AI Engineer World's Fair 2026
GTM Engineering: The Technical Bits — Everett Berry, Clay
Loading slides…
Read the talk
GTM Engineering: The Technical Bits
Everett Berry of Clay traces the engineering beneath fast-moving go-to-market teams: mutable account data, asynchronous tool orchestration, persistent agents, and customer outreach where small automation mistakes can damage real relationships.
From a talk by Everett Berry
At a glance
Ideas worth remembering
GTM engineering treats changes to data, automations, and campaigns as shipped systems that need release cadence, iteration, and operational ownership.
A usable account model combines first-party activity, third-party enrichment, entity resolution, selective refresh policies, and evaluations of competing data providers.
Orchestration must model asynchronous dependencies explicitly: a successful write in one tool does not mean a downstream tool is ready, so waits, checks, retries, fan-out, and joins belong in the control flow.
A long-running GTM agent can be assigned per account, remain dormant between triggers, reload current context, and preserve state across a deal cycle—but continual learning and next-best-action selection remain unresolved.
Customer-facing execution needs domain-reputation controls, reply routing, and cross-channel suppression because locally correct automations can combine into a harmful customer experience.
The final coordination problem is human: reps need to see what agents did, understand why, disagree when necessary, and arrive at customer conversations with the same account context as the automated system.
GTM becomes a shipping discipline
GTM engineering starts from a change in cadence. At Clay, Berry describes shipping new data, automations, and campaigns every two weeks, then iterating on them much like an engineering team iterates on product releases. The role exists to remove the technical constraints that historically kept go-to-market teams from moving that quickly.
More capable AI models increase what a GTM engineer can automate, but models are only one part of the system. Berry organizes the underlying work into four connected areas: data, orchestration, agents, and execution. Each layer depends on the previous one. An agent cannot choose a sensible action if the account record is stale, and a correct decision still fails if the outreach tools execute it out of order.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The account record is a changing model of the market
The data layer aims for what Berry calls a “perfect virtual copy” of the market: the target companies, the people inside them, and the current state of each relationship. Perfection is an aspiration rather than a stable endpoint. Companies get acquired, open offices, launch products, hire, and fire. Meanwhile, marketing, sales activity, meetings, expansion, and churn change the account’s relationship with the seller.
A new CRM therefore needs more than account names. Relevant contacts come first, followed by third-party account hierarchy, firmographics such as company size, technographics describing the software in use, and signals that suggest an account may be in market. Operational fields then record whether the account is a customer, expanding, churning, or scoring highly enough to pursue.
No single vendor supplies every field with adequate coverage. Berry’s concrete example is phone-number enrichment across a set of countries: using Forager alone filled only about half the desired records, so the workflow queried additional providers to fill the gaps. This technique, waterfalling, increases coverage by sending unresolved records to the next source rather than accepting the first provider’s omissions.
What does waterfalling make visible? The diagram shows that provider order is a decision rule, not merely a larger vendor list. Each source handles records still missing a value, while evaluations are needed to decide which providers deserve priority and which returned values are accurate enough to keep.
Waterfalling unresolved account data across providers
Read the diagram as text
- Accounts missing phone numbers. data-layer — The target field is empty for a set of country-specific records.
- Forager lookup. providers — The first provider fills roughly half of the example set.
- Still-unresolved records. data-layer — Records without a result continue through the waterfall.
- Additional providers. providers — Other sources attempt to fill the remaining values.
- Provider evaluations. quality — Coverage and accuracy determine which sources to trust and prioritize.
- Enriched account record. data-layer — Accepted values become part of the GTM data layer.
- Accounts missing phone numbers → Forager lookup: query first source.
- Forager lookup → Still-unresolved records: pass missing results onward.
- Still-unresolved records → Additional providers: query next sources.
- Additional providers → Provider evaluations: compare returned data.
- Provider evaluations → Enriched account record: accept trusted values.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Freshness has a cost, so fields need different policies
Coverage is only the first data problem. The layer must merge purchased third-party information with first-party activity and incoming signals, then resolve different representations of the same company across sources. Without entity resolution, one account can fragment into several records and the automation loses a coherent view of its state.
Continuous refresh is expensive when each update purchases data, so refreshing every field on every run is wasteful. The system needs field-specific policies: volatile facts should update more often, stable facts less often, and pushed signals should update the account when events arrive. This selective maintenance is what makes the virtual market model economically usable rather than literally perfect.
