When the bottleneck is the information coming in
An insurance company processes 14,200 claims a month. The average claim takes 8.4 days to process. 12.7% are rejected on first pass and sent back for rework, while each claim costs $38 to handle.
The team has already completed two Six Sigma projects. They improved handoffs, clarified responsibilities and reduced avoidable delays. Yet progress has plateaued.
Why? Because the main constraint is not necessarily human error. It may be the unstructured data arriving at the front door: scanned forms, PDFs, email attachments and supporting records that employees must interpret, re-key and reconcile before a claim can move forward.
That is where AI in finance operations can change the shape of the problem. Intelligent document processing (IDP), machine learning and workflow automation can help turn incoming information into structured, reviewable data. But the opportunity is not simply to automate more tasks. It is to identify where automation fits, prove that it improves the process, and put the right controls around it.
This guide maps an insurance claims back-office improvement to DMAIC (Define, Measure, Analyse, Improve and Control) and shows how process mining finance data can make hidden workflow behaviour visible.
A note on the figures: The insurance company and results in this article form an illustrative worked example based on the scenario figures provided. They are not presented as independently verified results from a named insurer.
The finance AI toolkit, mapped to DMAIC
AI works best when it supports a clearly defined improvement method. DMAIC helps a team avoid automating an unclear process or mistaking a faster step for an improved end-to-end result.
Define: agree on the problem and the boundaries
Before selecting a model or automation tool, define what the project is intended to improve. For this example, the initial problem statement might be:
Claims take an average of 8.4 days to process, with 12.7% rejected at first pass and an average handling cost of $38 per claim.
Then clarify the project boundaries. Does it cover one claim type, a particular region, or all claims? Does “processing time” start at claim receipt or at the point a complete document pack is available? What counts as a first-pass rejection? Which decisions remain with an authorised claims professional?
These details determine whether the baseline and subsequent results are comparable. They also draw an important line: an AI system extracting a policy number from a form is not the same as a system making a decision that affects a claimant’s entitlement.
Measure: build reliable visibility into the intake process
The Measure phase establishes what is happening now. In a document-heavy operation, measurement can start by using optical character recognition (OCR) and natural language processing (NLP) to extract data from scanned claims, PDFs and emails, reducing the need for manual re-keying.
The team can also automate timestamp extraction from workflow logs. That helps build a more precise view of when each event took place: document received, claim registered, information requested, review started, approval completed and payment issued.
Useful baseline measures include:
- End-to-end processing time, with a consistent start and stop definition
- First-pass rejection and rework rates
- Cost per claim, with cost assumptions documented
- Manual touches and handoffs per claim
- Document extraction accuracy by field and document type
- Straight-through processing (STP) rate
- Volume and age of work waiting in queues
Measurement should account for data quality. OCR output is only useful when the team knows where it is reliable, where it needs review and how errors are counted.
Analyse: find the causes and the real process
In the illustrative scenario, the team applies machine learning to three years of claims data and finds that 71% of first-pass rejections trace to four document fields. This is a strong lead for investigation, not proof, by itself, that those fields cause every rejection. The team still needs to check definitions, data quality, claim mix and the way rejection reasons were recorded.
A useful analysis combines several perspectives:
- Field-level analysis: Which missing, inconsistent or misread fields are associated with rework?
- Anomaly detection: Can the system flag potential duplicate claims or unusual patterns for investigation?
- Process mining: What do workflow logs reveal about actual pathways, queues and repeat work?
In this example, process mining reveals 14 workflow variants that were not documented. Some may be legitimate routes for different claim types; others may represent avoidable loops or inconsistent handling. A process map based on assumed procedures would not necessarily expose those differences.
Process mining does not improve a process on its own. Its value is that it helps the team see how work actually flows, then investigate the variants that matter. For broader context, a published Suncorp process mining case study describes analysis of claims workflow data to identify opportunities for process improvement. As with any case study, results from one organisation should not be assumed to transfer directly to another.
Improve: redesign intake before scaling automation
Once the team has identified the causes of delay and rework, it can test a redesigned intake flow. Intelligent document processing claims typically combine document classification, OCR, NLP, validation rules and a route for exceptions. The system can extract relevant fields, compare them with expected formats and send uncertain or incomplete records for human review.
Robotic process automation (RPA) may handle repetitive, rules-based actions, for example, transferring a validated field into a workflow system. It should not be used to conceal unclear process rules or bypass necessary approvals.
Simulation can help compare proposed designs before a live rollout. The team might test different confidence thresholds, pre-check queue capacity or routing rules to understand how each option could affect flow and review workload. A controlled pilot then provides evidence about what happens in practice.
A practical sequence is:
- Select one claim segment and a defined document set.
- Set field-level accuracy and exception thresholds.
- Keep a human review route for low-confidence or high-impact cases.
- Compare the pilot with a suitable baseline or control group.
- Review errors, rework, claimant impact and staff workload, not just speed.
Control: keep performance visible after launch
A successful pilot is not the end of the project. The team needs ongoing measures to detect whether the process remains within its intended performance range.
