AI Engineer World's Fair 2026
The Building Blocks of GTM Orchestration — Arman Vaziri, Ramp
Loading slides…
Read the talk
The Building Blocks of GTM Orchestration
Arman Vaziri explains how Ramp turns go-to-market intent into coordinated work by combining a shared customer-data layer, durable agent execution, searchable company knowledge, reusable skills, and human approval.
From a talk by Arman Vaziri
At a glance
Ideas worth remembering
Start with a specific operational bottleneck, then reuse its data mappings, tools, and procedures across adjacent teams.
A shared customer data layer must preserve entity relationships and provenance while making unstructured interactions searchable.
Durable execution turns tool and model calls into resumable activities, preventing a failed worker from restarting a long agent workflow from the beginning.
Retrieval, shared skills, and user instructions solve different problems: finding relevant facts, encoding organizational procedure, and shaping the output for individual adoption.
Multi-channel orchestration should route shared intent through proven vertical workflows, keep audience targeting consistent, and preserve human approval and policy guardrails.
The hard part begins after the campaign idea
Arman Vaziri, who leads product- and sales-led growth engineering at Ramp, defines go-to-market orchestration as the ability to describe a motion—a playbook, experiment, or evergreen campaign—and distribute its execution across outbound, advertising, web, and other channels. The desired interface begins with intent rather than a collection of channel-specific operating procedures. 0:13
The running example sounds simple: offer Pro V1 golf balls to golfers at East Coast construction companies. Operationally, that sentence implies much more. The system must identify the audience, define the incentive, write personalized outbound sequences, generate paid and web creative, possibly add in-app notifications, and keep the targeting consistent across those outputs. Vaziri concedes that a real description would require more than one sentence, but the user should specify the motion rather than manually assemble every artifact. 1:43
Three bottlenecks made that goal difficult:
- Fragmented data: CRM, product, and channel systems held inconsistent versions of the customer, making coordinated targeting unreliable.
- Operator busywork: Sales representatives already spent their days in meetings, outbounding, and selling; campaign administration competed with that work.
- Distribution cost: A good idea still required a proposal, enablement material, and persuasion across several teams. That coordination often operated on a timescale of months.
Ramp’s sequence was therefore bottom-up: improve ingestion and data consistency, build a narrow workflow that saves one team time or improves performance, reuse its patterns for adjacent teams, and only then coordinate several channels from a shared description of intent. Agents expand how much can be automated near each execution point, but they depend on the company context and operational foundations beneath them. 4:13
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A customer data platform gives every agent the same entities
A go-to-market agent may need to distinguish a prospect with almost no known information from an active customer with product history and support problems. Ramp addressed that breadth by building an internal customer data platform, or CDP. Its inputs include CRM records, product activity, enrichment and web data, internally modeled buying signals, external events such as funding announcements, and interactions including emails, meetings, calls, and page views. 5:26
Real-time events such as emails can enter through Kafka and be consumed into a Postgres-backed online layer. Postgres supplies transactional guarantees and referential integrity across entities originating in the CRM, product, and third-party systems. The platform also records provenance metadata such as where a value came from and when it arrived. That identity layer matters because an agent cannot safely coordinate a campaign if one system’s company, contact, and interaction records cannot be attributed to the corresponding entities elsewhere.
Structured rows are only half the material. Calls, emails, and notes contain valuable details that do not fit neatly into columns, so Ramp embeds them for search. Online batch jobs call external APIs to pre-ingest enrichment about companies Ramp can sell to, while offline jobs use dbt and Snowflake for warehouse computation before piping results back into the shared serving layer.
What does this foundation make visible? The diagram separates ingestion speed from storage form: real-time and batch paths converge on shared customer entities, while structured relationships and searchable text remain optimized for different jobs. Agents can then retrieve both without inventing their own joins or rebuilding identity resolution for every workflow.
Customer signals converge on a shared agent data layer
Read the diagram as text
- Customer and market sources. inputs — CRM, product, enrichment, web, buying signals, email, meetings, calls, and page views.
- Real-time ingestion. ingestion — Events such as emails flow through Kafka consumers.
- Online and offline batch. ingestion — API enrichment plus dbt and Snowflake warehouse computation.
- Postgres customer layer. storage — Maintains transactions, entity relationships, attribution, and source metadata.
- Embedded unstructured data. storage — Makes calls, emails, and notes searchable.
