Enterprise AI Myth #1: Just Vibe-Code a New App with Codex or Claude

Do you want new software, or smoother business?

More than one owner has asked me this year: our ERP and CRM are clunky — can you use AI to build us a replacement? The reasoning is understandable: the old system hurts, and Codex, Claude Code and Cursor can all write code now, so why not rebuild? Sounds smooth.

I usually ask one question first: do you want a new piece of software, or do you want your business to run with less friction? Most owners pause. The two answers look similar and are worlds apart. Building software is just an upgrade of the old software, with custom features. Removing friction is about where the work happens every day, who handles it, where the material comes from, who takes over when it jams, and where the result lands — none of which necessarily requires new development.

Most problems are not short of software. They are short of continuous action at the place of work: files live in group chats, confirmations happen in meetings, the spreadsheet is in another folder, the approver changed this week, the boss reshuffled priorities in one sentence, and someone forgot to update the status. A rebuilt system does not automatically absorb any of that — often it just adds one more node the company has to maintain.

Case 1 · Invoice approval at an engineering firm: one invoice drags a whole chain

The traditional build would be an invoice management system: upload, extract fields, match the PO, run approval, land on a dashboard. The demo would look great. But the real problem starts the moment an invoice hits the inbox: the PO number matches no project — now what? One PO maps to several invoices — which are normal, which duplicated? Supplier names are written inconsistently — who confirms they are the same company? Amount and GST do not reconcile — return it, or send to finance for review? The project manager changed — who approves now?

None of that lives on a page or in a recognition model. It is scattered across inboxes, project files, PO sheets, group-chat confirmations, finance reviews and approval records. Build a standalone invoice app and staff still carry material in, carry exceptions out, and ask around in the chat — the system decays into another spreadsheet. The effective move is to bring AI into the existing environment: as an invoice arrives it extracts fields, finds the project, links the PO, routes exceptions to the right person, and writes results back to the documents, sheets and approval records. Invoice approval is about how an invoice keeps moving forward after it enters the company.

Case 2 · Admissions at an education business: a CRM records state; deals do not advance themselves

An admissions CRM sounds complete: lead pool, client status, funnel, dashboards. But admissions work does not happen only inside a CRM — conversations live in chat apps, documents in email, statuses in Excel, key clients get discussed in meetings, and plenty of judgment lives in counsellors’ heads. When a client stalls, the CRM says "following up" — and the owner reading those words still cannot tell whether the deal is in danger.

Rebuild the CRM and the usual outcome is: counsellors fill in one more system, the boss watches one more dashboard, and real opportunities keep leaking. What helps is AI working where counsellors already work: reading statuses in Excel, checking chats for follow-ups, spotting missing documents in email, prompting the next action, surfacing at-risk clients. What admissions loses most often is the next step never being pushed in front of a person in time.

Case 3 · Store operations: a dashboard displays exceptions; it cannot get them handled

Store projects turn into pretty systems easily: home page, charts, rankings, daily sales, stock alerts. Demos look great. But what an owner wants daily is not which number turned red — it is: who saw this exception? Who handles it? Was it handled? With what result? Will it happen again? A dashboard can only tell you a problem appeared; it cannot perform the actions that follow.

Store operations close their loop better inside the daily work environment: the daily report generates, AI flags exceptions, exceptions route to owners, results write back into the report, and recurring issues get compiled for management. The worst outcome in store management: a beautiful dashboard, and nobody catching the actions.

The three cases, one structure: three layers

Invoices, admissions, stores — three industries that decompose into the same three layers on the ground.

• The tool layersolves point capabilities — reading an invoice, generating a report, writing code. Coding agents, LLMs and OCR live here. It makes one action faster; it does not catch the whole chain.
• The workflow layeris where companies actually jam: after extraction, where is the PO checked? Who gets the exception? Where does the result write back? This happens between email, chats, sheets, approvals and meetings — the workflow layer makes AI act continuously across those nodes.
• The operating loopdecides whether AI stays: continuously spotting exceptions, driving owners to handle them, recording the process, feeding review — so the next run gets smoother. Handling one invoice once is worth little; entering operations is what counts.
• Why projects drift:they mistake tool-layer capability for the whole deployment. Codex writes code, so rebuild the software; the model summarises, so build a knowledge base — all workable directions that stall at Demo without the workflow and the loop behind them.

A fast demo is not production

AI coding genuinely made development faster: what needed a project schedule can now be a passable version built by one person in days. But faster development does not automatically improve how a company works. After launch come the questions: will staff actually use it? Does old material migrate? What about confirmations in group chats, decisions in meetings, the approvals, tasks and calendars still living in the old systems? Who maintains it in two years?

Many companies end up in an awkward spot: the old system still runs, the new software also runs, and staff shuttle data between them — a faster way to add one more place to work. Personal software has multiplied a hundredfold in the vibe-coding era, and business software severalfold; the number that stabilised under sustained real use is vanishingly small. Enterprise software needs stability, extensibility and maintainability — not a DIY tool used by exactly one company.

What companies need is AI inside the work environment

Daily work already happens inside collaboration suites: people, org structure, permissions, documents, sheets, tasks, approvals, calendars, meeting minutes and message history all live there. That is not a chat tool — it is the shop floor of daily operations, and that is where AI should show up: not waiting inside a new system for staff to open it, but present where people already work.

Someone drops an invoice into the group chat — it catches it. A counsellor updates a sheet — it understands. A to-do appears in the minutes — it tracks it. An anomaly shows in the daily report — it alerts the owner. When results land, it writes them back. That is why we put the agent framework on top of the collaboration suite: the suite provides the environment; the framework deploys agents that understand context, call tools and push tasks. Traditional software makes people enter the system. Agents should enter people’s workflow.

A simple test

You will see plenty of rapid-build, AI-coding, agent-application pitches in the coming years. If a proposal talks pages, features, modules, databases and launch — but cannot explain how it enters your daily work environment and connects people, documents, sheets, tasks, approvals, meetings and existing systems — it is most likely just a faster outsourced dev project. It can produce a demo. It will not necessarily change how work gets done.

• Question 1:does it run where the company already works every day?
• Question 2:can it understand real context?
• Question 3:can it call the tools you already have?
• Question 4:can it push tasks forward?
• Question 5:can it write results back into the original environment? If these cannot be answered, it is probably just new software. If they run, it starts to be an agent inside the business.