AI SYSTEMS
AI Inside the Workflow — Not Beside It
Governed AI for document understanding, classification, routing, extraction, copilots and decision support — built as one stage inside a deterministic process, not as a chatbot bolted onto the side of it.
AI is used where interpretation is required — surrounded by typed inputs and outputs, confidence, validation, human review, cost tracking, logs and deterministic actions.
A bounded processing component. It interprets; it does not act.
- In
- Document + schema
- Out
- Fields + confidence
- Document type98Accepted
- Counterparty96Accepted
- Reference number71Review
- Effective date94Accepted
Original beside extracted value. One flagged field.
Record written in the business system, document generated.
Verified and loggedIllustrative — not client data. Field names, confidence values and the threshold are synthetic.
Where interpretation is actually required.
Every one of these is a step where a person currently reads something and decides what it means. That is the only place a model belongs.
Document understanding and extraction
Turn scans, PDFs and attachments into typed fields against a schema.
Email / message classification
Decide what an inbound message is about before it reaches a queue.
Case routing
Send work to the team that can actually resolve it.
Data enrichment
Fill gaps in a record from sources a person would otherwise read.
Risk / exception detection
Surface the handful of items that do not look like the others.
Internal copilots and knowledge retrieval
Answer from your own documents, with the source shown.
Human-reviewed decision support
Propose the decision and the reasoning; a person still makes the call.
AI-generated structured outputs
Produce a typed object a downstream system can act on, not prose.
The model is the smallest part of the system.
What makes an AI step safe to run in production is everything around it. Each of these controls exists to prevent a specific failure.
Confidence threshold
Prevents a low-certainty guess being treated as a fact. Below the line, the field routes to a person.
Human verification
Prevents an unreviewed interpretation reaching a customer or a ledger. Source and extraction sit side by side.
Approved action
Prevents the model from acting. It proposes structured data; a deterministic step performs the write.
Audit record
Prevents an unexplainable outcome. Input, output, confidence, reviewer and time are all recorded.
Model and cost tracking
Prevents silent drift and runaway spend. The model version is pinned and the cost per run is visible.
Fallback behaviour
Prevents a stalled operation. If the model is unavailable or unsure, the work continues down a human path.
What the model is never allowed to do
- Write directly to a system of record — a deterministic step does that.
- Act on a field that fell below the confidence threshold without review.
- Run against an unpinned model version in production.
- Complete a run without leaving an audit entry.
These are constraints in the workflow, not policies in a document. If the guard is missing, the run does not start.
One governed stage, end to end.
Document processing is the clearest example of the pattern: interpretation in the middle, deterministic steps either side, and a person on the exception.
AI document processing in production
An information-heavy intake process where documents arrive in inconsistent formats. Extraction is typed and scored, low-confidence fields are held for side-by-side review, and only approved data reaches the record system — so an unreviewed interpretation cannot become an operational fact.
- Document extraction
- Human-in-the-loop
- Audit trail
Often the honest answer is that you do not need AI.
A model is the expensive, non-deterministic option. If a rule can do the job, a rule is the better engineering decision — and it is cheaper to run.
AI adds value when
- The input is unstructured
- Free text, scans, mixed formats, inconsistent layouts.
- Meaning has to be inferred
- Two correct inputs can look completely different.
- The rule set would never end
- Every new case adds another exception to the logic.
- A person can check the output
- There is a reviewer and a definition of correct.
Deterministic automation is enough when
- The input is already structured
- A form, an API payload, a database row.
- The logic is stable
- The rule has not changed in a year and nobody argues about it.
- The outcome must be identical every time
- Finance, compliance, contractual thresholds.
- There is no reviewer
- Nothing should be non-deterministic if nobody is checking it.
The Operations Review will say which of the two your workflow needs — including when the answer is neither.
Models and components we build with
- OpenAI
- Anthropic
- Azure OpenAI
- OCR and document parsing
- Vector search
- Typed schema validation
- n8n
- Custom APIs
The model is a replaceable component. The schema, the thresholds, the review step and the audit record are the parts that have to outlast it.
Bring one information-heavy workflow.
The review will determine whether AI adds real value here, or whether deterministic automation is already enough. Either answer is a useful one to get before you build.