Architecture note 02
Applied AI
Supervised document extraction with a human-approval boundary.
- Pattern
- Extract, check against rules, approve
- Controls
- Evaluation on your own documents, audit log
- Boundary
- Nothing reaches the ERP without approval
AI built into real work, and tested before it gets there.
Knowledge search, document processing and supervised workflows that prepare actions for a person to approve, measured on your own data before they touch operations.
- Typical starting point
- A clear, repetitive, text-heavy task: purchase orders, support triage, contract review or policy search.
- You receive
- A deployed workflow or search tool, an evaluation report on your own data, logging and an agreed approval boundary.
- First step
- A feasibility assessment of one specific use case.
Sound familiar?
- “Our team spends hours copying information from documents into the ERP.”
- “Nobody can find the current version of a policy.”
- “Our AI pilot worked in the demo and nowhere else.”
What we build
Narrow, measurable, supervised.
Useful AI starts with one defined task and a clear idea of what a correct result looks like. We build it on data that has already been modelled and checked, which is why it behaves the same way in production as it did in testing.
-
01
Knowledge search
Retrieval over contracts, procedures or support history, combining keyword and semantic search, with a reference to the source of every answer.
-
02
Document extraction
Fields from purchase orders, invoices and contracts turned into structured records and checked against your rules.
-
03
Supervised workflows
A request is read, checked against your data and prepared as a draft or an action in the ERP or CRM, for a person to approve.
-
04
Integrations
Scoped access to the tools your team already uses, with logging of every input, output and approval.
-
05
Evaluation
A test set built from your own documents, run on every change, so quality is measured, not assumed.
-
06
Custom or fine-tuned models
Only when your data and the task justify it. Most problems are solved well without one.
Reference architecture
Documents flow through extraction and checks. Anything unusual goes to a person, and nothing reaches the ERP without approval.
How we keep it safe
Designed around the cost of a mistake.
-
01
Human approval where it matters
For consequential actions the system prepares and a person approves. The boundary is agreed with the business before go-live and can only widen when the measurements support it.
-
02
Deployment matched to data sensitivity
We agree the data flow, provider terms and access controls before building. Where needed this means enterprise endpoints or self-hosted models.
-
03
Audit trails
Inputs, retrieved sources, outputs and approvals are logged, so behaviour can be inspected and improved.
-
04
A way back
Every workflow can be switched to manual handling at any time without losing work in progress.
What we will not promise
- Autonomy.
Systems prepare work; people stay responsible for the decisions that carry risk.
- Zero errors.
Every model makes mistakes. We measure how often and route the uncertain cases to people.
- Fewer people.
The aim is to give skilled staff their time back for the work that needs judgement.
Next step
Start with one task.
Bring one repetitive process. We assess whether AI, conventional software or a mix is the right answer, and what it would need to work in production.
Discuss a use case →