# When merchants build for themselves

> When merchants build for themselves: Chapter 16 shows how a brief becomes a prototype, what changes for agencies and how to cost your own software projects.

- Source: https://www.agentic-commerce.sh/book/en/16-when-merchants-build-for-themselves/
- Author: Stefan Hamann
- Updated: 2026-11-16
- Language: en

![The expertise was already there. Now it gets the right compartments.](https://www.agentic-commerce.sh/book/media/4d4a8bfc11cfb662.webp)

*The expertise was already there. Now it gets the right compartments.*

## A real case: the merchant shapes its own ERP

A merchant I know is using a coding agent to develop an application intended to replace its existing ERP. It starts with the workflows the merchant knows from daily work. Rather than first fitting them into someone else’s standard design, the merchant can turn them into software and use the prototype to develop the idea further. [[S144](https://www.agentic-commerce.sh/book/en/f-sources-and-further-reading/#en-S144)]

The merchant’s role changes. Working with a coding agent, it can turn a workflow into software and try it before commissioning a complete external design. The application reveals whether a rule is clear and works in daily use.

Because an ERP connects purchasing, products, stock movements and accounting, the merchant’s niche knowledge is particularly valuable. It knows which exceptions recur and which supposed special cases are normal business. That knowledge can now become the software’s direct starting point.

Replacing the system also involves the business’s history. Germany’s GoBD rules require an equivalent transfer of tax-relevant data, including necessary links; procedural documentation also belongs in the project. Migration and tax coordination therefore remain distinct tasks. [[S148](https://www.agentic-commerce.sh/book/en/f-sources-and-further-reading/#en-S148), [S149](https://www.agentic-commerce.sh/book/en/f-sources-and-further-reading/#en-S149)]

The case points to a different division of work. Merchants shape solutions and identify where they need help. Developers, agencies and specialists bring the experience needed for dependable operation. Businesses gain independence as the distance between their knowledge and a usable application shrinks.

## A real case: a blocked migration finds a way forward

Another merchant needs functions tailored to its business model. Using Shopware and custom extensions, it is developing a solution previously beyond its economic reach. A coding agent helps the merchant turn its requirements into software and assess them through a working design. [[S144](https://www.agentic-commerce.sh/book/en/f-sources-and-further-reading/#en-S144)]

The business idea already existed; the merchant needed software to match. It can now describe a requirement, try a design and develop it around its workflows. It can shape the migration directly.

The extensions use the platform’s core functions. Product management and ordering do not need to be reinvented for each requirement. The merchant concentrates on the rules that make its business distinctive. Open code and a coding agent provide direct access to the system.

A working application changes the agency relationship because merchants can use it to explain their requirements. Advice, architecture, testing and maintenance start from a concrete solution. This brings business knowledge and technical experience together earlier in the project.

Both anonymized cases come from merchants I know and show what merchants can undertake with coding agents: an operational core system or the requirements of a complex shop. What fascinates me is business experience becoming direct working material for software. More companies can shape their own ways of working rather than adapt them to what they could previously afford.

## The brief becomes a prototype

At Werkpunkt, the idea starts with a spreadsheet. A maintenance buyer pastes a list of spare parts from an internal system; sales staff assemble article numbers, variants and customer prices. The desired quick-order flow is clear: upload the list, review recognised items, resolve ambiguity and prepare an offer.

Let a coding agent build the first version on a copy of the shop, using existing product and pricing functions. The initial screen already reveals the real work: an abbreviated description matches two revisions of the same machine. The customer must answer that question before the list becomes an offer. An attractive upload button would merely have moved the ambiguity along faster.

With the brief already expressed as a prototype, the agency can examine the business journey and agree with the merchant how to turn it into a maintained extension. For me, this is a major opportunity in agent-assisted development: ideas become tangible before a large project budget is committed.

Customer pricing makes this freedom tangible. An industry rule may run through several extensions. Access to the code lets the team trace it through to the order. Chapter 15 explains the technical foundation; what matters here is that the merchant’s own rule becomes visible and open to discussion.

## What changes for agencies

Coding agents allow agencies to deliver more of the work merchants previously had to put off because of cost. Routine development takes less time, creating room for more tailored solutions, better integrations and closer operational support. An agency’s value increasingly depends on how well it understands its client’s business and turns these possibilities into dependable solutions.