The practical consequence is severe: automated GTM plays inherit the quality and timing of this layer. A workflow can run exactly as designed and still target the wrong company, contact, or moment if its underlying entities are mismatched or stale.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Orchestration is a distributed data-engineering problem
Once activity spreads across a CRM, warehouse, sequencer, dialer, call recorder, and chat interface, each tool develops its own partial view of the account. Berry commonly encounters teams working through ten, twenty, or thirty tools. Orchestration keeps those views coordinated as emails, calls, account events, and rep actions flow back into the data layer.
The systems do not share one workload shape. One may need a single record immediately, while another accepts hundreds of thousands of records in a nightly batch. Employee count changes frequently; headquarters location changes rarely. A single input may need to fan out to several destinations, and failures become normal once that work is distributed.
The most instructive failure appears when tools synchronize behind the orchestrator’s back. Suppose an orchestration workflow creates a contact in Salesforce and then wants to enroll that person in Outreach or another sequencer. Salesforce and the sequencer have their own asynchronous sync. The contact is observably present in the CRM but not yet available in the sequencer, so immediate enrollment fails. The workflow must wait, check again, and loop until the downstream record exists.
Where does the delay enter? The diagram exposes two control paths: the orchestrator writes to the CRM, but a separate connector controls when the sequencer sees that write. A successful CRM operation therefore does not prove downstream readiness; the loop must test the destination’s state before acting.
Why creating a CRM contact cannot trigger immediate sequencing
Read the diagram as text
- Workflow trigger. orchestrator — An event or schedule starts the orchestration.
- Create contact in Salesforce. crm — The CRM accepts the new contact record.
- Independent CRM–sequencer sync. connector — A separate integration propagates the contact asynchronously.
- Check sequencer readiness. orchestrator — The workflow tests whether the contact has appeared downstream.
- Wait and retry. orchestrator — If the contact is absent, pause before checking again.
- Enroll contact in sequence. sequencer — Outreach begins only after the downstream record exists.
- Workflow trigger → Create contact in Salesforce: start contact creation.
- Create contact in Salesforce → Independent CRM–sequencer sync: record awaits propagation.
- Independent CRM–sequencer sync → Check sequencer readiness: contact may become visible.
- Check sequencer readiness → Wait and retry: not ready.
- Wait and retry → Check sequencer readiness: loop.
- Check sequencer readiness → Enroll contact in sequence: ready.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
General-purpose graph nodes encode the control flow
Clay’s response is a graph of general-purpose nodes rather than a chain of product-specific integrations. Different nodes run agents, call tools, evaluate conditions, execute code, or perform a map-reduce-style fan-out and aggregation. A trigger or schedule starts the graph; branches query several systems; their results come back together; and the combined context moves into the interfaces used by reps.
This graph does not remove distributed-system failures. It gives waits, retries, conditionals, fan-out, and joins an explicit place in the design. Berry presents the pattern as applicable whether the orchestration layer is bought or built: the important choice is to model control flow and data dependencies instead of assuming every integration behaves like a synchronous function call.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
One persistent agent follows each account over time
With data and orchestration in place, the account state now includes signals, changing fields, rep and agent actions, meetings, and feedback. Sales representatives often sort through this context manually. Berry’s proposed architecture assigns one agent to each account so the reasoning layer can maintain persistent account state across a deal cycle lasting weeks or months.
A long-lived account agent is dormant most of the time. A schedule, smart trigger, or heartbeat wakes it; it ingests the latest context from the data and orchestration layers; it reasons about the next action; and it records feedback for later decisions. The design must also map unstructured model output into structured CRM fields, where an incorrect value can affect later automation and customer communication.
What must persist while the agent sleeps? The lifecycle diagram makes clear that persistence belongs to the account state, not to a continuously running model call. Triggers restart computation around the latest account context, and feedback returns to the state used by later wake-ups.
Lifecycle of a long-running account agent
Read the diagram as text
- Dormant account agent. agent — The agent exists across a weeks- or months-long deal cycle but is inactive most of the time.
- Trigger or heartbeat. control — A schedule or meaningful account event wakes the agent.
- Ingest current account context. state — Load data, signals, meetings, and recent actions through the data and orchestration layers.
- Reason about the next action. agent — The agent uses the assembled context to make a decision.
