Artificial intelligence
Very useful, in one very specific place.
AI is exceptionally good at reading messy text and turning it into tidy data. Almost everything we build with it fits in that sentence. The rest of the time there's a cheaper and more predictable answer.
From €700
1-2 weeks
Where it helps
When the input has no format.
An ERP handles an order perfectly well when it arrives in the format it expects. The problem is that orders arrive as emails written by different people, as supplier PDFs whose layout changes every quarter, and as WhatsApp messages. That's where AI earns its keep.
Classification
Sorting incoming items by type, urgency or department without writing hundreds of manual rules.
Information extraction
Pulling references, quantities, dates and amounts out of documents that follow no fixed template.
Document analysis
Reviewing contracts, delivery notes or tender documents and flagging whatever departs from the norm.
Internal assistants
Answering questions about the company's own documentation, with a citation showing where the answer came from.
Email processing
Reading, sorting and drafting a reply, leaving the send button to a person.
Search across documentation
Finding the right detail in years of files without knowing which folder it's in.
Specialised agents
An assistant with one narrow job and limited access to what it needs, rather than a general chatbot plugged into the whole company.
ClassificationInformation extractionDocument analysisInternal assistantsEmail processingDrafted responsesSearch across your documentationSpecialised agents
Where it doesn't
Where we don't use it.
This is the part most sites skip, and it's the part worth reading. Saying no to AI in the wrong place saves a lot of money.
Calculations and fixed rules
If the result can be computed, compute it. A model that is occasionally wrong shouldn't decide an amount that always has the same formula.
Decisions with legal or financial consequences
Approving a payment, signing, applying an exceptional discount. AI can prepare the file; the decision belongs to a person.
Processes that aren't defined yet
If nobody can explain how the task is done, adding AI only makes the mess faster.
Clean, structured data
If the data is already in a database, a query does it better, cheaper and identically every time.
As soon as a task stops needing interpretation, we take it out of the model and move it into ordinary code. It's cheaper to run and gives the same answer every time.
How we set it up
With the person in the right place.
- The model proposes and a person confirms, for as long as the error rate is unknown.
- Everything the system decides is logged and reviewable.
- Accuracy is measured on real cases before supervision is removed.
- Sensitive data is handled with judgement: not everything needs to leave the company.
Got a process buried in documents?
That's usually where AI pays for itself. Tell us about it and we'll say whether it applies to you.