An early prototype makes an idea easier to discuss. The duration of a whole project also depends on data, decisions, integration and rollout. Merchants and agencies can agree which work an agent should handle and where the partner’s experience makes the greatest difference.

| Task | With coding agents | Agency contribution |
| --- | --- | --- |
| Clarify the idea | Merchant and agent build an initial version | Sharpen business logic and UX |
| Implementation | Routine development takes less time | Shape extension points, architecture and integration |
| Quality | More variants become possible | Check security, performance and real edge cases |
| Operation | More custom extensions remain in circulation | Maintain updates, monitoring and recovery |
| Commercial model | Effort and value change | Price scope, responsibility and benefits transparently |

An operations package names the extensions covered, response times and included updates. For outcome-based fees, merchant and agency agree on the benefit measured and how external influences are treated. A new pricing idea becomes an offer both sides can work with.

## Nordstern (fictional): advice that used to cost too much

Nordstern builds an outdoor trip planner for customers who plan outside store hours, extending advice that was previously economical for only a few orders. Destination, season, existing equipment and preferences lead to a suitable selection.

For one product group, the prototype uses verified attributes, stock and storefront prices. Missing facts prompt questions; experienced advisers try familiar trips and correct mistaken conclusions. Before launch, advice, basket and service must retrieve the same selection. Nordstern can then add more practical knowledge, creating distinctive advice whose depth would previously have been too expensive to offer.

Werkpunkt turns the earlier design into an extension where customers paste spare-parts lists. The agent matches articles while the shop supplies customer prices and delivery information. Unclear machine revisions go to a specialist before an offer is prepared.

A coding agent helps build the screen and connections; merchant and agency check familiar lists and common errors. Read-only customers can review results without ordering. This niche capability can later become a managed procurement service, built around known machines, approved parts and the customer’s purchasing rules.

Lichtfeld uses a coding agent to develop a collection experience from its images, product information and design rules. The team directs the atmosphere and tries alternatives in the working design.

Every displayed product stays linked to an actual article. Colour and finish choices retrieve valid variants, prices and availability; the basket uses that selection. Starting with one collection, Lichtfeld learns which presentation helps customers choose and how to maintain it before extending its style to further rooms and ranges.

## Three shortcuts that become expensive

One is changing the core despite an available extension point. The next update brings together two changes designed without knowledge of each other. Another is uncontrolled data access: a preview rarely needs the entire customer database. The third is a private fork without a maintenance plan. The first version arrives quickly; nobody feels responsible for later updates.

Use a simple workflow: state the task and limits, develop on a copy, review the change, accept it against a normal case and a counterexample, then release a limited scope. Record the rollback path and operational responsibility. This can start small.

## A calculation template for your own projects

## Model calculation · two fictional software projects

For a fictional ERP project, assume 80 hours of domain work at €70, 200 hours of development at €90, 100 hours of migration at €90 and 40 hours of acceptance work at €100. Add €500 for tools. The total before introduction is **€37,100**. If operations subsequently save €2,000 a month and additional maintenance costs €500, the monthly benefit is €1,500. Simple payback is approximately **25 months**.

A second fictional project concerns shop extensions: 160 hours of implementation at €80, 40 hours of review at €100 and €600 for tools total **€17,400**. Maintenance at €300 a month adds €10,800 over three years. The merchant compares that €28,200 with the benefit of the new features and the cost of the existing solution.

These assumptions describe neither of the two real cases. The calculation shows why a model subscription alone says little about the cost of a software project. Agents make more ideas available to try; operations determine which ones last.

## What platforms need to provide

Platforms now serve two kinds of developer: people and their agents. Stable extension points, clear errors and traceable business operations help both. AGENTS.md files or skills explain conventions and necessary checks. Test data and private previews shorten the path to a useful result.

The Lab’s Agent in a Box provides this project workspace; Chapter 13 explains its security boundary. [[S109](https://www.agentic-commerce.sh/book/en/f-sources-and-further-reading/#en-S109)]

Which experiments deserve the next investment? The next chapter shows how we approach that question in the Lab.
