It's planning season, and a lot of RevOps leaders are asking a reasonable question: why not just have Claude build our territory plan?
It's a fair question. You can export your accounts from Salesforce, drop the CSV into a chat window, describe your rules, and get back something that looks like a territory plan in a few minutes. We've talked to RevOps leaders this fall who got a working territory sheet out of Claude in about 12 minutes. That first result feels like magic.
We build AI agents for a living, and we're big believers in what they can do for GTM teams. But we've also watched a lot of teams try the DIY route this year. The first draft is the easy part. Getting from that draft to a plan that's balanced, explainable, agreed upon, live in your CRM, and still accurate in March is where a general-purpose agent starts to struggle.
This post walks through where a homegrown agent helps, where it breaks down, and why we think the smartest approach is to let your agent drive and let Carve do the territory work.
What a general-purpose agent gets right
Let's give the DIY approach its due. A capable agent is genuinely good at a few parts of territory planning:
- Turning fuzzy requirements into rules. You can say "keep strategic accounts with their current rep, balance on revenue potential, and give new hires lighter books" and the agent will understand you.
- Fast first drafts. It can group, sort, and assign a few thousand rows faster than any analyst with a pivot table.
- Exploration. It's great for asking "what if we split mid-market by vertical?" and seeing a rough answer before lunch.
- Flexibility. When requirements change mid-cycle, you can just rerun it. Some teams like working this way, in quick one-off bursts, rather than maintaining a system.
That's real value, and it's why we built Carve around an agent in the first place. Nobody should be hand-carving territories in spreadsheets anymore. The question isn't whether to use AI for territory planning. It's which parts of the job a general-purpose agent should own, and which parts need something purpose-built underneath it.
Where DIY agent territory planning breaks down
Territory planning isn't one task. It's a cycle: design, feedback, deployment, and adjustment. A chat window handles the first step well. The trouble shows up in everything after it.
Results you can't reproduce. Ask a general-purpose agent to carve the same territories twice and you may get two different plans. That's fine for brainstorming. It's not fine when a rep asks why they lost an account and the honest answer is "the model felt differently on Tuesday." Balancing targets like $2M of potential per rep within ±5% need math that runs the same way every time, with a reason recorded for every assignment.
Data that's bigger and messier than a chat can hold. Real planning means tens of thousands of accounts, parent-child hierarchies, duplicates, open opportunities that shouldn't move, and fit scores that need a hard ICP check before ranking. Agents working from a CSV tend to quietly drop rows, summarize instead of assign, or miss constraints buried halfway down the prompt.
Feedback that lives everywhere. Building a scenario takes an afternoon. Getting agreement takes weeks. With a DIY agent, manager feedback still arrives through Slack, email, and meetings, and someone in RevOps has to re-enter every override by hand. You end up with 8,000 revisions of a spreadsheet and no single source of truth. Most homegrown agents are single-player: one person, one chat, one laptop. Territory planning is a team sport.
Getting it into Salesforce. A plan in a chat window doesn't change anyone's book. Someone still has to export a CSV, run Data Loader, pause automations, update routing rules, and hope nothing needs rolling back. General-purpose agents also don't know your org's Salesforce quirks, like the routing conditions and ownership rules that only make sense to whoever set them up. That last mile is where many plans stall and where mistakes get expensive.
Plans that go stale. Reps leave, accounts move, strategy shifts. A one-time agent session doesn't know any of that happened. Unless someone rebuilds the prompt, re-exports the data, and reruns it all, the plan starts decaying the day it ships.
Upkeep that never ends. The prompt, the data pulls, the scoring logic, and the deploy scripts all become a system someone has to own. That usually lands on the same RevOps analyst you were trying to free up. And the "one-off burst" approach has a hidden cost: every burst starts from scratch, with no memory of last quarter's overrides or why an account moved.
Build at the surface, buy the infrastructure
We've written before about a simple rule for the AI era: build at the surface, buy the infrastructure. Build the parts that reflect how your business uniquely works, are highly variable, and are cheap to get wrong. Buy the parts where availability, performance, and risk matter.
AI has made specialized knowledge cheap. Your agent can explain territory design theory as well as a consultant. What AI hasn't solved is the rest of the infrastructure test: running reliably at scale, producing the same answer every time, and not wrecking account ownership for 60 reps when something goes wrong.
Territory planning has both layers:
- The surface is yours. Your segmentation philosophy, your balancing priorities, your rules of engagement, and the way you present the plan to your CRO. That's strategy, and it should be shaped by your team and your agent.
- The infrastructure should be bought. The assignment engine, the scenario versioning, the feedback and override tracking, the Salesforce deployment, and the ongoing rebalancing. That's where consistency, scale, and risk live.
So the choice isn't "agent or Carve." It's your agent on the surface, with Carve as the territory engine underneath.
What Carve does that a homegrown agent can't
Carve is a territory-planning agent built for the whole cycle, not just the first draft. You still describe what you want in plain language. The difference is everything that happens after that.
| Planning step | DIY agent | Carve |
|---|---|---|
| Design | One plan per prompt, hard to reproduce | Multiple scenarios in minutes, compared side by side |
| Explain | "The model decided" | A recorded reason for every assignment, plus AI-written narrative cards |
| Analyze | Ask for charts, one at a time | An analysis dashboard generated with every scenario |
| Feedback | Slack, email, and manual re-entry | Comments and overrides inside the scenario, kept across versions |
| Deploy | CSV exports and Data Loader | One-click deployment of ownership to Salesforce |
| Adjust | Rebuild from scratch | Bookbuilder keeps books balanced as reps and accounts change |
| Inputs | Whatever's in the CSV | Market Map fit scores and customer signals |
A few of these deserve a closer look.
Feedback becomes part of the plan. Sales leaders can move an account to a different rep directly in the scenario. Every override opens a thread showing what Carve originally proposed, so RevOps can keep it or revert it. The agent then applies feedback and sends a new version back out for review. That two-week back-and-forth on new-hire books shrinks to a few rounds in one place.
Deployment is built in. Once a plan is approved, Carve updates account ownership in Salesforce directly. No CSV handoffs, no toggling automations, no risky rollbacks.
The plan stays alive. Carve hands off to Bookbuilder, which keeps rep books balanced and current through the year. A rep leaves on Tuesday and their accounts are redistributed the same day, instead of going cold until next year's carve.
Your agent can still drive. This is the part DIY builders tend to miss. RevOps teams increasingly expect every tool to work like the agent they already use every day, and Carve does. Through the Gradient Works MCP, you can run the whole process from Claude, ChatGPT, Gemini, or Cursor. Pull scenarios, apply stakeholder overrides, and deploy territories without leaving your own agent. Your agent orchestrates. Carve supplies the territory expertise and the infrastructure.
That's the best of both: the flexibility of building at the surface, with none of the maintenance burden underneath it.
Let your agent drive, let Carve carve
AI agents are going to change how RevOps works, and territory planning is one of the best places to start. But a plan is only as good as what happens after the first draft: the feedback, the deployment, and the months after go-live.
You can spend 2027 planning season maintaining a homegrown territory agent. Or you can spend it on strategy, with your agent on the surface and Carve handling the infrastructure.
Try Carve free. There's no credit card required, and it works from the Carve UI or straight from your own agent over MCP.
