Vibe Coding AP: Build or Buy?

Part 1 of 5 in Telema’s series on vibe coding and AP automation.

With AI-assisted coding tools—often called vibe coding—a motivated employee can build a prototype that reads invoices, extracts data and suggests accounting codes in surprisingly little time. You no longer need to start every automation idea with a traditional software development project. For finance teams with a long list of repetitive tasks, that’s a real opportunity.

At Telema, we use AI to automate our own work too, and we think people should experiment with it. But there is a big difference between a useful internal tool and a system the finance team depends on every day.

An invoice is not just a document. It contains supplier relationships, negotiated prices, payment instructions, personal data and figures that end up in your financial statements. If the software misreads an invoice, loses track of its status or sends incorrect information to the accounting system, somebody in finance has to find and correct the problem. And the consequences may only become visible weeks later.

Where a prototype stops and real AP work begins

A demo usually runs on a handful of clean invoices. Daily accounts payable looks different. Imagine the same prototype after three months, processing hundreds of invoices from suppliers with different formats, references and payment terms:

• A supplier changes its bank account.
• The same invoice arrives twice, once by email and once as an e-invoice.
• A credit note refers to an invoice from a closed accounting period.
• An ERP update breaks the integration overnight.

These are not rare edge cases. They are part of normal AP operations. A prototype may be very good at reading an invoice but have no reliable way to decide whether it should be approved, held for review, matched against an order or sent to the ERP. These decisions need rules, permissions and a clear owner. An AI-generated piece of code is not a substitute for a controlled financial process.

Four questions before you build

Before connecting a custom AI agent to your invoice inbox or ERP, we would ask four questions. They matter whether you build in-house or choose an existing AP automation system:

1. Security and accountability: can untrusted invoice content influence the agent, what actions is it allowed to take and who approves financial decisions?
2. Data and confidentiality: where will supplier invoices be processed, who can access them and what may an AI provider do with their contents?
3. Audit trails and exceptions: can the team correct a wrong suggestion, reconstruct the decisions later and handle invoices that do not follow the happy path?
4. Cost and ownership: who maintains integrations, responds to incidents and pays for the solution after the initial prototype is finished?

Each of the next four posts in this series looks at one of these questions in more detail.

None of this is an argument against AI. Quite the opposite: small experiments are often the best way to discover useful automation opportunities. Let people prototype, test with sample documents and explore ideas. But before a tool starts processing live invoices, decide what must be checked by a person, what needs a recoverable audit trail and who is responsible when something fails.

For a company with a strong development team, building its own AP solution may be reasonable. For another, buying a maintained platform and adding custom AI for particular tasks could be much more efficient. The right decision depends on the business process, not on how impressive the first demo looks.

The question is not “Can we build it?” It is “Do we want to operate, secure and audit it for the next five years?”

Next up: invoice fraud and AI agents. Follow Telema on LinkedIn to catch Part 2.