We run SaaStr with 3 humans and 21+ AI agents. Two of the companies that just presented are already in our own stack. Ava from Artisan sent 7,000 cold emails for us in a six-week stretch. The “Amelia AI” running on our site is Piper from Qualified. So when these founders walked through what their agents do, I wasn’t watching a demo reel. I was watching tools I already run, plus the next wave coming for the ones I don’t.
Seven sessions, seven companies, one through-line: the go-to-market stack is collapsing from a pile of point tools into a smaller set of agents that actually do the work. Here’s each session, the top three takeaways, and what they add up to.
1. Vercel’s CPO: From Pages to Agents
Tom Occhino, Chief Product Officer
Tom helped build React and React Native at Meta before joining Vercel, where he runs product, engineering, and design (his CTO, Malte Ubl, spent 12 years at Google and created Core Web Vitals). His framing for the whole talk: the giants invested in foundational infrastructure at roughly the same level as product itself, and almost none of it was purpose-built. You could run Google search on Meta’s infra and Meta’s feed on Google’s. Every hour everyone else spends configuring infrastructure, building bespoke frameworks, and babysitting DevOps is energy that never reaches the product or the customer. He calls that “undifferentiated heat loss.” Vercel’s pitch is to be the general-purpose version of that foundation, now built for agents.
The thesis is that software is moving from pages to agents. For decades the UI was the trunk of the tree and every action branched off it: a user navigates to a page, takes an action, something happens. With agents, the UI drops down to the leaf nodes, lightweight surfaces where a human makes a decision or reads an outcome, and the trunk becomes headless autonomous software that wakes on a signal (a new lead, a fraud alert, a monitoring event), acts on its rules, and escalates only when it has to. He thinks every company ends up with at least two: an internal agent for employees that sits in Slack or Teams and replaces navigating legacy systems, and a customer-facing one for support, transactions, and self-service.
He doesn’t believe in buying agents off the shelf. “Agents as a service” fail for the same reason one-size-fits-all websites do, except the fit problem is worse, so every team builds its own. Vercel built hundreds in the last year and uses Vercel to build Vercel. The content agent turns Slack threads into blog drafts and now produces 96% of marketing first drafts. The lead-qualifying agent replaced a large SDR team, and those people moved into higher-impact roles. The support agent handles 93% of inquiries with no human, and the ones that slip past it get treated as a product bug or a config problem to fix rather than tickets to grind through.
Then he opened up the architecture, which is where the real lesson sits. Their GTM agent, Deal One, runs on a multi-layer stack. A Gong call ends, Gong fires a webhook to a Vercel function, and that kicks off a durable 10-step Workflow that finishes start to finish every time. It pulls the transcript, runs it through Claude via their AI Gateway, and produces a structured summary (topics, objections, deal stage, sentiment, stakeholders). That summary gets embedded into a hybrid vector-plus-keyword index alongside every other call, atomic objections, and cross-deal patterns, then exposed through a single secure MCP server so any agent (Deal One, Claude Code, whatever they build next) reads the same context through one auth boundary. Reps @mention Deal One in a deal channel and ask why a deal slipped or for coaching before a call, and get back cited answers that link to the exact moments in the Gong transcript or Slack thread. Each turn runs inside a Vercel Sandbox that gets snapshotted and resumed for the next message, so conversation state is durable with zero infrastructure to manage.
That stack matters because of the toolkit underneath it, which Vercel built to fill its own gaps and now sells as the agent stack:
- AI SDK: integrate pretrained or custom models in a few lines, and swap models without touching app code. You can try a new model the day it ships instead of cutting a new release to test it.
- AI Gateway: one interface to 100+ models, no juggling keys, accounts, or rate limits, with fallback routing so a rate limit or upstream outage fails over to another provider instead of silently dropping requests. Deal One reaches Claude through this.
- Chat SDK: one TypeScript codebase that targets Slack, Teams, Discord, and WhatsApp, so you stop building the same integration two or three times.
- Fluid Compute: serverless scale with server-like flexibility. A nice-to-have for normal apps and close to mandatory for AI ones, which spend most of their time waiting on a model to stream back and shouldn’t burn compute idling.
