Intelligent Automation
Automate the decision, not the mouse
Document understanding, durable workflows and decision services that survive real input — calibrated confidence, a small exception queue, and posting into systems of record that is idempotent. Every automated decision leaves a replayable audit trail.
Engagement at a glance
- Capabilities
- 05
- Deliverables
- 08 artefacts
- Stack groups
- 04
- Ways to start
- 03
Engagement options
- Automation teardown3 weeks
- First straight-through line10–14 weeks
- Automation platform6 months+, rolling
What breaks
The failure modes we keep meeting
Cheap to design around at the start. Expensive to discover once the system is carrying real traffic.
The bot breaks when the screen moves
Screen-scraped automation binds to pixel positions and DOM selectors that were never a contract. A vendor ships a UI patch, the ERP moves a field, someone switches on a Citrix session at a different resolution — and Monday morning the queue is empty because the bot has been clicking on nothing since Friday. Integration belongs at the API, not the glass.
Confidence scores that mean nothing, so everything gets reviewed
Most extraction stacks emit a raw model score and call it confidence. Nobody measures calibration, so 0.9 is picked because it looks like a lot. The result is an automation programme with a licence fee, a vendor, a steering committee — and a human still opening every single case.
No trail, and the process was wrong to begin with
Six months in, audit asks why a particular invoice was paid. The answer lives in a bot log that rotated, a rule someone edited in a console, and a model nobody versioned. Worse, the fourteen-step approval chain was never questioned — automation has simply made a bad process faster and harder to change.
Capabilities
What the work actually consists of
01
Document understanding and IDP
Invoices, KYC packs, claims, purchase orders, contracts. Extraction is schema-constrained and every field carries provenance back to the page region it came from, so a reviewer can see what the system read rather than trusting a number.
- Layout-aware extraction into a strict JSON schema — Textract or Document Intelligence for structured forms, a vision model such as Claude or Gemini for the long tail and for contract clause reading
- Classification and splitting of mixed scan batches, with page-level routing so a 400-page bundle becomes typed documents rather than one blob
- Line-item table extraction reconciled against header totals — a mismatch is an exception signal, not a silent overwrite
- A labelled golden set per document type (typically 300–500 documents), with inter-annotator agreement measured before any accuracy claim is made
- Template-drift detection on vendor-by-vendor field accuracy, so a supplier changing their invoice layout raises an alert instead of a quiet regression
02
Durable workflow orchestration
A case that spans days, three systems and two humans is a distributed transaction wearing a business-process costume. We model it as durable execution so a deploy, a timeout or a downstream outage does not lose state.
- Temporal or Restate workflows with durable timers, bounded retries and idempotency keys on every external call
- Saga-style compensation for partial failures — a posted ledger entry with a failed downstream notification is reconciled, not orphaned
- Versioned workflow definitions so month-long cases started on the old code path complete correctly after a release
- Signals and queries for mid-flight human decisions, and timers that escalate on SLA breach rather than waiting for someone to notice
- Replay tests that run recorded histories against new code in CI — the only honest way to catch a non-deterministic workflow change
03
Decision automation — rules plus models
Policy belongs in a versioned, testable decision layer, not scattered across application code and a model’s output. Models supply scores and extracted facts; rules decide. That ordering keeps the decision explainable when it matters.
- Business rules as versioned artefacts — DMN in Camunda, or a decision-table DSL in code with full unit coverage
- Model scores enter as features into the rule layer, never as the final authority on a policy question
- Shadow mode and champion/challenger on every rule change, with a simulation run against replayed historical cases before release
- Decision logs capturing inputs, rule version, model version, threshold set and outcome as a single immutable record
04
Straight-through processing and the exception desk
The point of the work is a small, well-routed exception queue — not the absence of one. Thresholds are derived from calibration curves and the cost of being wrong, set per field and per document class rather than globally.
- Confidence calibration measured properly (reliability curves, expected calibration error) and thresholds chosen against error cost, not intuition
- Exception queues with reason codes, SLA clocks and routing by skill — so backlog can be diagnosed instead of merely counted
- A review console that puts the source region beside the extracted field, keyboard-first, built for hundreds of cases an hour
- Every correction captured as a labelled example and fed back into the next evaluation run
- Straight-through rate, review minutes per case and sampled silent-error rate carried as SLOs with alerting
05
Integration and audit
Automation that cannot be reconciled is a liability. We post through published interfaces, make every write idempotent, and keep an append-only record that answers the auditor’s question without a forensic exercise.
- Integration via SAP OData and BAPI/IDoc, Oracle Fusion REST, Dynamics and NetSuite APIs, Salesforce Bulk API — GUI automation only where genuinely no interface exists
- Idempotent posting with correlation keys, plus nightly reconciliation that detects duplicates, gaps and drift against the system of record
- Append-only decision and document event log, hash-chained where the regulatory bar requires tamper evidence
- PII redaction, residency routing and retention policy applied at ingestion rather than bolted on at review time
- Per-case evidence packs — source document, extracted fields, rule and model versions, reviewer actions — exportable for audit
Outcomes
Targets, and where each number comes from
The basis line under every figure is the point of this section. Where a number is a design standard rather than a measured delivery, it says so — we would rather be checkable than impressive.
