressure-test where it breaks, then hire consultants for the context side and domain experts for the correctness side.
The Old “Customer Success” Has Been Split and Moved
Five sessions, leaders from companies growing faster than almost anything in B2B history, and a shared conclusion. The work that used to sit inside customer success didn’t disappear. It moved, split into context and correctness, and got attached directly to revenue.
The companies treating AI as a cost-efficiency play are mining the short-term win. The ones treating it as a structural reboot are the ones hitting $100M ARR in months. The function isn’t being disrupted. As Perez put it, it’s being promoted, and the seat at the table is real for anyone ready to claim it.
Have a question for Dear SaaStr? Submit it at saastr.ai/ai-mentor