- Sandbox: a secure, ephemeral environment for running untrusted or model-generated code without touching production. Deal One executes there.
- Workflow SDK: durable orchestration with retries, state persistence, and observability. Optional for traditional software, foundational for agents, which need to pause, resume, and wait on a human in the loop.
His advice for your first agent: find low-value, highly repetitive work, document exactly how it should be done in painful detail (an agent is only as good as the process you hand it, and if you can’t write it down, neither can it), then point a coding agent at that document and build.
Top 3 Takeaways
- The UI is becoming the leaf and the agent is becoming the trunk. Build headless first, then add the thin human surface on top.
- The architecture is the moat. Durable workflows, one shared context layer through an MCP server, and model fallback are the difference between an agent that holds up in production and a demo that breaks.
- Document before you automate. If you can’t write the SOP in painful detail, the agent can’t run it, so the writing is the actual work.
2. Artisan’s CEO: The AI BDR That Runs Your Outbound
Jasper Carmichael-Jack, CEO
Full disclosure: we use Ava. SaaStr sent roughly 7,000 emails through her in a six-week window and got a 3.6% positive response rate, all on Ava 1.0. So I’m not a neutral observer here.
Jasper’s pitch is that outbound used to be a taped-together stack: a sequencer, a data tool, an enrichment tool, a copywriting tool, all stitched into one fragile workflow. When everything is fragmented, nothing has accountability and you can’t see what’s actually driving ROI. Artisan consolidates the stack and leads with cost transparency. The UI shows cost per lead, cost per meeting, and the underlying credit cost for every outcome. Ava 2.0 now handles replies and books meetings fully autonomously, with optional human escalation rules. She won’t cold call (it’s illegal for outbound, and humans are still better at it), so she queues calls with per-person talking points in a native dialer.
His framework: outbound comes down to who, what, and when. The who runs on two databases: a 280M-plus B2B contact set (trimmed from 450M down to ~272M to cut the junk, waterfall-enriched for email and phone across a couple dozen providers and bounce-tested) and a second one covering every local business with a Google Maps profile. Your own CRM data beats both because it carries relationship context. The what is where Ava spends her effort: she pulls mutual former employers, mutual investors, work history, and company-level signals, and you can stand up custom research agents that send a swarm of web researchers after whatever you specify, then she picks the best angle per message. The when comes from person- and company-level website de-anonymization plus intent signals like funding and hiring, with social engagement and custom AI signals rolling out.
Pricing moved to credits at roughly 2 cents each, landing around 30 to 60 cents per lead enrolled and down to ~20 cents for a bare-bones campaign, and it’s self-serve now with $300 in free credits and no credit card. A campaign Jasper ran against YC founders makes the point: he coached Ava to write in lowercase with a casual, human tone, none of it timed to any signal, and the who and the what alone pulled a 4% response rate.
The part that matters for anyone evaluating this category: Cook Unity got terrible results for two full months, stuck with it, kept iterating, and now sits under $50 cost per meeting booked. Jasper calls it “outbound market fit.” Even with product-market fit, people may not reply because they don’t want your product yet, in which case you move to warm, CRM-driven outbound rather than net-new cold. On the “Stop Hiring Humans” billboards: the strategy is to get attention, then convert it to pipeline, and controversial converted better than polite.
Top 3 Takeaways
- Outbound is who, what, when. Fix the data and the message first; timing is the multiplier on top.
- Give it real iteration time. Outbound market fit can take two-plus months. Don’t kill the program in week three.
- Demand cost-per-outcome visibility. If your tool can’t show cost per meeting, you can’t manage it.
3. Lightfield’s CEO: The CRM That Builds Itself
Keith Peiris, CEO
Keith built Tome at Meta-scale before starting Lightfield, which now has around 3,000 customers. He ran a live demo on real Claude with a synthetic dataset of a process-automation company selling to manufacturers.