- Straight-through rate on high-volume, well-structured document flows
- 60–85%Straight-through rate on high-volume, well-structured document flowsTypical range for mature IDP programmes on a single document type after calibration — not a launch-week figure, and lower for contract-style documents
- Silent error rate on auto-approved cases
- < 2%Silent error rate on auto-approved casesDesign target agreed per engagement and enforced by a sampled QA audit of auto-approved decisions
- Cycle time for cases that clear threshold
- Minutes, not daysCycle time for cases that clear thresholdStructural, not measured — most of the elapsed time in a manual flow is queue time rather than work time
- Replayable from inputs, rule version and model version
- Every decisionReplayable from inputs, rule version and model versionDesign target — a property of the append-only event log we build, verified by replay tests rather than claimed from a delivery
Stack
What we reach for, and when
Defaults, not dogma. The list below is what we would propose on a blank page; an existing estate, a procurement constraint or a team's operating experience all legitimately move it.
- Amazon Textract
- Azure AI Document Intelligence
- Google Document AI
- Claude and Gemini vision models
- LayoutLMv3 / Donut (self-hosted)
- PaddleOCR, docTR
- PyMuPDF, pdfplumber
- Temporal
- Restate
- AWS Step Functions
- Camunda 8 / Zeebe
- Apache Kafka, Amazon SQS
- Airflow for batch ingestion
- DMN (Camunda), GoRules Zen
- MLflow model registry
- OpenFeature for staged rollout
- scikit-learn calibration tooling
- Great Expectations for input contracts
- SAP OData, BAPI/IDoc
- Oracle Fusion REST, NetSuite SuiteTalk
- Salesforce Bulk and Platform Events
- PostgreSQL append-only event tables
- Next.js review console
- OpenTelemetry, Grafana, Loki
- HashiCorp Vault, AWS KMS
Deliverables
What you keep
Everything below lands in your repositories and your accounts, under your licence, with the reasoning written down. There is no runtime you have to keep renting from us.
08 artefacts, handed over
- Process teardown: volumes, touch time, rework and error rates per step — with the steps we recommend deleting before anything is automated
- Labelled golden set per document type, with an inter-annotator agreement report and a held-out evaluation split
- Extraction service with published schema, calibration report and a per-field, per-class threshold policy
- Durable workflow definitions in code, with replay tests covering timeout, partial-failure and compensation paths
- Decision service with versioned rules, decision logs and a shadow-mode comparison report against current practice
- Exception review console with reason codes, SLA routing and correction capture wired back into evaluation
- Integration adapters with idempotency keys and a nightly reconciliation job against the system of record
- Runbook and SLO pack: straight-through rate, review backlog, sampled silent-error protocol, kill-switch and escalation paths
Engagements
Three honest ways to start
Sized so the first one can end. Each option is designed to produce something usable even if we never work together again.
Automation teardown
3 weeks
A process everyone complains about and nobody has measured, or an existing RPA estate with an unexplained break rate.
Includes
- Volume, touch-time and rework baseline drawn from real logs, not interviews alone
- Document sample audit — variability, quality, template count, exception drivers
- Automation feasibility ranking with the honest straight-through ceiling for each candidate
- Target architecture and a costed first increment
First straight-through line
10–14 weeks
One document type or one case type taken end to end into production, with the platform pieces the second one will reuse.
Includes
- Extraction, calibration and threshold policy for the chosen document type
- Durable workflow, decision service and system-of-record integration
- Exception review console and reason-code taxonomy
- Shadow run against live volume before any auto-approval is switched on
- SLOs, dashboards, runbook and handover to the operating team
Automation platform
6 months+, rolling
Several processes queued behind the first, needing shared orchestration, review tooling and governance rather than one-off builds.
Includes
- Shared workflow, decision and review substrate with per-process configuration
- New process onboarding at a predictable cadence
- Model and rule change governance, including shadow-mode gates
- Quarterly calibration review and threshold retuning as document mix drifts
Questions
The awkward ones
The questions that decide whether this is worth starting — answered the way we would answer them on a call.
No. Keep the bots where a GUI genuinely is the only interface, but wrap them as activities inside a durable workflow so failures are visible, retried and compensated. Then measure break rate per bot and retire the worst offenders onto APIs first. Migration is ranked by fragility, not by ideology.
Adjacent work
These are the practices this one leans on and the reference architectures that show it assembled.
Related solutions
Reference architectures
Next step
Bring us the problem you keep deferring.
A 45-minute working session on your Intelligent Automation work. We will tell you what we would build, what we would not build, and roughly what it costs. No deck.