The 3 episodes below are this week's deep dives; the other 2 are written up in full in the Notion archive.
The medallion data architecture, from Kushal Sharma (head of internal AI at Circle) on LeanScale: the layered structure that makes your GTM data trustworthy enough for AI to run on. Three layers, each with a clear "done" state.
Underneath sits lineage: knowing that if a bronze field changes, which downstream silver and gold tables break. dbt maps this for you.
First step this week: pick one metric your GTM team argues about (NRR is the classic - monthly vs annual churn weighting changes the number) and trace it through your layers. Success means a new hire reads the definition and gets the same number. Failure mode: over-engineering gold early, promoting artifacts that haven't survived a business change yet, then maintaining reports the business already abandoned.
The dominant thread across every episode this week: the agent is the easy part, and the data layer underneath is the actual work. Sharma on LeanScale says it outright - "the agent was never the hard part" - and Lynch on GTMnow arrives at the same place, arguing the market misconception is that agents are easy to build. She distinguishes building one agent (trivial with self-serve tooling) from scaling agents across a GTM org, which needs shared data, governance, and infrastructure like calling and email deliverability. Totten on GTM Science lives the payoff: Vercel's agents work because they're fed dense customer context and telemetry, and he's blunt that clients who can't see results usually have bad underlying data, "100% of the time." For a RevOps leader, the sequencing rule is: fix your data model and orchestration before buying another agent, or you're automating garbage.
The second thread is a reframing of what RevOps is. Sharma rebranded his RevOps team as the AI org, and both he and Jiao on RevOps Lab treat RevOps as a product-owning data function rather than a CRM admin support center. Sharma's warning: teams that don't take control of the data story become "a group of people that just supplies prompts in ChatGPT" - relevant, but not leading. The AI strategy is RevOps's to take precisely because they already sit in the systems and the data.
One tension worth flagging: Lynch's augmentation thesis - make a rep 10x, hire more reps - sits against the AI-native economics Rike and Buchanan lay out on AI to ROI, where model COGS crush margins and buyers increasingly build rather than buy. A CRO betting on aggressive rep expansion should stress-test whether productivity gains survive the cost of the agent infrastructure producing them.
The three episodes worth your time this week, with the mechanics behind the takeaways above.
Kushal Sharma, head of internal AI at Circle, argues GTM AI runs on infrastructure, not models.
Two concepts matter for data-driven revenue teams. The semantic layer is metadata attached to your data model: not just "this column holds NRR" but how NRR is calculated, whether churn is weighted monthly or annually, and the business logic encoded in it. His test: if a human reading it can't understand the data, neither can your AI. The context graph solves retrieval by traversing connected nodes to pull only relevant context per query instead of scanning an entire knowledge base, making answers cheaper and more accurate. Most of this pipeline is traditional software, not an LLM, which keeps it cheap, reliable, and model-agnostic.
The maturity ladder is the actionable core: don't buy a vector database at 20 records, hardcode context in a skill instead. Buy one at 2,000, when scanning everything to find the relevant 10% gets too costly. Circle's homegrown AI SDR generated close to seven figures, but Sharma is careful to note the tech unlocks speed, while product and brand produce the revenue.
David Totten, VP of Field Engineering at Vercel, runs technical customer touch across a consumption-based business where post-sale is most of the revenue. His most portable move: killing CRM data entry. Instead of engineers spending an hour a day updating opportunities, an agent runs sentiment analysis on customer engagements and pushes dates, contacts, and product feedback into the CRM via API, with a human reviewing and approving. This closes the gap between what ICs see day to day and what leaders can see, without the manager-chasing-rep tax.
On retention, churn in pro and enterprise stays low not through CSM coverage but through in-product telemetry. Agents flag customers burning through commits faster than expected or hitting cost spikes, and the product recommends fixes, often a cheaper architecture. His view: churn happens because customers use one slice of the platform suboptimally, not because they dislike it. The expansion mechanic works the same way: many customers overspend on premium options when a cheaper route would do the same job, so the expansion play is showing customers how to spend less, which builds trust that drives consumption elsewhere.
The organizing philosophy: treat GTM like a product, with sentiment analysis, sales-trend analysis, and a shift from mostly reactive to proactive, acting on a signal rather than waiting for a complaint.
Loreal Lynch, CMO of Nooks (formerly CMO of Jasper), argues that bolting AI onto pre-AI processes fails because today's GTM stack was built for humans to click through, not for agents to execute. Her core claim: agents make reps more productive, so the response is to hire more reps, not fewer. Teams on an agent stack see roughly 3x meetings booked and 2x pipeline on existing headcount, the same logic as the 10x engineer who ships more rather than getting cut.
The trap she names: agents are easy to build, hard to scale. A prompt saved as an agent is trivial; running agents across a GTM org needs shared data, governance, and execution infrastructure like email deliverability and dialing. RevOps leaders grasp this immediately; others learn it the expensive way.
The most concrete GTM mechanic is the CMO-CRO alignment model she runs with her CRO counterpart. Instead of separate sourced-pipeline targets, which trigger the "you're not closing my leads / you're not sending me good leads" fight, they share one pipeline goal both are accountable to. Their enterprise motion drops sourced attribution entirely because a deal might touch a dinner, an exec LinkedIn connection, and a case study before closing, so crediting a single source is counterproductive. Pipeline meetings look at the top-line number by segment and region, then ask jointly where the gap is and whether marketing programs can close it.
Full write-ups are in the Notion archive.
The LeanScale Podcast, "The Agent Was Never the Hard Part" with Kushal Sharma. The audio earns the hour because Sharma explains semantic layers, context graphs, and vector databases from first principles in plain language, with Jake Toepel translating the jargon in real time - it's the rare technical masterclass a non-engineer RevOps leader can actually follow and act on.
Get the next one on Monday. Free, one email a week.