The premise: connect mail, calendar, data warehouse, and call recorder, and the CRM fills itself. Accounts enriched, opportunities auto-created off emails and calls, contacts populated by about ten vendors, no manual entry. Then he worked a stalled Johnson Controls deal in real time. He asked Lightfield “why is this deal stalled,” and it ran code in a sandbox to compare the deal against closed-won and closed-lost patterns, found they had never met the CIO, found the actual CIO, drafted an intro email in his voice, and sent it. Then he promoted that one action into a natural-language automation: “run this every time a deal reaches POC without an IT contact.” It runs as Python in the background with role-based permissions you approve at activation.
Then he used the learning to prospect: surfaced 10 matching heads of IT, filtered them by legacy factory-floor software signals pulled from job postings and LinkedIn posts, and built a custom multi-step sequence off prior winning sequences and his own writing style.
The automations are not rigid rules. They run as Python in the background, which Keith framed as Apex with a natural-language front end and far more power, so you build and edit workflows in plain English as you learn from one deal at a time.
On governance, the first question every RevOps leader asks: Lightfield stores everything at a foundational level (emails, events, calls) and puts a CRM schema on top, then tracks version history on every field, attribute, and object with role-based access control, so you can see whether an agent or a human changed something and roll it back. The agent, the UI, and external systems all run through one Lightfield API, so a rep’s code execution is bound to the same data access and rate limits as that rep. One smart default keeps the database clean: outbound only syncs into Lightfield when someone replies, so you don’t drown in 500,000 dead contacts. Migration off Zoho or HubSpot takes about two hours, training runs 30 to 45 minutes, and anyone who has used ChatGPT picks it up. That is why adoption holds: there’s no CRM busywork left to avoid.
Top 3 Takeaways
- The CRM should do the data entry, not your reps. Adoption follows when the busywork is gone.
- Turn one good play into an automation. Work a deal, then promote the winning motion to a rule.
- Your closed-won and closed-lost history is prospecting fuel. The patterns in won deals point straight at new pipeline.
4. Attention’s CEO: Your Conversations Are the Growth Engine
Anis Bennaceur, CEO
Attention is a Series B company at roughly $15M ARR, growing 4 to 5x year over year, with customers like Scale AI and Lovable. Anis’s argument: your prospect conversations already hold the secrets to your next quarter of pipeline, and most teams let that intelligence evaporate. His own team built what he calls a go-to-market machine that compounds, and the fuel is every recorded conversation.
He walked through three plays. First, refresh your ICP on a schedule. Build an agent that pulls recent closed-won deals every month or quarter, analyzes the conversations for buyer function, tenure, company size, geography, firmographics, technographics, and intent, then turns a broad ICP into specific persona pockets (his example: “the frustrated veteran stuck on an incumbent”). Your ICP from last quarter is already stale.
Second, predict reply rate before you send. For each target, Attention pulls everything from past conversations (even ones from 9 or 12 months ago), runs fresh deep research on the web, ranks it, has the agent write the email, then builds synthetic personas from real buyers and runs them through a probabilistic model to predict the reply rate as if you had sent 100 similar emails. If it lands below ~5%, the agent rewrites and re-scores until it clears the bar, then drops the message in Slack for a human yes or no before it sends.
Third, a proactive agent that surfaces deals at risk, re-engages prospects who are ghosting, and re-opens deals you lost nine months ago, then takes the action rather than just flagging it. He noted that a call repository tells you what happened; this is built to get the next thing done.
Top 3 Takeaways
- Your ICP is a living document. Re-derive it from recent won deals, not last year’s deck.
- Predict before you send. A reply-rate model plus synthetic personas beats spray and pray.
- Closed-lost is not dead. The intelligence buried in old conversations is your cheapest re-engagement fuel.
5. Qualified VP Demand Gen: AI SDRs Done Right
Sarah McConnell, VP of Demand Generation
Disclosure again: the “Amelia AI” on our website is Piper, built around a custom avatar of our Chief AI Officer. For SaaStr, Qualified reports 130 meetings booked, $2.5M in pipeline, and $1M in closed-won revenue. On the site she runs in always-on “pin mode” rather than as a corner chatbot, present without covering up content.
