By Shankar Ganapathy, Co-founder, Boomerang AI. Published November 16, 2026.
If you are a GTM engineer in 2026, someone has asked you to build an agent.
It might be an account research agent, a pre-call brief, a "who should we go after this week" digest, or a full outbound workflow. And you can build it. With Claude, an MCP server or two and a weekend, the agent part is no longer the hard part.
The hard part is what the agent knows.
The agent is the cheap layer now
The numbers point one way. In the State of GTM Engineering 2026 survey of 228 practitioners, 71% use AI coding tools, alongside Clay at 84% and their CRM at 92% (report). Retool's 2026 Build vs Buy report found 35% of teams have already replaced at least one SaaS tool with a custom build, and 78% expect to build more internal tools this year (Retool).
And the systems your agent calls are opening up. Salesforce made hosted MCP servers generally available in April 2026, and HubSpot's remote MCP server is generally available with write access to CRM records (Salesforce, HubSpot).
So the prompt, the workflow and the surface are now things you can own. We argued in Sales teams will buy APIs, not tools that this is where the market is heading. This post is the practical follow up: if you are the one building, what should you buy?
The data layer is not cheap
An agent is only as good as its context. Most GTM agents have plenty of one kind and almost none of another.
They have firmographics and contacts: who works where, what the company does, what it raised. Enrichment vendors sell that, and it is easy to call.
They have signals: hiring, funding, intent, website visits. Also easy to call.
What they rarely have is relationships: who on your side already knows the buyer, how well, and who should make the ask. That is the part that decides whether outreach gets a reply, and it is the part you cannot generate with a better prompt.
Relationship data is hard for three reasons:
- It is spread across systems you do not own. Your team's work history, your CRM, calendars, call transcripts, your investors' portfolios and boards, your customers' champions and your partners' account teams.
- It has to be scored, not just listed. A LinkedIn connection is not a relationship. Two people who worked together for four years and met last month are.
- It goes stale. People change jobs constantly. When Armis audited its target contacts, about 30% were no longer in their roles and 25% did not hold the right titles (Armis case study).
You can build a scraper. You cannot build four years of who worked with whom, or your board's relationships, in a sprint.
What to build and what to buy
| Layer | Build or buy | Why |
|---|---|---|
| The surface (Claude, Slack, your internal app) | Build | It should match how your team sells |
| The workflow and prompts | Build | This is your playbook, and it changes weekly |
| System of record | Buy | Keep the CRM boring; build around it |
| Contacts and enrichment | Buy | Commodity data, priced per record |
| Signals and intent | Buy | Needs scale and feedback loops |
| Relationship graph | Buy | Years of data across people you do not employ |
A useful test for each layer: if you rebuilt it, would it be better than what you can buy, or only cheaper this quarter?
A worked example: Digideck
Digideck, a Boomerang AI customer, sells interactive sales presentations to sports and entertainment, a market where everyone knows everyone. It has three new-business reps, and every fall two of them spend most weeks at trade shows.
Jacob Anderson, Digideck's Sr. Growth Marketing Manager, did not buy a new tool for reps. He built a workflow in Claude on top of the Boomerang MCP (case study):
- Ask. Claude queries the Boomerang relationship graph through the MCP for the full customer and prospect base, more than 6,300 relationships.
- Enrich. Claude adds public knowledge, such as ownership groups, parent companies and shared leadership, to find sister companies of existing customers.
- Map. Claude builds an artifact with each rep, each target account and the path Boomerang found.
- Act. Reps pick an account, see which customer knows the buyer, and reach out with what Digideck calls the reverse name drop.
The one rep not at trade shows opened three to four new opportunities within weeks of launch. Digideck now uses the same approach on roughly 40 open deals, finding a customer the quiet buyer knows to restart the conversation.
The split is the point. The workflow, the prompt and the artifact were Digideck's. The relationship data came from Boomerang.
How to wire a relationship graph into your agent
If you are building, a few things make the difference:
- Call relationships at the account step, not at the end. Use paths to decide which accounts to work, not just how to open them.
- Ask for the sender, not just the path. The useful answer is "this board member should ask, and here is why," not "you have 14 connections."
- Keep the evidence. Pass through why a path is strong (worked together, recent meetings) so the rep, and the agent, can judge it.
- Never send on someone's behalf. Draft the ask for the connector to approve and send themselves. Relationships do not survive automation that skips the human.
- Refresh on job changes. A champion moving to a new company is a new path, and an old path is now a dead one.
Where Boomerang AI fits
Boomerang AI is the relationship layer. It maps warm paths from your team, investors and board, customers and partners into each target account, scores them by evidence and recommends who should make the ask. Through the Boomerang MCP and API, your agent can ask Rudy who can get you to a buyer, who should send the ask and how strong the path is, from Claude, Clay or your own app.
MCP access comes with the Team, Company and Enterprise plans, and the REST API with Company and Enterprise (MCP and API).
Build the agent. Buy the graph.
Read next: The missing step in your Claude Code GTM workflow, What is a GTM context graph? and Sales teams will buy APIs, not tools.
Sources
- State of GTM Engineering 2026 (228 practitioners): https://stateofgtme.com/
- Retool, 2026 Build vs Buy report: https://retool.com/blog/ai-build-vs-buy-report-2026
- Salesforce hosted MCP servers GA, April 2026: https://developer.salesforce.com/blogs/2026/04/salesforce-hosted-mcp-servers-are-now-generally-available
- HubSpot remote MCP server GA: https://developers.hubspot.com/changelog/remote-hubspot-mcp-server-is-now-generally-available
- Digideck case study: https://www.getboomerang.ai/customers/digideck-claude-boomerang-mcp-abm
- Armis case study: https://www.getboomerang.ai/customers/scaling-smarter-inside-armis-agentic-ai-playbook-for-sustainable-pipeline-growth
Frequently asked questions
Should GTM engineers build or buy AI agents?
Usually both. Build the surface and workflow your team uses, because that is your playbook. Buy the data layers that depend on scale or years of collection, such as your CRM, contact data, signals and relationship data.
What is a relationship graph in sales?
A map of who on your side, including your team, investors and board, customers and partners, knows people at target accounts, scored by how strong each relationship is. Agents use it to find warm paths and pick the right person to make an introduction.
Can I connect Boomerang to Claude?
Yes. The Boomerang MCP server lets Claude, Claude Code or ChatGPT query warm paths, recommended senders and path strength from your workspace, read only and over OAuth. MCP access comes with the Team, Company and Enterprise plans.
Why not scrape LinkedIn connections instead?
A connection list shows who accepted a request, not who knows whom. It misses former colleagues who never connected, board and investor relationships, and customer champions, and it does not tell you how strong a relationship is or whether it is current.
What did Digideck build with the Boomerang MCP?
A Claude workflow that queried more than 6,300 customer and prospect relationships, added public ownership data to find sister companies, and gave each rep an artifact showing which customer knows each buyer. One rep opened three to four new opportunities within weeks.




