AI & Data
Automation that handles
the exceptions
The happy path is 20 percent of the work. Everything that matters in back-office automation is the exception queue, the partial failure, the vendor who sends a scanned fax, and the auditor who wants to know who approved what.
Overview
Process automation that survives an exception, an audit and a system upgrade.
Intelligent automation here means a specific combination: document extraction where the input is unstructured, a decision layer with explicit rules, an orchestration engine that keeps state, and robotic process automation only where an API genuinely does not exist. Most engagements use three of the four, and the mix is decided by the input, not by a preferred vendor.
Common targets are invoice and remittance processing, goods receipt matching, KYC and onboarding packs, claims intake, order entry from PDFs and email, and reconciliation. These share a shape: high volume, structured outcome, unstructured input, and an existing team that knows every edge case and has never been asked to write them down.
We measure straight-through processing rate and cost per transaction, and publish both weekly during rollout. A process at 71 percent straight-through with a well-run exception queue beats one claimed at 95 percent where the failures are quietly absorbed by someone in finance at month end, unpaid and unrecorded.
- 78.4%
- Straight-through invoice processing at Halden Group, from 22%
- £1.36
- Cost per processed invoice, down from £4.90 at baseline
- 31 hours
- Weekly manual effort removed from goods receipt matching
Capabilities
What this covers
Six areas we staff properly. If your problem sits outside them, the honest note at the foot of this page says so.
Document extraction and validation
Field extraction from invoices, bills of lading, remittance advices and scanned forms, with per-field confidence and validation against master data. Low-confidence fields route to a reviewer, not into the ledger.
Decision rules and policy engines
Business rules held in a versioned, testable engine that finance and operations can read, rather than buried in code or a model. Rule changes are reviewable and every decision records which version applied.
Process orchestration
Long-running workflows with durable state, compensating actions and timers, so a process that stalls at step seven for two days resumes correctly rather than restarting or dying without anyone noticing.
Exception queue design
The queue is a product, not a dumping ground: ranked by value and age, showing the extracted fields beside the source document, with one-click correction that feeds back as labelled training data.
RPA where no API exists
Screen-driven automation built with retry logic, health checks and a maintenance budget stated upfront. We treat every robot as technical debt with an expiry date, and name the release that should remove it.
Controls and audit evidence
Segregation of duties, approval thresholds, immutable event logs and evidence exports your internal audit team can sample. Built against the control descriptions rather than reverse-engineered afterwards.
Deliverables
What you get
- Process map with volumes, exception taxonomy and cost per transaction baseline
- Document extraction pipeline with confidence thresholds and validation rules
- Versioned decision rules with a test suite your finance team can read
- Orchestration workflows with durable state and compensating actions
- Exception review console with correction capture and queue analytics
- Control matrix and audit evidence exports mapped to your policies
Stack
What we build it with
- Camunda
- Temporal
- Azure AI Document Intelligence
- Amazon Textract
- UiPath
- Drools
- Apache Kafka
- PostgreSQL
- Elasticsearch
- Python
- Keycloak
- Grafana
Process
How the engagement runs
Two-week increments against a written definition of done. You can stop at any increment boundary and keep everything built so far.
Process mining and baseline
We measure the current process from system logs rather than interviews: volumes, cycle time, rework rate and cost per transaction. The baseline is the contract for later claims.
Exception taxonomy
We sit with the team doing the work and enumerate every reason a case goes sideways. That list, usually 30 to 60 entries long, determines the whole design.
Pilot on one stream
One document type or supplier group, run in parallel with the manual process, so the two outputs can be compared line by line before anything is switched off.
Scale with controls
Further streams added as the straight-through rate holds, with controls, approvals and audit evidence validated by your internal audit function before volume increases.
Operate and shrink the queue
Monthly review of exception causes, each one either fixed at source, absorbed by a rule, or accepted as permanently manual and documented as such.
When this is the wrong engagement
If the process changes materially every quarter, or exists mainly to work around a system already scheduled for replacement, automating it locks in the workaround and the underlying system should be fixed instead.
FAQ
Questions we get asked
- Is this just RPA with a new name?
No, and we use RPA sparingly. Robots that drive a user interface break with every vendor release and accumulate maintenance cost. We prefer APIs, database integration and event streams, and where a robot is genuinely the only option we state what it costs to keep alive.
- What happens to the people currently doing this work?
In every engagement we have run, the team moves to the exception queue and to supplier or customer follow-up, which are the parts needing judgement. You need that redeployment plan before launch; automation announced without one gets quietly sabotaged, and fairly so.
- How accurate is document extraction on poor-quality scans?
On clean PDFs, field-level accuracy above 98 percent is routine. On faxed or photographed documents it drops sharply, sometimes below 85 percent on handwritten fields. We handle that with confidence thresholds and human review, not by pretending the number is higher.
- Can we start with one process to prove it?
Yes, and we recommend it. A single document type with 500 or more transactions a month gives a readable result in eight to ten weeks. Choose a process with a clear owner and a measurable baseline; low volume will not produce a signal worth arguing about.
Related