DEV Community

gewumind
gewumind

Posted on

When Your CRM Is a Spreadsheet: Building a Customer Ledger on a Ready-Made Backend

Your company's customer data lives in three places. A spreadsheet, whose status column nobody has updated since the first month. A chat log — WhatsApp, WeChat, whatever the team already lives in — where the real conversations happen. And a notebook of promises made on site visits.

None of the three has an owner. When a sales rep leaves, their sheet stops moving; a customer marked "following up" stays that way for weeks. The data is not missing — it is scattered, and nothing in your stack is built to hold it together.

Then the boss asks the question: "Who hasn't been followed up this month?" Answering it takes an afternoon of scrolling sheets, cross-checking chat history, and pinging people over who owns which batch of customers. By the time the number is assembled it is already stale, and slicing it by status, rep, or region is out of the question.

If you are the developer handed "build us a customer management", or the tech lead comparing options, the decision that matters is where the shared layer underneath the customer table comes from.

Why it happens

The two obvious exits both fail.

Buying a SaaS CRM hands your customer records and your login to a vendor: data and accounts live on someone else's service, and migrating out means re-importing field by field.

Rebuilding from scratch trips over the same wall. The customer table — name, status, owner, follow-up log — takes days. What eats weeks is the layer underneath it, the part every back office ends up rewriting:

  • Permissions — who sees which customers, which buttons, which API endpoints — and how the backend enforces it.
  • Data dictionary — where statuses like lead / following / won / lost live as one shared set of values, not string literals copied across three codebases.
  • Menu and page skeleton — the sidebar, pages, and components that make a module feel like a product.
  • Stats and reports — stat cards, list and crosstab views, and an export that lands in the morning standup.
  • Open APIs — how a third-party app gets an identity, calls only authorized endpoints, and leaves a log behind.

Large companies have a team that builds that layer once and reuses it for years. Small teams do not: "build us a customer ledger" quietly becomes "build us a system", and the maintenance bill starts before the business value does.

The goal is not really to write less code. It is to make business data visible and usable: the boss asks, the system answers, and the answer is filtered by who is asking. Everything else in the list above is plumbing — and plumbing that already exists is plumbing you stop paying for.

The foundation

GewuMind Boot's one-line positioning: "AI's backend foundation, not an AI code generator." The kernel is open source under Apache-2.0, and after a baseline import it ships 45 system tables plus the platform's menus, roles, dictionaries, configuration, and page-component data. The baseline already contains the five capabilities above. Deployment is self-hosted: the release package runs without Node or Maven, and your data stays in your own database (MySQL 8 built in).

You install gewumind-skill into an agent you already use — Claude Code, Cursor, opencode — describe the requirement in natural language, and the agent runs a four-stage pipeline: requirement → design doc → code generation → one-click publish. The output is a complete business module — SQL, API endpoints, pages, plus menu and dictionary data — pushed to the running backend in one step; the new menu and pages are visible right after publish. The artifacts are plain files, fully equivalent to hand-written code — no black box. The official docs demonstrate the same mechanism with a one-sentence device-management module (device registry, heartbeat monitoring, firmware versioning): this is a production delivery path, not a demo.

A customer ledger is a one-line requirement on that pipeline:

Build a customer management module: customer list, follow-up log and opportunity
tracking; customer status managed by the data dictionary; role-based permissions;
stat cards and reports.
Enter fullscreen mode Exit fullscreen mode

Here is what the ledger looks like on the foundation, in the three places it gets used.

Scenario 1 — The ledger itself: customers, follow-ups, opportunities

Sales reps and managers use it: the AI delivers three business tables — customer, follow-up log, opportunity — with their endpoints and pages, plus menu entries and dictionary items, published in one step. Reps and managers open the same system and see different slices, by design.

Customer status — lead, following, won, lost — is a built-in data dictionary, not hardcoded values. Business tables store only the value; the display layer translates it to the label. Renaming "lost" to "not closed" is a dictionary edit: dropdowns and columns update everywhere, and no row of history changes. One source of truth keeps a ten-person team from drifting into three spellings of the same status.

Permissions come from built-in roles: a role is a set of permissions — create the role, grant the permissions, assign the users. Three layers, all enforced: menu (what shows in the sidebar), page buttons (front-end), and API endpoints (back-end). The last is the security baseline — front-end control hides the button, it does not close the door. Row-level scoping goes further: a role can see all data, a custom selection checked on the organization tree, own level and sub-levels, or own level only — and every query, from lists to stats to reports, is filtered by the user's organization scope. A regional rep sees their region; a branch manager sees theirs and the teams under it. Visibility stops depending on good behavior.

The workbench home page ships with stat cards and charts — by default platform data plus sample charts — and the same stat capability is extended by the AI with customer and follow-up cards.

Scenario 2 — Follow-up reports, when the boss wants numbers

"Who hasn't followed up this month?" should stop being answered by flipping through files. The foundation's built-in reporting — a report component (MagicReport) plus a data API that can connect straight to any business table — points directly at the customer and follow-up tables the ledger module created. Presets cover the shapes a back office needs: list, crosstab, Chinese-style multi-header, multi-sheet, and subtotal styles. One-click Excel export means the monthly follow-up report, or the "won this month" list, lands in the morning standup as a file, with no data warehouse in between.

Scenario 3 — Open APIs, feeding the mini-app and third-party systems

When a mini-app or a partner system needs customer data, the built-in open API gateway is the answer — and it is a capability of the free/open-source tier. Each third-party application gets its own appKey/appSecret; you authorize it per endpoint, with per-endpoint rate limits. The third party exchanges credentials for a token and then calls only the endpoints it was authorized for. Every call leaves an access record — application, endpoint, duration, result code, IP, timestamp — for reconciliation and rate-limit tuning.

The official line: external systems never touch your backend account system — they hold one application identity and call only the endpoints they were authorized for. Database credentials are never handed to anyone outside the company.

What to look for

Before you evaluate — or build on — any foundation like this, the honest limits matter as much as the feature list.

  • It is a foundation plus AI delivery, not an off-the-shelf CRM. The customer, follow-up, and opportunity modules are artifacts the AI generates on the foundation. Industry sales machinery — lead distribution rules, multi-stage conversion, commission schemes — is something you define yourself; there is no built-in sales-process suite.
  • Reporting targets small and mid-size back-office needs. It is not a data-warehouse or BI replacement for large-scale multi-dimensional analytics.
  • The workbench shows platform stats out of the box, not your customers. The default home page carries platform stat cards (users, roles, menus, operation logs) and sample charts. Customer and follow-up stats are extended on top of the same stat and reporting capability, delivered by the AI — do not expect them on first login.
  • Licensing is open core, and the split is specific. The kernel is open source under Apache-2.0; some enterprise capabilities (workflow engine, SSO, shared sessions, message channels, module management) are enabled at runtime by an offline commercial license key. Without a key they are disabled or run in a 30-day trial, after which the instance degrades to the base edition.
  • It is a B2B back-office foundation. If you are building a high-concurrency consumer-facing online service, this is not the layer to build on; the built-in database is MySQL 8.

CTA

If you'd rather have AI deliver modules on top of a mature backend than generate code from zero, take a look at GewuMind Boot.

Top comments (0)