01
Business questions
Natural language questions over Power BI, SAP and spreadsheets, with access profiles and every answer citing its source.
Pillar 3 · Agents on your data and processes
Agents that do real work, not demos.
They run on the platform your company already pays for, inside your infrastructure, with a record of every decision. They scale only once the number is visible.
The difference between answering and doing
A generic assistant answers about the internet. An agent on your data answers about your business, with the figure and its source. Useful, but still just a query.
The next jump is execution: matching invoice against purchase order and goods receipt, entering the emailed order into the ERP, building the reporting somebody rebuilds by hand every week. It is what RPA never solved, because it requires reading formats and deciding on exceptions.
Gartner estimates that more than 40 % of agent projects will be cancelled before 2028. That is why this starts with a measured pilot: if it works it scales, and if not, we do not move forward.
What we build
02
Process automation
Invoice, purchase order and goods receipt matching. Orders from email into the ERP with their approval flow. Document validation in the field.
03
Document reading
Extraction from PDFs and forms in different formats, with criteria that require judgement and exceptions a rigid flow does not cover.
04
Model agnostic
Claude, Gemini or Copilot Studio, depending on the case and the platform your company already has under contract. The recommendation does not depend on a commission.
Blumar case
BluHarina management agent and artisanal quota and debt agent: both live in 4 weeks.
How what the agent does is controlled
An agent with write permission is governed like an identity with operational capability, not like a desktop tool.
Least privilege
The agent accesses what it needs for its use case and nothing more, under the same access profiles that apply to people.
Action logging
Every decision leaves a trail: what it read, what it wrote, on what criterion and who approved it. It is what an auditor asks for and what a change committee accepts.
A way to stop it
Scope limits and a switch. An agent that cannot be stopped is not ready to touch a production system.
How it starts
A bounded pilot, one data domain and one management user.
The result is measured in cost avoided, the currency a committee understands without translation. In weeks, not quarters.
- One process, not a whole department
- A real user who uses it every day, not a demo
- Measurement from day one, against how it was done before
- If it does not work, it stops there and the learning is written down
Frequently asked questions
- Which platform do you build on?
- The one your company already pays for. Most projects run on Microsoft 365 and Copilot Studio because that is what is already contracted; when the case justifies it we use Claude or Gemini. The choice is explained before building, with its cost.
- How is consumption cost controlled?
- It is estimated before building and instrumented afterwards. The cost does not depend only on what it does: it depends on who talks to it, with which licence and how often. The calculation is handed over in writing with the pilot.
- Do I need my data in order first?
- For the pilot's domain, yes, and it is usually half the work. That is why we pick a bounded domain where getting it in order is achievable in weeks. If it is not in shape, we say so before starting.
- Does this replace people?
- It is not what we sell nor what we have seen. What changes is where the time goes: fewer hours building reports and matching documents, and the same people on the exceptions, where their judgement is worth something.
The pilot is worth it even if it never scales.
It ends with a number from your own operation, compared against how it was done before. That number is what justifies everything that comes after.