Sarah’s whole message is that the tech is the easy part and the foundation is not. Before you launch an AI SDR, do five things. One, decide where you actually need coverage: after-hours, other geographies and languages, or all inbound from scratch. Two, onboard and train it like a human SDR. Give it the same materials a human would get; an hour of setup buys you off-brand answers. Three, set targets based on the use case, not a fantasy. She benchmarked against human SDR meeting volume, then adjusted for the actual after-hours lead volume and conversion. Four, define SLAs so the AI SDR, human SDR, AE, and marketing aren’t all hitting the same lead at once. Five, evaluate and coach continuously through a scorecard with thumbs up and thumbs down.
On the build: themes handle branding, the agent profile sets the persona, onboarding crawls your whole site (which doubles as an audit, since it surfaces the stale, outdated content you forgot was live and makes you clean it up) and lets you upload offline PDFs and slides, plus snippets for material that isn’t on your site at all, like competitive battlecards. Guides are natural-language guardrails (never give exact pricing, always note that pricing depends on the customer). Segment-specific guides change tone and content in real time when reverse IP recognizes an enterprise versus an SMB account. Then goals and a scorecard. Her most counterintuitive note: write the instructions in plain human language. Using another LLM to generate them actually performs worse.
She closed with the three highest-ROI use cases. High-intent buyers who won’t fill out a form (trigger follow-up off page visits or prior conversations). Maximizing ad spend (offer content instead of only a demo form, so you capture leads that would otherwise bounce). And pre- and post-event outreach, where Qualified saw a 70% increase in event pipeline within one quarter, because human SDRs hate working event leads and an AI SDR will work every one of them at scale.
Top 3 Takeaways
- Onboard the AI SDR like a hire. Same materials, same coaching loop, or it underperforms a human.
- Write your guardrails in plain English. Don’t let another model write the instructions for this one.
- Point it at the leads humans won’t work. Event lists and no-form-fill traffic are where it prints pipeline.
6. Aurasell’s CEO: Contact to Contract on One Platform
Jason Eubanks, CEO, with Carol Lee Robarts, Founding AE
Jason was CRO at Harness, Twilio, and Meraki before starting Aurasell, and his framing is that AI-era CROs got quietly turned into CIOs, owning a sprawling tool estate instead of selling. His old stack tells the story: 22 products, more than $3M a year in software fees, and 11 ops people stitching integrations, patching fragile workflows, and chasing a single view of the customer. Asked by his board to keep growing while cutting burn, he ran an internal audit called Project X-Ray, drilling into every department, logging every activity each person did and every tool they touched, and hunting for overlap. Two findings drove the whole company: reps were living in 10 to 12 products a day, and the industry benchmark is that B2B sellers spend only 24 to 30% of their time in front of customers. That second number is the entire pitch.
Aurasell’s answer is a unified data architecture (structured and unstructured together), a conversational context layer across all channels and signals, and an automation layer of prebuilt and custom agents that runs contact to contract. The platform ships with 900M contacts and 85M accounts, auto-enriched. You can run it as your AI-native CRM and auto-migrate off the old stack, or layer it on top of Salesforce or HubSpot with bidirectional sync and adopt at your own pace.
Carol Lee, the founding AE, demoed her real first 41 days, in which she closed $2.7M. On day one a territory agent had already built her territory, prioritized by ICP, with no spreadsheets. AI columns ran custom research at scale (which accounts hired a new CRO in the past year, and the background on that hire). Contacts pulled from the database ranked by propensity to engage, auto-enriched with LinkedIn, email, and phone for one-click dialing. Sequences were prompt-built and personalized per persona. On live opportunities, an “AI block” scores deal completion and quality against the sales methodology, flags a missing champion, and coaches the next move (for example, tie a stated conversion-rate gap to revenue to strengthen the business case). On top sits an agent builder for custom background agents.
His warning on the bolt-it-together approach: every niche tool drags in a siloed database, and the context (the conversations, the activities) gets stuck there. Run agents on a fraction of the context and you get “agentic thrash,” low-quality automation or agents that step over each other.
Top 3 Takeaways
- Measure selling time. If you don’t know what percent of rep time is actually in front of customers, that’s your first KPI.
- Context beats integration. Field-level sync isn’t enough; agents need the full conversational graph or they thrash.
- Ramp is the real prize. A founding AE closing $2.7M in 41 days is a ramp-time story as much as a tooling story.
