Vercel's Internal AI Agents Explained: Lead Agent, d0, Vertex (2026)
What Vercel's Lead Agent, d0, Vertex, Athena and Deal One do, their sourced results, where sources conflict, and how to build similar agents for your team.
Vercel runs a fleet of internal AI agents that handle inbound sales, data questions, support tickets, forecasting and content. This explainer covers each named agent: what it does, its headline result with a source, and how a PM, ops, RevOps or CX leader can build something similar. We use the same format as our breakdown of Ramp's internal AI tools. Every claim about Vercel is attributed. Where sources disagree, we give both figures and say so.
What AI Agents Does Vercel Use Internally? (Quick Answer)
Vercel says it runs “more than a hundred agents in production” (Vercel, June 2026). The best documented are:
- Lead Agent (inbound SDR)
- d0 (data and text-to-SQL)
- Vertex (support)
- Athena (sales cockpit)
- draft0 (content)
- V (orchestrator)
- An abuse-detection agent
SaaStr also reports two sales agents that Vercel has not described itself: Deal One and Playbook Platform.
Across the platform, agents now trigger around 29% of deployments on Vercel, up from under 3% a year earlier [C]. COO Jeanne DeWitt Grosser explains how the company picks what to automate: “If you can document a workflow, it's now pretty straightforward to have an agent do it” (Business Insider via AOL).
Sourcing note: First-party claims come from Vercel's blog and its platform page. Third-party claims from SaaStr, Business Insider and Tomasz Tunguz are attributed by name. Bracketed letters such as [C] refer to the Sources list at the bottom of the page.
Every Vercel AI Agent at a Glance (Summary Table)
Vercel's documented agents cover sales, data, support, RevOps, content, orchestration and trust and safety. The table below lists each one with its most-cited number and the type of source behind it.
| Agent | Team | What it does | Headline number | Source |
|---|---|---|---|---|
| Lead Agent | Sales / SDR | Researches and qualifies inbound leads, drafts replies, routes support questions | “32x ROI” at about $5,000/yr per Vercel. Tunguz reports about $1,000/yr. | [C][E] FP; [J] TP |
| d0 | Data | Answers data questions in Slack (text-to-SQL) using a single bash tool and a Cube semantic layer | 30,000+ questions/month; 45% come from other agents | [C][D] FP |
| Vertex | Support | Resolves tickets across the help center, docs and Slack, and escalates the rest | 87%–93% resolution, depending on the source | [A][C][D][E] FP; [G][I] TP |
| Athena | RevOps | Answers pipeline and forecast questions from Snowflake and Salesforce | Pipeline coverage “nearly doubled” | [C] FP |
| draft0 | Marketing / Content | Turns Slack threads into draft posts and runs a review pipeline on drafts | — | [A][C] FP |
| V | Cross-functional | Routes Slack tasks to the right agent | — | [C][D] FP |
| Abuse detection | Trust & Safety | Flags risky content | — | [A] FP |
| Deal One | Sales (AE) | Uses Gong and Salesforce for call notes, Slack coaching, CRM updates and lost-deal postmortems | — | [G] TP only |
| Playbook Platform | Sales | Matches buying signals to sales plays and drafts outreach | — | [G] TP only |
FP = Vercel first-party. TP = third-party report. Only SaaStr describes Deal One and Playbook Platform.
Lead Agent: Vercel's AI SDR for Inbound Leads
Lead Agent is Vercel's AI SDR. It researches inbound leads, qualifies them, drafts replies and routes support questions away from sales. It is the most-discussed agent in the fleet because its numbers are easy to compare against human headcount.
What it does
According to Business Insider, Lead Agent uses OpenAI Deep Research to investigate each inbound lead, and a manager reviews its work in Slack [H]. SaaStr reports that its inputs include LinkedIn, BuiltWith, the company's website, the CRM and Slack, and that it runs on Vercel's Fluid compute [G].
How it was built
Vercel documented the workflow before building the agent. Starting in June 2025, three engineers shadowed a top-performing SDR for six weeks and recorded how that person researched, qualified and responded to leads [H]. This is Grosser's documented-workflow point in practice: the agent copies a real workflow from a strong performer instead of an imagined ideal.
From a 1,000-line prompt to 14 rules
The first prompt was about 125 lines. It grew to about 1,000 lines as edge cases piled up. Tomasz Tunguz reports that the agent now runs on 14 rules, and a second agent watches for rule breaches [J]. Grosser's summary: “So now actually, our inbound is 14 rules” [J].
