Implementation projects: automations and agents

The problem

Repetitive, rule-based work gets done in a spreadsheet and someone’s memory, week after week. Mistakes don’t show up right away. They usually surface months later, when someone compares numbers and can’t explain the gap.

The same applies to larger work processes, such as order handling, quote preparation or management reporting. The work runs fine while the person who knows it is around, but on holiday or under pressure details get missed, and the rules live in that person’s head and nowhere in writing.

What we do

We build a significant work process from design to go-live. Three examples of typical projects, as fictional illustrations, not descriptions of completed client work:

• Order handling. Situation: orders arrive as emails and attachments, and the details get typed by hand into several systems. The solution extracts the data and checks it against agreed sources. The employee gets a fully drafted order ready to check and approve. Orders that don’t fit the pattern always go to a person.
• Quote preparation. Situation: prices, past quotes and technical details live in different places, and quotes sit waiting. The solution gathers the source information and produces a quote draft. The employee gets a draft whose sources can be checked. Pricing and the final quote are always approved in-house.
• Management reporting. Situation: management only sees the picture once someone has pulled the reports together by hand from several locations. The solution combines the agreed data into one shared report and flags exceptions. Management gets a finished view where every figure traces back to its source. What needs attention always stays a human decision.

The build follows the same pattern regardless of the process. First we scope the content and the decision point in design and scoping, the paid first phase of the implementation. Then we build the solution, test its edge cases and check the result before go-live. Go-live includes training the team on the new way of working and agreeing who owns monitoring and maintenance afterward.

Clear calculation and decision rules are implemented as rules. AI is used where it adds value to the task. Our own production systems already run on the same principle: our campaign-copy generator refuses to write anything that breaks a character limit, new-product data is pulled straight from the source system’s own API instead of copied by hand, and our product-data quality check catches translation and feed errors before a customer ever sees them.

The build is, for example, a Python pipeline connected to the Claude API or the Agent SDK, scheduled inside your own environment. MCP connections to your source systems are handled as a separate part of the work. Model choice is matched to task difficulty, repeated parts are cached, and large batches run overnight. Cost efficiency is part of the design from the start, not a line to explain on the invoice afterward.

What you own

The code in your own repository, the scheduling in your own environment, and the AI account in your own name. Billing goes straight to you, and we take no margin on it. Maintenance is a separate, optional agreement; without it, the system keeps working, because it’s yours. The same goes for the rules: they’re written into visible code, not left to live on in one person’s memory.

What we don’t promise

• That building starts without a scoped design phase on significant projects. The content and the decision point are agreed first, before anything gets built.
• Automation without a human checkpoint before production. Any writing action is walked through, never assumed safe.
• To “guarantee” ad profitability. Profitability is assessed from your own sales and ad data, not general averages, and decisions are made on the measurement.
• A customer-service bot that answers without human oversight. First response can be automated, but responsibility for the answers stays with a person.
• Full write access without explicit scoping. Permissions are reviewed tool by tool before anything goes to production.
• One solution for all repetitive work. Some tasks fit training or a ready-made MCP connection better than a custom agent. We say which before we start building.

First step

Tell us about one significant work process and its current state. We’ll go through it together before we design a solution.

Get in touchNext step: MCP connections to your business systems