- GTM agents and workflows. execution — Consume consistent entities and relevant company context.
- Customer and market sources → Real-time ingestion: streaming signals.
- Customer and market sources → Online and offline batch: enrichment and warehouse data.
- Real-time ingestion → Postgres customer layer: consume and attribute.
- Online and offline batch → Postgres customer layer: pipe computed data back.
- Customer and market sources → Embedded unstructured data: chunk and embed text.
- Postgres customer layer → GTM agents and workflows: structured context.
- Embedded unstructured data → GTM agents and workflows: retrieved text context.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Pre-meeting briefs expose the difficult joins and failure modes
Rather than begin with universal orchestration, Ramp solves one team’s workflow and then scales sideways. Automated outbound and meeting preparation recur across teams, while work such as QBR generation may remain team-specific. The first implementation can therefore solve a concrete operational burden while revealing which data mappings, tools, and procedures are reusable. 8:08
Pre-meeting briefs for account managers provide the concrete example. Before a customer conversation, the brief gathers who is attending, what the meeting concerns, the account manager’s goal, product usage, account vitals, the agenda, open support tickets, and any email indicating what the customer wants to discuss. The observable change is simple: instead of assembling this context between back-to-back meetings, the account manager receives one prepared artifact.
Producing that artifact begins with a meeting event, then hydrates it with other data and maps attendee emails and meeting titles to an account. This is a “sneaky hard problem” at Ramp because the same person can work on behalf of several businesses. The mapping is therefore fuzzy rather than a dependable one-email-to-one-account lookup. Once resolved, Ramp persists the association so every downstream workflow does not repeat the same expensive and potentially inconsistent matching.
The remaining work runs as a durable thread built on Temporal. Each tool call and model call becomes an activity. If a worker dies partway through gathering account context, execution resumes from the accumulated state instead of repeating the entire thread. Tool access can be scoped by agent configuration, and human-in-the-loop tooling can pause a run, collect input, and resume it. 10:08
Where does durability change the workflow? The diagram shows that identity resolution and completed activities become reusable state. A crash interrupts the worker, not the logical thread; a human decision also becomes a controlled pause rather than an abandoned run.
A pre-meeting brief survives interruptions
Read the diagram as text
- Meeting event. trigger — Supplies timing, title, and attendees.
- Account resolution. identity — Fuzzy-matches attendee emails and meeting titles to the relevant account.
- Durable Temporal thread. execution — Tracks accumulated state across tool and model activities.
- Context-gathering activities. execution — Collect product usage, vitals, tickets, email context, and agenda material.
- Optional human input. control — Pauses execution, records input, and resumes.
- Prepared meeting brief. output — Presents the assembled account and meeting context to the account manager.
- Replacement worker. recovery — Resumes from persisted state after a worker failure.
- Meeting event → Account resolution: hydrate and match.
- Account resolution → Durable Temporal thread: persist resolved account.
- Durable Temporal thread → Context-gathering activities: schedule as activities.
- Context-gathering activities → Optional human input: pause when review is required.
- Optional human input → Prepared meeting brief: resume with input.
- Context-gathering activities → Prepared meeting brief: assemble context.
- Durable Temporal thread → Replacement worker: recover accumulated state.
- Replacement worker → Context-gathering activities: continue unfinished work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Retrieve only the relevant corpus, then let users shape the output
Meeting transcripts, emails, enablement materials, product knowledge, and playbooks contain much of the context needed for the brief. Ramp combines real-time ingestion with batch jobs that chunk and embed these materials in a vector database. Agents search them with a mixture of vector, attribute, and keyword retrieval, scoped to an account or current question. This avoids placing the entire raw corpus into model context, which would be expensive and inefficient. 11:16
Retrieval supplies facts, but it does not decide what a useful brief looks like for every account manager. Ramp therefore built a skill library where users express their preferred format and information in text. The background agent combines system-owned meeting-preparation skills with those custom instructions. Vaziri identifies this customization as particularly valuable for adoption: people could preserve the shape of the work they already found useful instead of accepting one centrally imposed brief.