The takeaway is to compress a workflow into rules you can enforce. A short rule list can be audited, tested and explained to the person who oversees the agent. A 1,000-line prompt cannot.
What happened to the SDR team
Sources describe the change differently:
- Business Insider: nine of 10 SDRs moved to outbound, and one person oversees the agent [H].
- SaaStr and Tunguz: the team went from 10 people to 1.25 [G][I].
- Vercel's blog: “Vercel's lead qualification agent helps one SDR do the work of 10” [A].
The Business Insider account describes redeployment to outbound, not layoffs.
Cost and ROI (sources conflict)
Vercel says Lead Agent “costs about $5,000 a year to run, returns 32 times that” [C], and its platform page lists “32x ROI at $5,000 a year” [E]. Grosser repeated the figure to Tunguz: “Our sales development agent is a 32x ROI” [I]. By our own arithmetic, not Vercel's, 32x on $5,000 implies about $160,000 a year in return.
Tunguz separately reports that the agent costs “about $1,000 per year” [J]. The sources do not explain the gap. SaaStr adds that maintenance takes about 20% of one engineer's time [G], and neither cost figure clearly includes that labor.
The open-source reference architecture
Vercel publishes a lead-processing agent template and an MIT-licensed repo at vercel-labs/lead-agent. The repo is built on Next.js, AI SDK, Workflow SDK, Slack Bolt and Exa. It describes itself as a “reference architecture.” It is not the internal agent itself.
How to copy it
- Shadow or record your best inbound rep. Write down every decision they make: what they research, what disqualifies a lead, and how they phrase replies.
- Start long, then distill. Let the first prompt grow, then compress it into roughly 10–20 hard rules.
- Add a check before anything is sent. Use a watcher agent, a human approval step in Slack, or both.
- Fork the reference repo instead of starting from scratch.
- Redeploy freed headcount, as Vercel did with outbound. Measure ROI against the fully loaded cost of an SDR, not just salary.
d0: Vercel's Data Agent and the 80% Fewer Tools Lesson
d0 is Vercel's text-to-SQL agent. Employees ask it data questions in Slack, and it writes and runs the queries. According to Vercel's December 2025 post, it ran on Claude Opus 4.5 through the AI SDK, with Sandbox, AI Gateway, Slack Bolt, a Cube semantic layer and a single bash tool [B]. It also uses Workflow and Snowflake [A].
It was built lean. Vercel says one person built it “in a few weeks using 20% of their time” [A].
The before and after numbers
Vercel removed about 80% of d0's tools and published the results [B]:
| Metric | Before | After removing ~80% of tools |
|---|---|---|
| Average time per question | 274.8s | 77.4s |
| Success rate | 80% | 100% |
| Tokens used | ~102k | ~61k |
Why fewer tools worked
A single bash tool on top of a well-modeled semantic layer gives the model one clear way to get an answer. Each narrowly scoped tool adds a decision the model has to make and another way to fail. The semantic layer defines metrics, joins and business logic, which makes the tool list matter much less. Vercel walks through the change in a community meetup [F].
Scale: an agent other agents call
d0 handles “more than 30,000 questions a month” [C]. According to VP of Data Abhi Sivasailam, 45% of those questions come from other agents [D]. That makes d0 shared infrastructure for the rest of the fleet, more like an internal API than a Q&A bot for people.
How to copy it
- Build the semantic layer first. Invest in Cube or dbt metrics before you invest in prompts.
- Start with one general-purpose tool, not twenty.
- Run queries in a sandbox so mistakes stay contained.
- Put it in Slack, where people already ask data questions.
- Design it to be called by other agents from day one.
The eve Data Analyst template is a reasonable starting point. It is not d0 itself.
Vertex: Vercel's AI Support Agent
Vertex is Vercel's support agent. It answers tickets using the help center, docs and Slack, and escalates what it can't resolve [C]. SaaStr reports that three engineers built it in-house in about two months [G].
Its reported resolution rate depends on the source:
- 87% “of initial questions” [A]
- 92% [C]
- 91% [D][E]
- 93% per SaaStr and Tunguz [G][I]
Vercel's own posts put Vertex's resolution rate at 87–92%. SaaStr and Tunguz report 93%. The 87% figure refers specifically to initial questions, which may account for some of the spread.