At intake, a predictive model might score the likelihood of rejection and route higher-risk claims to a pre-check queue. That score should guide attention, not silently make an adverse decision. The team can also monitor anomalies in process KPIs and use an automated control plan to assign owners, thresholds and response actions.
Control measures could include:
- Field-level extraction accuracy and exception rates
- Rejection, rework and payment error rates
- Claim cycle time and queue age
- Model overrides and human-review outcomes
- Outcomes by claim type, channel and relevant customer group
- Model or data changes, with validation and approval records
For any AI application affecting customers, governance should reflect the use case, its level of autonomy and its potential impact. The EIOPA opinion on AI governance and risk management discusses safeguards including data governance, human oversight, explainability, accuracy and robustness. Insurers should also assess applicable laws and regulatory guidance in each jurisdiction where they operate.
Worked example: what the improvement could look like
Suppose the redesigned process is piloted and validated against the team’s agreed definitions. The illustrative before-and-after results are:
| Measure | Before | After |
|---|---|---|
| Average processing time | 8.4 days | 3.1 days |
| First-pass rejection rate | 12.7% | 3.4% |
| Processing cost per claim | $38 | $19 |
| Straight-through processing rate | 22% | 68% |
At the stated volume of 14,200 claims per month, the first-pass rejection rate equates to about 1,803 rejected claims before the improvement and about 483 afterward. That is roughly 1,320 fewer first-pass rejections per month, assuming the claim volume and rejection definitions remain comparable.
The $19 reduction in processing cost per claim implies a potential difference of $269,800 per month at the same volume. That is a useful estimate for a business case, not a guaranteed cash saving: finance should confirm whether the cost reduction represents budget savings, released capacity or both.
The improvement also comes with real implementation costs. The business case should account for:
- IDP licence and ongoing operating costs
- Workflow redesign, system configuration and testing
- Data privacy and security review
- Training and change management for the 40 affected FTE
- Ongoing model monitoring, exception handling and governance
A credible benefits case compares these costs with validated benefits, including service outcomes and capacity released, not just a projected reduction in manual touches.
What can go wrong, and how to plan for it
AI can improve the flow of information, but it introduces risks that DMAIC teams need to make visible.
OCR accuracy can fall on handwritten or poor-quality scans. Measure accuracy by document type and field, not as a single blended percentage. Send low-confidence records to a review queue and track the errors that reach downstream steps.
A model trained on one claim type may not generalise. Differences in forms, terminology, languages or evidence requirements can change performance. Validate separately across relevant claim segments before expanding scope.
People may resist a process that changes their roles. Staff may reasonably question how automation will affect their work or whether they can trust its output. Involve the people doing the work in process design, explain what automation will and will not decide, and provide practical training for new review and exception-handling tasks.
Regulators may question automated decisions. Keep a clear record of what the system recommended, what a human reviewed, what action followed and why. Build escalation and appeal paths into the process, particularly when a decision could adversely affect a claimant.
The principle is straightforward: automate routine handling where evidence supports it; preserve accountable human judgement for uncertainty, exceptions and high-impact decisions.
Where Ci Flow fits: from insight to managed improvement work
The AI models in this example live in the insurer’s finance analytics and workflow environment. Ci Flow is a separate product: it is where the improvement team can manage the work of turning those insights into a governed, measurable improvement project. It is not presented here as the system hosting the insurer’s claims models.
A finance operations team can use a shared improvement workspace to keep project stages, actions, decisions, risks and benefits connected. For a DMAIC initiative, that can mean tracking the problem definition, baseline, root-cause evidence, pilot plan, control measures and phase-gate reviews in one place.
Ci Flow describes support for AI-assisted DMAIC project tracking, digital workspaces, portfolio reporting and real-time KPI boards for daily management. Teams can use these capabilities to connect operational evidence to assigned improvement actions and maintain visibility of work across a portfolio, while keeping the model execution in the organisation’s finance analytics stack.
For process owners, project tracking is most useful when it keeps the evidence and the next action together: a queue metric changes, an owner investigates, a countermeasure is tested, and the result is reviewed against the baseline. Ci Flow’s platform information outlines its methodologies, project workspaces, portfolio views and tools. Its security page provides details about its security arrangements; organisations should review those details as part of their own supplier and data-security assessments.
Build capability, then scale with control
For shared services leaders, finance operations managers and AP/AR process owners, the lesson is not “add AI and expect a better process.” It is to use AI in finance operations as part of a disciplined improvement effort: measure the current state, analyse actual workflow and root causes, test targeted automation, and control performance after implementation.
That is how an AI Lean Six Sigma back office can pursue faster processing without losing sight of accuracy, customer impact or accountability.
Explore Ci Flow to manage improvement work from project evidence through sustained control.
Build your capability with Lean 6 Sigma Hub’s CSSC-accredited Green Belt or Black Belt certification, featuring practical case studies and applications for process improvement.
Kaizen. Kai-Care. Kai-Done. Lean Six Sigma
