Consulting
AI architecture and integration
Two moves, always in this order. First we put the data your operation runs on into a shape software can be trusted to act on. Then we build the system that acts: taking over a real process end to end, on your infrastructure, with a human in the loop exactly where you want one.
Data foundation
One shape and one home for the facts your business runs on: clean schemas, a single access layer, a full audit trail. Not a migration for its own sake: the minimum structure an agent needs to be trusted.
Agents that act
Not a chatbot over your documents. A system that reads the current state, makes the decision, writes the result back into your tools, and escalates the cases it should not decide alone.
Integration with what you already run
Your ERP, CRM, warehouse system and inbox stay where they are. We connect to them rather than replace them, so nothing depends on your team changing how they work.
Guardrails and audit trail
Limits enforced in code, not in a prompt. Every decision logged with the data it was based on, so you can answer "why did it do that" months later, and satisfy an auditor when asked.
Self-repair, not support tickets
Upstream systems change and formats drift. That is what normally kills an integration six months in. Ours notices, diagnoses the cause and repairs itself within limits you agree with us up front: a renamed field it fixes on its own, anything structural it proposes and waits for a human. A live dashboard shows what it handled, what it repaired and what it escalated.
Yours, with no retainer
Deployed on your servers and your own AI provider account, in your name, documented and handed over. No seats, no lock-in, and no monthly maintenance fee. You keep an agent you can ask to change the system, and you only hire us again if you actually want to.
Warranty, then support only if you want it
Ninety days of warranty from cutover: if it does not do what the specification says, we fix it at our cost. No ticket count, no hourly billing, no argument about whose fault it is. After that, support is paid and entirely optional. There is no standing fee and no contract that renews itself while you are not looking.
Migration without stopping the operation
Large scopes are staged. Old and new run side by side on the same inputs until the numbers agree, then we move one process at a time, with a rollback available at every step. Nobody in your operation has a bad Monday because of our project plan.
Why AI projects stall
It is almost never the model.
The pilot demoed well and then never shipped, because the moment it had to act on real data it could not be trusted. The same SKU under three IDs, orders that only exist in an inbox, a number that means something different in each system. We start there. Then the agent works inside your systems with limits enforced in code, complete decision logs and human sign-off where it matters, and if your data is sensitive, all of it can run on your own premises and never leave the building.
What we automate
The internal desk work: management and operations
Most of what we replace is not a person. It is the vacancy you have spent four months failing to fill, and the overtime your team is absorbing while you try. It is the part of the week that is reading, re-typing, checking and chasing: the work nobody was hired to do and everybody ends up doing.
Order and invoice handling
Orders arriving as email, PDF and portal exports get read, validated and entered. Invoices get matched against purchase orders and receipts, and only the genuine exceptions ever reach a person.
Reporting and month-end
The spreadsheet somebody rebuilds every Monday morning, and the pack that eats the last three days of every month: assembled on schedule, reconciled against source, and delivered without anyone having to ask for it.
Master data upkeep
Customers, suppliers, articles and price lists kept consistent across every system that holds a copy, with conflicts raised for a decision instead of silently overwritten by whichever import ran last.
Reconciliation and stock checks
Three systems each insist something different is true. The system compares them continuously, resolves what it can prove, and puts the rest in front of the person who is actually allowed to decide.
Request triage and routing
Incoming requests read, classified and routed to the right queue: the routine majority answered directly, the rest escalated to a human with the context already gathered rather than as a one-line ticket.
Planning and follow-up
Replenishment, delivery slots, shifts and routes proposed against live constraints: plus the chasing emails nobody enjoys sending, sent on time, every time, to everyone who is late.
Compliance and audit packs
The same evidence assembled from the same six places every quarter, with the trail of where each number came from attached to it, so the audit is a download rather than a fortnight.
The reports nobody reads until they matter
Margin drift, an unusual credit note, a supplier quietly slipping on lead time. Continuous checks that stay silent while everything is normal and speak up the moment a number leaves the range it should be in.
How the work actually runs
Five moves, in this order, on every engagement. The first four are the delivery. The fifth is what keeps the thing working once we are no longer in the building.
-
01
One database, not nine
Everything the business runs on gets pulled into a single structured MySQL database: the ERP, the warehouse system, the spreadsheets, the exports somebody emails round on Fridays. One schema, one set of definitions, one place to look. From that point a number means the same thing everywhere, dashboards take an afternoon instead of a quarter, and anything downstream, whether a person or a machine, can be trusted to read it.
-
02
Find what actually hurts, then rebuild that one piece
We sit with the people doing the work and find the part of the operation that genuinely costs you, which is rarely the part anyone had budgeted for. Then we rebuild that specific piece of software properly, instead of proposing to replace everything you own. Small surface, real fix, and something you can see working inside a month.
-
03
Reconnect it without stopping the operation
The new piece has to take over from the old one while the company keeps trading. Old and new run in parallel on the same inputs until the outputs agree, then we move one process at a time, with a rollback that takes an afternoon at every step. There is no weekend where everybody holds their breath and no Monday where nobody can invoice.
-
04
Put an AI on top of the database
With the data finally in one shape, we deploy an agent that reads it directly and builds on demand: a new report, a new check, a new automation, the chart somebody needs for a board meeting on Thursday. Your team asks for it in plain language instead of raising a ticket and waiting a quarter. It works inside permissions and limits enforced in code, and everything it does is logged.
-
05
A weekly review that comes to you
Every week the system reports on itself: what failed, what it repaired on its own, what has drifted, and what it proposes changing. During the warranty we act on it. After that it is your call: read it and decide, ask the agent to make the fix, or bring us back for the ones worth doing properly. What does not happen is the thing quietly rotting while everyone assumes it is fine.
