TL;DR — An AI can generate a CRM demo in an afternoon; the permissions, email sync, imports and migrations behind a real one are where the weeks and tokens go. Start from an open-source system like Twenty, then spend AI effort on agents, MCP and the integrations that are actually yours.
Every few weeks a client asks whether an AI coding agent could simply write them a CRM — or a small ERP — tailored to how they work. It can write one. The question is whether it can write one you would still want to maintain in a year, and whether that is the cheapest way to get there. Usually it is not.
The from-scratch trap
Generated code makes the first screens look finished. Contacts, companies, a deal board — a day of prompting. The trouble starts with the parts nobody demos:
- Permissions. Who sees which deals, which fields, which exports — and an audit trail when someone asks who changed what.
- Email and calendar sync. Threading, duplicates, shared inboxes, token refresh, the meeting that moved twice.
- Imports and migrations. The spreadsheet with 40,000 rows and three date formats, then every schema change after go-live.
- The fix loop. Each bug fix is another round of prompts, tokens and review — and the code is yours alone to understand.
We have written before about what tokens really cost. A from-scratch CRM spends them on problems that were solved years ago.
What open source already solves
Take Twenty, an open-source CRM. Out of the box it covers what a company usually needs first:
- Pipelines and records. Companies, people, opportunities and tasks, with board and table views.
- Workflows. Automations triggered by record changes or schedules, without writing a job runner.
- Email and calendar. Meetings and threads attached to the people and deals they belong to.
- A data model you can change. Custom objects and fields — defined in the UI, or as versioned code with its SDK.
The same logic holds for ERP: established open-source suites such as ERPNext and Odoo Community have already absorbed a decade of accounting and inventory edge cases.
Cost and control
- No per-seat licence when self-hosted. Twenty runs on your own server with Docker Compose; the cost is infrastructure, not headcount.
- Your data stays in your infrastructure. That simplifies the conversation about KVKK and GDPR.
- Read the licence. Twenty’s core is AGPLv3, with a few enterprise files under a commercial licence — fine for internal use, worth a check before you redistribute.
- A community behind it. Security fixes and upgrades come from a large project, not from one person re-prompting.
Where AI earns its place
The interesting work moves from rebuilding a CRM to making one smarter. Twenty Cloud workspaces ship with a native MCP server, so assistants like Claude, ChatGPT or Cursor can read and write CRM data over OAuth. On top of that:
- Agents with a narrow job. Enrich a new lead, summarise a meeting into the deal, draft the follow-up for a human to send.
- Integrations that are actually yours. ERP, billing or support systems, connected through the API.
- Small, reviewable code. A custom object or workflow step is exactly the scoped task coding agents do well.
When building from scratch is right
Sometimes it is. If the system is the product, if the domain model fights every CRM’s idea of a contact and a deal, or if licence terms conflict with how you distribute software, custom wins. We build those too — with the same discipline as any internal tool. It just should not be the default.
What this buys
A proven core on day one, no seat fees, data you control — and AI spent on the edges where your business is genuinely different. Do not reinvent the wheel; make it smarter.
Frequently asked questions
Is it worth building a CRM from scratch with AI?
Rarely. AI gets a demo working quickly, but permissions, email sync, imports, audit history and migrations take most of the effort, and you own every line afterwards. Adapting an open-source CRM and spending AI effort on integrations and automation usually ships sooner and costs less to maintain.
Can AI assistants work with an open-source CRM like Twenty?
Yes. Twenty Cloud workspaces ship with a native MCP server, so assistants such as Claude, ChatGPT or Cursor can read and write CRM data over OAuth, and every installation exposes an API that AI agents and automations can call.
Working on something similar?
T-Square architects, builds and operates production systems for learning, AI and custom software products. Talk to a senior engineer for a second opinion.