Why Your AI Sales Agent Isn’t Working — The Model Isn’t the Problem, Your Data Is

What is a GTM agent and why companies want one

A GTM agent (Go-to-Market agent) is an autonomous AI system that executes sales process steps without human involvement: qualifies leads, personalizes outreach, summarizes account status, suggests next best actions, and updates CRM records.

This is not a chatbot or a copilot. A GTM agent runs in a loop: observes data → makes a decision → executes an action → logs the result. No “send” click from a rep.

Why do companies want this? The numbers speak for themselves. According to the Anthropic State of AI Agents 2026, over 57% of organizations use agents for multi-stage workflows, and 80% report measurable ROI. The Halkwinds AI Agent Adoption Report 2026 shows that 45% of enterprise teams have at least one agent in production — in 2024 that number was under 3%. McKinsey State of AI 2026 reports that 44% of organizations are scaling AI across the enterprise.

The pressure is real. So is the problem.

Why coding agents work and GTM agents don’t

A coding agent operates on a single, coherent context: a code repository. Files are in one place, they have a deterministic structure, change history is complete (git), dependencies are explicit. The agent has a full, unified picture of reality and can act correctly.

A GTM agent operates on chaos. Its context is a CRM — and a CRM in most companies is a graveyard of half-truths: incomplete notes, duplicate records, missing dates, manually entered data full of typos.

That’s a fundamental difference, and it’s exactly why coding agents deliver real value from day one while GTM agents stall — or worse, run confidently but incorrectly.

As the 24seven.news article put it well: GTM agents don’t work not because the model is too weak — but because sales data is fragmented, duplicated, and cut off from external signals.

An AI agent is only as good as the data it operates on. Give it bad context — you get confident but wrong answers. That’s not a model problem. That’s a data problem.

Three concrete reasons why CRM data is bad

1. “Happy ears” — reps log what they want to see

“Client interested,” “demo scheduled,” “opportunity for Q4.” Reps naturally log positive signals. Objections, delays, churn signals — they often never make it into the system, or get filed as “follow-up needed.”

This isn’t a people problem. It’s human nature. The issue is that an agent operating on this data sees a more optimistic pipeline than reality. It qualifies a lead as hot because the interaction history looks positive — without knowing the prospect cancelled the call three times and hasn’t replied to the last four emails.

2. Identity resolution — the same customer in three places

Cisco as a prospect might appear in your CRM as “Cisco,” “Cisco WebEx,” and “AppDynamics” (a Cisco company). Three separate records, each holding a fragment of the history.

An agent analyzing one account doesn’t see the full picture. It doesn’t know your company already talked to a different unit of that same customer two years ago, that there was a pilot, that there was a champion who left. Without identity resolution — the process of merging duplicate records into a single unified company profile — the agent draws wrong conclusions from incomplete data.

3. External signals never make it into the CRM

Series B funding round. New VP of Sales hired last week. Job posting for “Head of Revenue Operations.” Board changes. Acquisition by a competitor.

These are buying intent signals — and they’re not in a standard CRM. The agent doesn’t know that an account that went dark for six months just raised $20 million and is now building out a sales team. It doesn’t know because nobody typed it into Salesforce.

What a working agent looks like — minimum requirements

For a GTM agent to work correctly, it needs four things. Not optional — mandatory:

Identity resolution — one record per company, regardless of how many entities, domains, and names that organization has. The agent must operate on a unified profile, not a fragment.

External signals linked to accounts — funding data, personnel changes, job postings, industry news must flow into the agent’s context automatically, in real time. Not once a quarter via manual LinkedIn import.

Full interaction history — emails, call notes, demo recordings, results from previous campaigns. Everything. Not just what the rep decided was worth logging.

Live data access, not historical snapshots — the agent needs to see the account state today, not the state from the last sync 48 hours ago. Continuously synced data is the minimum.

Without these four elements, the agent will work — but incorrectly. And that’s worse than not working at all: confident wrongness costs more than no answer.

Worth noting: even well-equipped agents require oversight. Halkwinds shows that 67% of production deployments include a mandatory human review gate. That’s not a weakness — it’s intentional design.

What to actually do about it

There are two paths, both viable, but with very different costs.

Option A: integrate your existing CRM

If you have Salesforce, HubSpot, or Pipedrive with years of history — you can try to layer agents on top. But be honest about the costs: this is a 6–12 month project, it requires a dedicated team for data cleanup, identity resolution is a separate project, and integrations with external signal sources (Crunchbase, LinkedIn, Apollo) need to be built from scratch.

For a 10–100 person company, that’s often not realistic. McKinsey reports that 32% of companies chose NOT to buy software because they could build it internally — but building a data layer for agents on top of a legacy CRM is far heavier than building new tooling.

Option B: start with a CRM designed for agents

If you’re starting fresh or can afford a migration — it makes sense to set up a CRM that assumes an agentic model of work from the ground up: unified data, built-in identity resolution, automatic ingestion of external signals, full history in the agent’s context.

Khirby is an open source, self-hosted CRM designed exactly for this scenario. Not a SaaS with an agent bolted on the side — an architecture where the agent is a first-class citizen and data is organized so it can actually use it correctly. You can self-host it and keep full control of your data.

Summary

A GTM agent isn’t futurology — it’s today. But it only works when data is complete, unified, and current. Most CRMs don’t meet those conditions — not because they’re bad, but because they weren’t designed with an agent as the user.

If your sales agent deployment isn’t delivering results, before you swap the model — check the data. The problem is almost certainly there.

Book a demo | Self-host it

Sources

1. Anthropic: The 2026 State of AI Agents Report

2. Halkwinds: AI Agent Adoption Report 2026

3. McKinsey: The State of AI in 2026 — On the Road to ROI

4. 24seven.news: Why Coding Agents Work and Go-to-Market Agents Don’t (Yet)