Your team shouldn’t spend
Friday compiling reports.

Not always, and we say so on the first call. Narrow jobs that repeat thousands of times, such as pulling fields out of documents, classifying tickets or reranking search results, often run better on a small model built for that one job: it runs on a modest server of your own, answers a prepared batch in under a second in our tests, and has no per-token fees. When the layout never changes, plain rules or a script are cheaper still. Large language models stay where the task needs reasoning, open questions or conversation. Quickscratch picks the smallest tool that does the job reliably, because that is also the one that is cheapest to run and easiest to keep private.

Answers are built only from your own sources, and each one cites the document it came from. Before the model writes anything, the retrieved passages are checked for relevance; when they don’t answer the question, the assistant rewrites the search from the conversation, and if the sources still don’t hold the answer, it says it doesn’t know rather than inventing one. Keyword search runs alongside meaning-based search, so exact codes and names aren’t blurred. For answers where every word matters, a second model checks the first before the user sees it. And every behaviour you care about becomes a test that runs before each release.

You choose where the system is hosted: an EU cloud region, your private cloud or your own servers. Files, the search index and logs stay where you put them. When an assistant calls a cloud model, a small local model first finds and masks personal data such as names, ID numbers and account numbers, so the cloud model sees the task, not the person; the details are restored in the answer on your side. Sensitive workloads can run entirely on local models, in which case nothing leaves your infrastructure at all. Your data trains only your own model, if you order one.

Your data trains only your model, if you order one. Never ours, never a model vendor’s.

In most cases, yes. From 2 August 2026, Article 50 of the EU AI Act requires that people are told they are interacting with an AI system, unless that is obvious from the context. Every assistant we build discloses it from the first message.

Regulation (EU) 2024/1689, Art. 50

Switching is a configuration change. We test the new model against your data first and only switch if it’s better.

The core is ours and stays ours. What you ordered — the business logic — is yours: the way your data is collected, processed and turned into a result. Generic building blocks are not business logic. A class that writes files to Google Drive stays ours and we reuse it elsewhere; a class that processes your data in your particular way belongs to you. Models trained specifically for you are yours as well — that is custom work, and we are only co-authors. By default we build, host and maintain the product, which is how security fixes and core improvements reach you quickly. If you would rather run everything yourself, you can buy a licence to use the core in your own projects, with credit to QuickScratch. That licence also lets you end our support and continue on your own — we just need it confirmed in writing.

You don’t have to take it on faith. Payment is split into three milestones, the last one after acceptance tests agreed before the build starts, so you never pay in full for something that doesn’t work.

MCP (Model Context Protocol) is an open standard that lets AI tools such as Claude, ChatGPT or Cursor read and update your systems directly. Quickscratch is building a ready MCP core, available as early access, not yet as a finished service. The first stage covers reading and updating records in your database, spreadsheets or CRM. The next step is shaping the data for the model: querying a large table, filtering it and computing statistics, so the AI tool gets a compact, relevant context to analyse instead of raw rows. The point is that your team can ask questions of its own data from the AI tool it already uses.

Yes, and it changes nothing about how we work. You bring the workflow, we take on the technical side and explain every decision in plain language, so you can judge it for yourself. Experience with AI is welcome, not required.

If a task repeats every week and takes someone hours, it’s usually worth automating. If it’s a one-off, we’ll say so.