- Write agent state and actions. state — Agent-specific CRM values preserve outputs without overwriting human or deterministic fields.
- Collect feedback. learning — Results and human responses inform later decisions.
- Dormant account agent → Trigger or heartbeat: event or schedule arrives.
- Trigger or heartbeat → Ingest current account context: reload latest state.
- Ingest current account context → Reason about the next action: supply decision context.
- Reason about the next action → Write agent state and actions: record decision or action.
- Write agent state and actions → Collect feedback: observe outcome.
- Collect feedback → Dormant account agent: retain state until next trigger.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Agent-written fields need clear ownership
Berry recommends separating CRM fields updated by agents from fields updated by deterministic systems or people. That separation reveals who produced a value, prevents competing writers from silently overwriting one another, and gives feedback or review workflows a specific agent output to inspect.
A closed-lost reawakening agent shows why timing belongs in the architecture. It combines call information, email, CRM data, and warehouse context, but it does not immediately pursue an account after a loss. A time-based trigger introduces a deliberate cooling-off period before the agent reassesses the opportunity.
The unresolved frontier is continual learning and reliable next-best-action selection. Berry says GTM has not fully solved how an agent should update its own view of what works across accounts. Persistent state and feedback create the prerequisites for learning, but they do not establish that autonomous updates will improve decisions safely.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Execution turns system errors into human costs
The final layer sends the message, places the call, or coordinates another customer-facing action. This is where a technically valid decision becomes consequential. Berry’s displayed figures put cold-email replies around 0.5% to 1%—perhaps one reply from one hundred sequenced contacts—and describe LinkedIn as three to four times more effective than cold email. These figures come from charts discussed in the talk; Berry himself was uncertain that the cited email sample covered twenty million messages and said some displayed channel rates looked elevated, so the numbers support the narrow point that response margins are small rather than a precise universal benchmark.
Small margins raise the cost of bad agent behavior. If automated outreach says the wrong thing, it can lose the rare response the campaign might have earned. If it sends directly from a rep’s company address and recipients ignore or reject the campaign, the company’s domain reputation can suffer.
One workaround spreads outreach across multiple domains. That protects the primary domain from carrying all campaign risk, but creates a routing problem: replies arriving on alternate domains must return to the rep’s main inbox or working interface. Clay also uses a rep-proxy arrangement for some accounts, connecting multiple rep inboxes to a sequencer and sending on their behalf; other accounts use the multi-domain pattern.
Multichannel execution adds cross-channel suppression. When a call connects and books a meeting, the system should stop the email sequence and may need to remove the person from a lifecycle marketing campaign. Otherwise, individually successful tools create a collectively absurd customer experience: a booked prospect keeps receiving automated requests to book.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The hardest rollout problem is the human–agent interface
The closing Q&A moves the bottleneck from infrastructure to coordination. Agents can provide the reasoning and decision layer for work previously performed by sales reps, but the rep may disagree with an action or may not know it happened. Because a person still joins the eventual prospect call, the automation must keep that person informed enough to understand the account and intervene when needed.
That interface needs more than a stream of agent outputs. It must make actions visible, reconcile disagreement, and coordinate ownership between automated systems and the rep. Otherwise, adding a capable decision layer merely creates a second actor with a different view of the account—the same inconsistency orchestration was meant to eliminate among tools.
The discipline also covers inbound GTM, not only outbound campaigns. Inbound systems must route a lead correctly, qualify the account, and recover relevant historical context before deciding what happens next. The four-layer model therefore applies across GTM: maintain the account model, coordinate system state, reason over context, and execute while keeping the responsible human in the loop.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
A practical companion explaining Clay’s two-week sprints, core internal stack, release notes, team structure, and examples of the workflows its GTM engineers maintain.
Related talks
- The Building Blocks of GTM Orchestration — Arman Vaziri, Ramp
Extends the orchestration layer with durable execution, searchable unstructured information, reusable skills, and pre-meeting briefs.
- Build the AI GTM Agent That Knows the Buyer Before the First Message
Develops the buyer-context problem using CRM records, conversations, job changes, LinkedIn activity, and website signals.
- 12-Factor Agents: Patterns of reliable LLM applications
Provides a production-oriented treatment of explicit control flow, structured tool calls, deterministic software, and human contact around agents.
Full transcript
Loading transcript…