How to copy it
- Ground answers in your docs and in past resolved tickets.
- Define clear escalation criteria before launch.
- Measure initial-question resolution separately from full-ticket resolution.
- Feed escalated tickets back into your documentation.
Escalated tickets often hold your most useful product signal. Use them to fix recurring issues instead of only deflecting them.
Athena: The RevOps Sales Cockpit Built Without Engineers
Athena answers pipeline and forecast questions using data from Snowflake and Salesforce. Vercel's RevOps team built it “in six weeks without engineers” [C]. Vercel says “pipeline coverage nearly doubled after it went live” [C].
Athena shows that a non-engineering team can ship a production agent when the data layer and agent platform already exist.
How to copy it
- Start with the five forecast questions your CRO asks every week.
- Connect your warehouse and CRM with read-only access.
- Have RevOps own the agent instead of waiting in the engineering queue.
Deal One: AI Deal Coaching From Call Notes (Per SaaStr)
According to SaaStr, Deal One is a sales agent for account executives. Vercel has not described it first-party, so everything in this section comes from SaaStr's report [G].
SaaStr reports that Deal One:
- Pulls data from Gong and Salesforce
- Writes call notes
- Delivers coaching in Slack
- Proposes CRM field updates
- Writes postmortems on lost deals
As described, Deal One connects the sales conversation, the CRM record and rep coaching without manual data entry.
How to copy it
A Deal One-style agent needs three layers:
- Capture every customer call.
- Analyze those calls with insights you can trace back to the source.
- Push structured updates to Salesforce or HubSpot, and to Slack.
BuildBetter covers all three layers:
- Call recording, with or without a bot
- Analysis across every customer call, with citations back to the moment in the conversation
- CRM updates and delivery to Slack
It can also look for lost-deal patterns across calls, support tickets and Slack threads together. That matters because a deal is often lost over a product gap that also shows up in support.
Vercel has not said it uses BuildBetter. This is a suggestion for your own build.
Vercel's Other Agents: Playbook Platform, draft0, V and Abuse Detection
Vercel's smaller documented agents cover sales plays, content, routing and trust and safety.
Playbook Platform
According to SaaStr, Playbook Platform matches buying signals to sales plays and drafts outreach [G]. To copy it, write down three to five plays tied to specific signals before you automate anything.
draft0
draft0 turns Slack threads into draft posts [A] and runs a review pipeline on drafts [C]. To copy it, start with internal threads that already contain an expert explaining something well.
V
V is an orchestrator that routes tasks from Slack to the right agent [C][D]. You only need an orchestrator once you run three or more agents, so don't start here.
Abuse detection
An abuse-detection agent flags risky content [A]. Internal agents handle trust and safety work too, not only go-to-market tasks.
The platform underneath: eve
eve is Vercel's agent framework (GitHub; announcement). Vercel argues that “anyone can build agents but it takes a platform to run them.”
Where to Start If You're Not Vercel
The pattern behind Vercel's agents is consistent: document a real workflow, give the agent clean data and few tools, enforce rules, and keep a human in the loop. Here are five lessons you can apply this quarter.
Lesson 1: Document it first
Every Vercel agent with a detailed origin story started from a documented workflow. Lead Agent began with six weeks of shadowing a top SDR [H]. As Grosser put it earlier, a documented workflow is the starting point. Pick one repeatable workflow that has a clear owner and a measurable output.
Lesson 2: Fewer tools beat more
After removing about 80% of its tools, d0 got faster, hit a 100% success rate and used 37% fewer tokens [B]. Start with one general tool and good data.
Lesson 3: Rules over prompts
Lead Agent went from about 1,000 prompt lines to 14 rules, with a watcher agent checking for breaches [J]. Rules are easier to test and easier to explain.
Lesson 4: Small teams ship
Small teams built all three of these agents:
- d0: one person at 20% of their time [A]
- Athena: RevOps only, no engineers [C]
- Vertex: three engineers in about two months [G]
Data and platform readiness limit you more than headcount.
Lesson 5: Keep a human in Slack
A manager reviews Lead Agent's output in Slack [H], and Vertex escalates what it can't resolve [C]. Put human review where people already work.
