← All articles

Build vs. buy for GTM agents

Buy an AI-native CRM that does everything, or build agents on the tools you already have. Building no longer means starting from scratch.

ALL-IN-ONEAI-NATIVE CRMBUYBUILDYOUR STACK

Introduction

In Agent Jobs in GTM we mapped every job an agent can take on in sales, RevOps and customer success. Knowing the jobs is the first step. The second is deciding how to get each one done: buy a product that does it for you, or build it yourself.

Two years ago that was an easy call. Building meant engineers, months and a system you had to maintain. Today most of the parts already exist, and a coding agent can put them together in days. That changes the answer for most teams.

The map, briefly

The map scores every job on two scales. Autonomy is how much of the job the agent owns, in five steps: it answers a question, it drafts something for a person, it watches and flags, it acts inside rules you set, and finally it owns a goal. Difficulty runs from very easy to very hard and reflects how much context the job needs, what a mistake costs and how many people it touches.

Together the two scales give four groups. Leveraged rep, the Leverage Lab, is low on both: the agent summarizes a call or drafts a follow-up and the rep decides. Leveraged team is harder but still low in autonomy, like forecasting or win/loss, where the agent analyses every account and the team decides. Background agents run on their own and flag what a person should act on. Agent-led revenue is high on both: agents reply to buyers, send follow-ups and update the CRM themselves.

The higher a job scores on both scales, the more it depends on your context and your rules. That is where the build or buy decision matters most.

What buying means

Buying means an AI-native CRM like Lightfield, Day AI or Dreamteam. It holds your data, builds the context and runs the agents, all inside one product. You sign up, connect your inbox and calendar, and a lot works on day one.

The cost is the switch. The AI-native CRM becomes your system of record, so you move your data, your team and your processes into it. You get the agents and the rules its makers chose, on their roadmap. If a job you care about isn't on it, you wait.

What building means

Building means you keep the CRM you have and assemble the rest from tools that already do one thing well. Claude writes and runs the agents. OpenRouter gives you one API to every model, so each job can use the model that fits it. Supabase holds your data and runs the small app around it. Deepline builds your lists. HubSpot or Salesforce stays your system of record.

Build or buy?

CRMDataContextAgentsBUY · AI-NATIVE CRMLightfield · Day AI · DreamteamCRMHubSpotSalesforceListsDeeplineDataSupabaseContext graphNorthstarModelsOpenRouterAgentsClaudeBUILD · YOUR STACK

The piece most teams are missing is the context. An agent that drafts a follow-up or flags a quiet deal needs the full customer story: what was said on calls and in email, what was promised and what changed. That is what Northstar, our open-source context graph, provides. It connects your calls, email and CRM into one record per customer, and any agent can read it through MCP, the API or the CLI.

Why start from scratch?

Building no longer means writing everything yourself. Nobody should build their own CRM, model API, database or memory layer in 2026. Those parts exist, they are maintained by people who do nothing else, and most of them are open source or priced by usage.

What you write is the part that is yours: which jobs you hand to agents, the rules they follow, the skill each one runs and where its output lands. That is a few files, not a product. It is also the part an all-in-one product can't know for you.

How to decide, job by job

Don't make the decision once for the whole team. Make it per job, because the answer changes as you move across the map. Four questions settle most cases.

Does a tool already do the job well? If a product ships the job today and does it properly, buying is usually faster than building. Call summaries are the clearest case: Gong and Fathom do them well, and your CRM's own AI is catching up.

What does a mistake cost? In the Leverage Lab a person reads every output before it goes anywhere, so a weak draft costs a minute. In agent-led revenue a mistake reaches a buyer. The higher the cost, the more you want control over the context the agent reads and the rules it follows.

How much of your own context does the job need? Forecasting needs your pipeline. Tracking every promise needs every call and email. Sending a follow-up on its own needs the full history with that customer and your team's sending rules. The more a job depends on context only you have, the less an off-the-shelf agent can do for you.

What would you have to give up? Buying an AI-native CRM to get one job means replacing your system of record for everyone. If you already run HubSpot or Salesforce and the team works in it, building around it is usually the smaller change.

Applied to the map, the pattern is clear. Leveraged rep jobs are mostly worth buying. Leveraged team jobs are mixed: forecasting has strong products like Clari, while testing a hypothesis against past deals has none yet. Background agents and agent-led revenue are mostly worth building, because they depend on your context and your rules, and several of them, like tracking every promise or sending follow-ups on their own, have no dedicated product at all.

The database shows this for every job: what you can buy today and where it stops, and what it takes to build it yourself, with the tools, the steps and the skill.

Summary

Buying an AI-native CRM gets you agents fast, at the cost of moving your whole system of record into one product and its roadmap. Building used to mean starting from scratch. It no longer does: your CRM, a context graph like Northstar and developer tools like Claude, OpenRouter, Supabase and Deepline cover most of the work, and what you write is the part that is yours.

Decide job by job. Buy where a tool already does the job well and a mistake costs little. Build where the job needs your context and your rules, or where nothing on the market does it yet.