Every night, the system fans out background agents to compute preparation per account. Each run can use the online Postgres CDP, the vector database, system-level meeting skills, and user-authored instructions. The layers have distinct jobs: structured storage resolves entities and current state; retrieval finds relevant text; shared skills encode the organization’s procedure; user instructions define the individual’s preferred result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Vertical workflows become multi-channel orchestration
Once meeting preparation works, adjacent workflows become cheaper to build. A post-meeting agent can use the transcript to draft a follow-up or pre-fill a CRM opportunity after detecting a possible expansion discussion, then ask a representative for a thumbs-up before creating it. The shared foundation does not remove approval; it moves preparation and data entry ahead of the approval point. 13:03
Horizontal reuse still requires adaptation. Email embeddings, call transcripts, and custom instructions may generalize across teams, but account executives working on pre-sales opportunities need more third-party prospect data and less existing-product usage data than account managers. Extending the system therefore means adding the right ingestion, integrations, and skills—not merely pointing a new team at the old prompt.
Ramp also exposes the background agents’ tools and skills through its internal GTM MCP, allowing employees to chat with the same data and build their own automations. Those experiments double as demand signals. When an employee connects to the tools and creates a prompt, skill, or small application, they reveal a real problem and a candidate solution that the central team can productionize for others with similar needs.
This returns to the golf-ball campaign. The earlier vertical systems become execution endpoints inside Ramp Revenue: audience construction for SDRs, personalized copy and outbound sequences, landing pages, images, and other creative. Channel owners review and sign off before release. Vaziri says the Pro V1 experiment “works really well,” but the recording supplies no metric, comparison, or experimental conditions, so that remains an anecdotal internal result rather than a quantified effectiveness claim. 15:42
What turns one intent into several coordinated artifacts? The diagram shows a shared audience and campaign context feeding established vertical workflows rather than one unconstrained agent improvising every channel. Human owners remain at the release boundary, while the common substrate keeps targeting and campaign context consistent.
The intended outcome is faster and safer shipping across channels, with room to balance new experiments against known returns. Guardrails can enforce compliance, rules of engagement, context-sensitive exclusions, and repetition limits. These are necessary because coordinating more channels also increases the chance of contacting the same person repeatedly or executing an action that conflicts with policy.
Campaign intent fans out through reusable vertical workflows
Read the diagram as text
- Campaign intent. input — Offer Pro V1 golf balls to golfers at East Coast construction companies.
- Ramp Revenue. orchestration — Internal application that routes intent into established vertical solutions.
- Shared target audience. shared-context — Identifies the golfers to contact with consistent targeting.
- Outbound sequence. channels — Produces personalized copy and sequences for SDRs.
- Landing page. channels — Creates a destination for campaign traffic.
- Images and creative. channels — Produces campaign assets for web and paid channels.
- Channel-owner review. control — People responsible for each channel inspect and approve its output.
- Coordinated campaign. output — Runs approved artifacts with common targeting and guardrails.
- Campaign intent → Ramp Revenue: describe the motion.
- Ramp Revenue → Shared target audience: resolve target segment.
- Shared target audience → Outbound sequence: personalize contacts.
- Shared target audience → Landing page: keep targeting aligned.
- Shared target audience → Images and creative: supply campaign context.
- Outbound sequence → Channel-owner review: submit for sign-off.
- Landing page → Channel-owner review: submit for sign-off.
- Images and creative → Channel-owner review: submit for sign-off.
- Channel-owner review → Coordinated campaign: approve under guardrails.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Smaller teams should build the first useful vertical slice
The closing question asks where a smaller company should begin. Vaziri’s answer is deliberately narrower than the architecture just presented: find a specific problem that can be automated and solve it first. Ramp’s own earlier effort involved two people building automated outbound—pulling available data, generating personalized copy, and placing it into sequences. That work created value while teaching the team how the mechanism might extend elsewhere. 18:32
A smaller team does not get a year to design a perfect orchestration architecture. The practical strategy is to build vertical solutions around real work, learn which data and execution patterns recur, and connect those solutions afterward. The final advice matches the architecture’s history: orchestration is assembled from proven workflows, not designed in the abstract before anyone receives a useful result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Related talks
- Building Durable, Production-Ready Agents with OpenAI SDK and Temporal
A practical companion for the durable-execution mechanism used here, including persisted model calls, worker recovery, state management, and activity-based tools.
- Don't Build Agents, Build Skills Instead
Develops the reusable-skill model that helps Ramp separate shared procedures from user-specific instructions and tool connectivity.
- AI in GTM at Notion — Flora Liu
Offers a complementary GTM architecture built around shared customer context, signal-driven Temporal workflows, human approval, and outcome feedback.
Full transcript
Loading transcript…