Suggested first agent by team
| Team | First agent to build |
|---|---|
| RevOps | An Athena-style pipeline Q&A agent |
| CX | A Vertex-style deflection agent |
| Sales | A Lead Agent forked from the open-source repo |
| Product / PM | A customer-insight agent over calls and tickets |
For PMs, the documented workflow often already exists, spread across customer calls, tickets and Slack. Pulling it into one place is the document-it-first step. For a walkthrough, see how to build a customer insight agent like Ramp. For another company's approach, read how Notion uses internal AI agents.
FAQ: Vercel's AI Agents
What AI agents does Vercel use internally?
Vercel says it runs more than 100 agents in production [C]. The best documented are Lead Agent (inbound SDR), d0 (text-to-SQL data agent), Vertex (support), Athena (RevOps cockpit), draft0 (content), V (orchestrator) and an abuse-detection agent. SaaStr also reports Deal One and Playbook Platform [G].
Is Vercel's lead agent open source?
Partly. Vercel published an MIT-licensed lead-processing agent and a template. The repo calls itself a “reference architecture,” not the exact internal agent.
How was Vercel's Lead Agent built?
Starting in June 2025, three engineers shadowed a top-performing SDR for six weeks to document the workflow [H]. The prompt grew from about 125 lines to about 1,000, then was distilled into 14 rules, with a watcher agent checking for breaches [J].
How much does Vercel's AI SDR cost?
Vercel says Lead Agent costs about $5,000 a year to run and returns 32 times that [C][E]. Tomasz Tunguz reports about $1,000 a year [J]. The sources differ. SaaStr adds that about 20% of one engineer's time goes to maintenance [G].
Did Vercel replace its SDR team with AI?
Not according to Business Insider. It reports that nine of 10 SDRs moved to outbound and one person oversees the agent [H]. SaaStr and Tunguz describe the change as 10 people to 1.25 [G][I], and Vercel says the agent “helps one SDR do the work of 10” [A].
What model does d0 use?
Vercel's December 2025 post says d0 runs on Claude Opus 4.5 through the AI SDK and AI Gateway, with a single bash tool and a Cube semantic layer [B]. The model may have changed since then.
What is Vertex's resolution rate?
Between 87% and 93%, depending on the source and how resolution is measured [A][C][D][E][G][I].
What is Vercel eve?
eve is Vercel's open-source agent framework, announced in June 2026 [C] (GitHub). Vercel also offers an eve Data Analyst template, which is not d0 itself.
Sources
- [A] Vercel blog (Dodds & Grosser), “Anyone can build agents but it takes a platform to run them,” 2026-02-09 — first-party: https://vercel.com/blog/anyone-can-build-agents-but-it-takes-a-platform-to-run-them
- [B] Vercel blog (Andrew Qu), “We removed 80% of our agent's tools,” 2025-12-22 — first-party: https://vercel.com/blog/we-removed-80-percent-of-our-agents-tools
- [C] Vercel blog, “Introducing eve,” 2026-06-17 — first-party: https://vercel.com/blog/introducing-eve
- [D] Vercel Ship 2026 recap, 2026-10-15 — first-party: https://vercel.com/blog/vercel-ship-2026-recap
- [E] Vercel platform page — first-party: https://vercel.com/platform
- [F] Vercel community meetup on d0, 2026-05-26 — first-party: https://community.vercel.com/meetups/how-vercel-removed-80-of-their-agent-s-tools-with-d0/41002
- [G] SaaStr, interview with Jeanne DeWitt Grosser — third-party: https://www.saastr.com/vercel-took-a-10-person-sdr-team-down-to-1-the-whole-thing-costs-5000-a-year-with-vercels-coo-jeanne-dewitt-grosser/
- [H] Business Insider via AOL, 2025-10-27 — third-party: https://www.aol.com/articles/vercel-trained-ai-agent-best-163002264.html
- [I] Tomasz Tunguz, 2026-09-15 — third-party (quotes Grosser): https://tomtunguz.com/single-digit-thousand-dollar-ai-sdr/
- [J] Tomasz Tunguz, 2026-10-06 — third-party: https://tomtunguz.com/how-to-automate-inbound/
Last updated October 2026. We will update this page as Vercel publishes new figures.
Build Your Own Customer-Insight Agent
Vercel's lesson is to document the workflow first, and for product teams that workflow already lives in customer calls, tickets and Slack. BuildBetter captures those conversations in one place and turns them into PRDs and tickets.
