AI workflow automation for EU back-office teams

Capabilities

Four workflows we automate end to end.

The same four shapes cover most of the manual work that eats a week.

Lead routing and follow-up

Inbound requests are classified, routed to the right owner and chased automatically. An LLM step reads the message and picks the queue, then rules confirm it before anything is sent. The follow-up stops the moment a reply lands.

Document intake and extraction

Contracts, invoices, forms and CVs arrive by e-mail or from a shared folder. They are parsed into structured records, with an extraction step reserved for the fields a parser is not certain about, so a human checks the doubtful ones only.

Reporting and internal alerts

Scheduled reports are assembled from the systems you already run and delivered where the team already reads them. A number crossing an agreed threshold raises an internal alert instead of waiting for somebody to notice the number moved.

Approvals with human review

Every request passes a deterministic rule before it reaches an approver, and the approver decides. We do not automate the sign-off: a judgement call stays with a named person, and the audit trail records what was proposed and what was approved.

Scope

What we automate, and what we don't.

Automation earns its keep on rules. Judgement stays with people.

What we automate

Repetitive, rule-shaped work: routing, copying, extracting, reconciling, scheduling and chasing. If a step can be written as a rule or a checklist, it belongs in the workflow, and it stops being somebody's Friday afternoon.

What we don't automate

Decisions that carry accountability, anything that needs a negotiation, and work whose rules nobody can state out loud. We do not replace a role with a bot; we remove the steps around the role so the role can be done properly.

Where judgement stays

Where a step needs judgement we add an LLM step with guardrails and a human-review path instead of blind automation. The model proposes, a rule or a person confirms, and the outcome is logged — so the workflow waits for a decision instead of guessing at one.

Other services from the same studio: custom AI software development, AI integration services, and EU AI Act compliance for Regulation (EU) 2024/1689 preparation. Every engagement starts with a fixed first milestone, scoped and priced before kickoff, and runs against the engagement model published on the process page.

Delivery

How it runs.

Four weeks from kickoff to a first milestone in production, then operation.

A fixed first milestone

One written scope, one deployed workflow, one success metric, and an explicit out-of-scope list. It is signed off before any code ships, and the figure written into the scope does not move once signed.

Monitoring from day one

Uptime monitoring and error alerts are wired on the first day, not bolted on at the end. You get the operational dashboard, not only the repository, so the workflow is observed while it is being built.

Backups from day one

Daily encrypted backups with a monthly restore drill, running from the first week of the engagement. A backup that has never been restored is not a backup, so the restore is rehearsed on a schedule.

Data stays in the EU by default, every subprocessor is named on the subprocessors page, and a GDPR Article 28 Data Processing Agreement is available on request. The stack is deliberately boring, with no vendor lock-in, so the workflow stays maintainable by somebody other than us. What we have shipped is on the work page, our defaults are on the security page, and the artifacts a procurement or DPO team asks for are listed on the trust index.

The shape of the work is not a house style: the human-in-the-loop pattern is what the NIST AI Risk Management Framework describes for govern-and-manage functions, and the guardrails around an LLM step are the OWASP Top 10 for LLM applications risk list read as a checklist.

Trust signals

Proof close to the decision.

Three signals buyers can verify before they start a conversation.

EU data residency by default

Product runtimes run on our own infrastructure, EU-first by default. See our security posture.

Founder-led delivery

The person you contact stays close to the work, from written scope through operation. See how we work.

Open-source track record

OpenStore is MIT-licensed and public, with 320 GitHub stars and 120 forks. Inspect the foundation.

Questions

What a process owner asks before starting.

Which processes are worth automating first?

Start with the work that is frequent, rule-shaped and repeated every week — lead routing, document intake, recurring reporting and approval chasing. If a task runs many times a week, follows a documented rule and produces a checkable output, it is a good first candidate. We name one first milestone, not a full back-office replacement.

How do you handle a step that needs human judgement?

We do not automate the judgement itself. Where a step needs judgement we add an LLM step with guardrails and a human-review path instead of blind automation: the model proposes, a rule or a named person confirms, and the outcome is written to an audit trail. The workflow waits for the human decision instead of guessing.

Where does the data live?

In the EU by default. Product runtimes run on our own infrastructure with EU-first data residency, TLS in transit, and backups encrypted at rest. Every subprocessor is named on the subprocessors page, and a GDPR Article 28 Data Processing Agreement is available for customers who need one.

How long until the first milestone ships?

Four weeks from kickoff. The first milestone is a fixed scope with one deployed workflow, written down and signed off before any code ships, plus an explicit out-of-scope list. At the four-week mark you get one production-ready slice end to end, and you decide whether to extend, refine or stop.

What does it cost?

Every project starts with a fixed first milestone, scoped and priced before kickoff. The figure for your project is written into the scope during validation and it does not move once signed. We tell you whether the work is a fit in one reply, and we say plainly when it is not.

Will this replace the tools we already use?

Usually not. We automate the process across the systems you already run rather than asking you to migrate onto new ones. Where a step needs judgement we add an LLM step with guardrails and a human-review path, and where a rule can do the job we use a rule. The result is a boring, well-understood stack you can maintain.

One paragraph is enough to start.

We take on a limited number of contracts each year. If your paragraph names these three things, one reply can tell you whether it’s a fit.

The problem What is done by hand today, and by whom.
The metric The one number that will say it worked.
The deadline The date it has to work by.

Write the paragraph Book a call

Reply within one business day, from the person who writes the code.

AI workflow automation for back-office teams that live in e-mail and spreadsheets.