7. Relevance AI’s CEO: From Co-Pilots to Co-Workers
Daniel Vasilev, Co-CEO
Daniel delivered the cleanest maturity model of the day. Relevance is a Series B company split between San Francisco and Sydney, with 100-plus enterprise implementations and agents running millions of tasks a year, and his reframe is that building agents is more like hiring than buying software: you define the role, set the scope, grant tool access, and manage performance. The mismatch he sees everywhere: teams buy a co-pilot but expect a co-worker. Boards say “do more with less, run work autonomously,” then hand reps a productivity assistant and wonder why nothing changed. That gap comes from the approach, not the technology.
He breaks autonomy into four levels. L1 Assist is discrete tasks you delegate. L2 Co-pilot chains playbooks and skills into multi-step work. L3 Autopilot runs agents triggered by events and signals, working without a human kicking off each one. L4 Self-driving targets a specific business metric, tries variants every day, and prunes the losers. Most teams are at L1, some at L2, and very few are at L3, which is exactly where the leverage sits right now.
To deploy L3 you need three things. A documented process, because L3 is not for experimentation; if you can’t teach a human to do it, you can’t teach the agent. A clear owner, because agents are accountable to no one and someone has to own the outcome. And a narrow scope: one process, one trigger, one outcome, proven fast so the business buys in. Good use cases are high-volume, repeatable, with clear inputs and defined outputs, already documented, and measurable against a top-three priority, not a side quest.
On technology, agents need three things. Tools that are vendor- and model-agnostic and plug into everything, with no read or write gaps, so the agent stays as flexible as a human. Excellent instructions, where the job description matters more than it would for a human hire. And governance. Once you have a workforce running, and his example is a demo-qualification agent that escalates the prospects it can’t handle to a second agent, a sales strategist, governance handles security and permissions and becomes the on-ramp to L4, because an agent can’t self-improve without clean data on what worked. The wins he’s seen come from process, not the platform.
Top 3 Takeaways
- Match the expectation to the level. Don’t promise a co-worker and ship a co-pilot.
- L3 needs an owner and a documented process. No accountability, no autonomy.
- Governance is the on-ramp to L4. No clean data on outcomes means no self-improvement.
The Pattern Across All Seven
Seven different categories (infrastructure, AI BDR, CRM, conversation intelligence, AI SDR, full GTM platform, and an agent platform) and roughly the same five messages.
1. Documentation is the new prerequisite. Vercel, Lightfield, Qualified, and Relevance all said a version of the same thing: an agent is only as good as the written process behind it. The model stopped being the bottleneck. What matters now is whether you can describe the job in writing. The teams winning here can write the SOP. The teams stuck can’t.
2. Consolidation is real and it’s coming for your stack. Artisan collapses the outbound stack. Lightfield and Aurasell collapse the CRM stack. Relevance collapses the long tail of point automations. Aurasell’s old reality (22 tools, $3M a year, 11 ops people) is the before picture for a lot of B2B teams in this room.
3. Context is the real moat. Almost everyone is running on Claude or staying model-agnostic on purpose. The differentiation is the unified data and conversation layer feeding the agents. Field-level integration isn’t enough, and partial context produces what Aurasell rightly called agentic thrash.
4. Autonomy is a ladder you climb one rung at a time. Relevance named the rungs (L1 to L4), but every company showed one: Vercel’s headless agents, Artisan’s autonomous reply handling, Lightfield’s natural-language automations, Aurasell’s background agents. The skill is picking your rung honestly instead of pretending you’re three rungs higher than you are.
5. Humans move up the stack, not out. Vercel redeployed its SDR team into higher-impact roles. Lightfield turns data hygienists into closers. Qualified frees human SDRs from the leads they’d never work anyway. It’s the same pattern we run at SaaStr with 3 humans and 20-plus agents: the humans own the judgment, the agents own the volume.
What I’d do Monday morning is simple. Pick one repetitive, high-volume, well-understood process. Write the SOP in painful detail. Assign one owner. Wire up one agent against one trigger. Measure it against a number someone on your exec team already cares about. That’s the whole playbook, and six of these seven sessions said it in their own words.