The economics
Compare it to the job description you were about to post.
The honest comparison is not "AI versus no AI". It is this system against the person you would otherwise hire to do the same work, so here it is, side by side.
| An operations hire | The same work, automated | |
|---|---|---|
| Year one | €50,000–€70,000 fully loaded | One fixed project fee, typically €12,000–€25,000 |
| Year two | The same again, plus the raise | Nothing. You own it outright. |
| Hours covered | About 1,700 a year, in office hours | 8,760 a year, nights and weekends included |
| Absence | Holiday, sick leave, notice periods | None. It does not take leave and does not resign. |
| A volume spike | Overtime, temps, or another hire | The same system, more throughput |
| When they leave | The process knowledge walks out with them | Documented, in your repository, in your name |
| Time to productive | 3–6 months to hire, onboard and train | About a month for a standard scope |
Year one
An operations hire
€50,000–€70,000 fully loaded
The same work, automated
One fixed project fee, typically €12,000–€25,000
Year two
An operations hire
The same again, plus the raise
The same work, automated
Nothing. You own it outright.
Hours covered
An operations hire
About 1,700 a year, in office hours
The same work, automated
8,760 a year, nights and weekends included
Absence
An operations hire
Holiday, sick leave, notice periods
The same work, automated
None. It does not take leave and does not resign.
A volume spike
An operations hire
Overtime, temps, or another hire
The same work, automated
The same system, more throughput
When they leave
An operations hire
The process knowledge walks out with them
The same work, automated
Documented, in your repository, in your name
Time to productive
An operations hire
3–6 months to hire, onboard and train
The same work, automated
About a month for a standard scope
These ranges are typical for mid-market operations roles in Northern Europe and for the projects we deliver. They are not a quote. We work out the real number for your process (including what it is costing you today) before you commit to building anything.
Our flagship program
The AI-First Program
A fixed-scope path from scattered data to your first AI system in production. It opens with a free call, every phase is priced before it begins, and it is delivered mostly in writing (two meetings per phase, kickoff and handover) so your calendar stays yours.
-
01
Solution design
fixed price · agreed on the call
The project starts here. We interview the people who actually run the process, look at the systems behind it read-only, and write the design: what gets built and in what order, what it connects to, where a human stays in the loop, and the cost and the saving for each piece. The document is yours whether or not we build it, and if we do, this is phase one, not a preliminary.
-
02
Data Foundation
1–3 weeks · fixed price
We restructure the databases your operation already runs on so that people (and machines) can finally rely on them: clean schemas, one access layer, a full audit trail. You get a readiness report and a concrete roadmap of what to automate first.
-
03
First AI System
2–6 weeks · fixed price
Your first production AI system, deployed on your own infrastructure and your own AI accounts: strict guardrails enforced in code, complete decision logs, human sign-off where it matters. It takes over one workflow and runs it end to end.
-
04
Cutover & Handover
fixed price · then it is yours
One process at a time, we reconnect your operation to the new system, old and new run in parallel until the numbers match, then we cut over. You get the documentation, the dashboard, and the agent that keeps the system repaired and adjusts it when you ask. There is no retainer after this. If you want us on standby we can arrange it, but nothing is designed to need us.
A standard scope runs about a month end to end. Larger ones take longer and are staged so the operation never stops. Either way each phase stands on its own, stop after any of them and keep everything, because the code, the documentation, the infrastructure and the AI accounts are in your name from day one. There is nothing to hand back and nobody to keep paying.
Delivery
A month for the standard case. Staged for everything else.
Speed is a function of scope, not of ambition. We are explicit about which one you are in before you commit to anything, and the first call is usually enough to tell.
-
01
Standard scope: about a month
One process, systems we can reach through an API or a database, and a decision rule your team can state out loud. Audit, data layer, build, cutover, roughly four weeks, at a price agreed before we start.
-
02
Freeze the definitions first
For anything larger, the first move is agreeing what each number means, once, in writing. Nearly every migration that goes badly went badly here, months before anyone noticed.
-
03
Run both in parallel
Old and new take the same inputs and are compared automatically, every day, until the outputs agree. Nothing is switched on the strength of a demo or a spreadsheet of test cases.
-
04
Cut over one process at a time
Each process moves on its own day, with a rollback that takes an afternoon. Your operation keeps running throughout. There is no weekend where everybody holds their breath.
After delivery
Warranty first, then nothing you are obliged to pay
Most consultancies make their real margin on the maintenance contract, which quietly makes fragility profitable. We would rather build something that does not need us and be hired again because you want more.
Ninety days of warranty
From cutover. Anything that does not do what the specification says, we fix at our cost. No ticket count, no hourly billing, no argument about whose fault it is.
Then it repairs itself
Upstream systems change and formats drift; that is what normally kills an integration six months in. Ours notices, diagnoses the cause and repairs itself within limits agreed with you up front: a renamed field it handles alone, anything structural it proposes and waits for a human.
Support, paid and optional
If you want us reachable after the warranty, it is an hourly rate or a small prepaid block that you use or do not. No standing fee, no minimum, and nothing that renews itself automatically.
Extensions are quoted, not metered
A new process, a new integration, a change of scope: a fixed price for a defined piece of work, decided before it starts. You always know what you are agreeing to before you agree to it.
Straightforward to hire, easy to stop
It starts with a call that costs you nothing. After that: defined scope, defined price, defined delivery, and then it ends. Everything is built in your name and on your accounts, it comes with ninety days of warranty, the system repairs itself, and the agent we leave behind can change it when you ask. We would rather you call us back because you want more, not because the thing stops working without us.