Insurance Underwriting Automation: Architecture, AI Models, and ROI (2026)

Key Takeaways

• Architecture matters as much as speed. Effective insurance underwriting automation needs integrated architecture, explainable decisioning, reliable data, and measurable outcomes not just faster workflows.
• A rules-plus-ML hybrid model is the practical foundation. Business rules enforce eligibility and compliance while machine learning models identify risk patterns and support pricing neither one works well alone.
• Four distinct AI model types do the work. Predictive AI, computer vision, NLP/document intelligence, and generative AI each handle a different part of the underwriting lifecycle no single model covers everything.
• Straight-through underwriting has limits by design. It performs best on standardized, lower-complexity risks; unusual, high-value, or ambiguous submissions still require underwriter judgment and human oversight.
• ROI depends on tracking the right metrics. Quote turnaround time, referral rates, automation rates, override frequency, and portfolio outcomes matter more than automation speed alone.

Insurance underwriting automation integrates data ingestion, validation, risk scoring, decisioning, and policy workflows. It connects core insurance platforms with rules engines, AI models, and external data sources. Insurance underwriting automation in 2026 looks markedly different from earlier pilots. Insurers are moving beyond isolated automation tools and limited proof-of-concept deployments, and now require governed systems that support production-scale underwriting across multiple insurance products.

A rules-plus-ML hybrid underwriting model provides a practical foundation for these systems. Business rules enforce underwriting guidelines, eligibility requirements, and regulatory controls. Machine learning models identify risk patterns, estimate exposure, and support pricing decisions. Modern platforms also process unstructured submissions, property images, loss histories, and third-party risk signals. Application programming interfaces transfer this information into policy administration and underwriting workbenches.

Automation handles predictable risks through straight-through underwriting. Complex, high-value, or exceptional submissions remain with experienced underwriters for further assessment. Regulatory governance must operate across the complete AI lifecycle. The NAIC AI Model Bulletin expects insurers to maintain controls covering fairness, transparency, documentation, and risk management.

Therefore, effective insurance underwriting automation requires more than faster workflows. For P&C underwriting AI, this means integrated architecture, explainable decisioning, reliable data, and measurable business outcomes.

What Is Insurance Underwriting Automation?

Insurance underwriting automation uses software, business rules, data integrations, and AI models to evaluate risk. It supports decisions from submission intake through pricing, approval, referral, and policy issuance. The system collects applicant information from forms, emails, documents, images, and connected databases. Document intelligence converts unstructured files into validated fields for downstream processing.

Rules engines check eligibility, coverage limits, exclusions, and underwriting guidelines. Predictive models assess risk probability using historical and real-time data. Decision engines then approve, decline, price, or refer each submission. Straight-through underwriting processes eligible cases without manual intervention. Exception workflows route complex or uncertain risks to human underwriters with supporting evidence.

Unlike basic workflow automation, insurance underwriting automation coordinates decisions across the complete underwriting lifecycle. It connects policy administration systems, external data providers, AI services, and underwriter workbenches. This approach improves processing speed while maintaining governance and human oversight. It also creates consistent, traceable decisions across products, channels, and underwriting teams.

In practical terms, automation handles repetitive analysis. Underwriters retain authority over exceptions, complex risks, and commercially sensitive decisions.

Why Insurance Underwriting Automation Matters

Insurance underwriting automation helps insurers process growing submission volumes without increasing operational complexity. This is especially true for commercial underwriting automation, where submission complexity and data variability are highest. It replaces fragmented activities with connected, governed, and measurable underwriting workflows.

Manual underwriting depends heavily on emails, spreadsheets, documents, and disconnected external databases. These processes delay risk assessment and create inconsistent decisions across teams.

Automation addresses these limitations through several business improvements:

• Faster decisions: Automated intake and verification reduce submission processing and quote turnaround times.
• Higher productivity: Underwriters spend less time collecting data and reviewing routine applications.
• Consistent risk selection: Centralized rules apply underwriting guidelines uniformly across products and distribution channels.
• Improved pricing accuracy: Predictive models combine historical experience with current risk indicators.
• Lower operating costs: Straight-through processing reduces manual effort across eligible, lower-complexity submissions.
• Stronger governance: Decision logs support compliance reviews, model monitoring, and internal audits.
• Better broker experiences: Faster responses improve submission visibility and strengthen distribution relationships.

The strongest business case combines efficiency, pricing discipline, and portfolio performance. Faster processing alone cannot correct unreliable data or poorly governed models.

Insurers therefore need scalable architecture, explainable decisions, and structured human oversight. These capabilities convert workflow automation into sustainable underwriting value.

How Insurance Underwriting Automation Works

Automated underwriting coordinates data, rules, AI models, workflow engines, and policy administration systems. Each component supports a defined stage within the underwriting lifecycle.

The process follows six connected stages:

Application programming interfaces connect submission channels with internal and external data sources. Document intelligence extracts information from forms, emails, spreadsheets, images, and supporting records.

Rules engines validate eligibility and enforce underwriting guidelines. Predictive models estimate risk using approved variables and historical outcomes. Decision engines combine these results to approve, decline, price, or refer submissions. Workflow orchestration then assigns tasks, records decisions, and updates policy administration systems.

Human review remains essential for complex, uncertain, or high-value risks. Underwriters receive structured evidence, model explanations, and identified exceptions. This connected process reduces manual handoffs while preserving accountability across every underwriting decision.

1. Data Collection and Risk Assessment

Brief Overview

Data collection captures the information required to evaluate an applicant's exposure and insurability. Sources include proposal forms, broker emails, loss histories, property records, images, and connected devices.

How Automation Supports the Process

NLP submission ingestion extracts relevant fields from structured and unstructured documents. Validation rules identify missing values, inconsistent entries, duplicate submissions, and unsupported file formats.

APIs transfer standardized information into underwriting platforms. Initial risk assessment then compares submission details against product eligibility and risk appetite requirements.

Key Business Benefits

• Reduces manual data entry and document review
• Improves submission completeness before assessment
• Accelerates initial risk classification and triage
• Creates consistent data for downstream models
• Identifies missing information earlier

Example or Use Case

A commercial property insurer receives applications through emails containing forms, spreadsheets, and photographs. Automation classifies each document and extracts location, occupancy, construction, and coverage information.

The system checks mandatory fields and flags conflicting property details. Complete submissions continue automatically, while exceptions return to brokers or underwriters for clarification. This workflow improves intake speed without weakening data controls.

2. Data Enrichment and Verification

Brief Overview

Data enrichment supplements submitted information with trusted internal and external risk data. Verification confirms whether applicant details are accurate, complete, current, and suitable for decisioning.

How Automation Supports the Process

APIs connect underwriting platforms with property, vehicle, identity, credit, weather, geospatial, and claims databases, often sourced from providers such as Verisk and LexisNexis Risk Solutions. Matching services compare external records against information provided within the application.

Automated controls identify address mismatches, outdated valuations, ownership conflicts, and undisclosed loss histories. Confidence scores determine whether verified data proceeds automatically or requires human review.

Key Business Benefits

• Improves risk-data completeness and reliability
• Reduces manual third-party database searches
• Detects inconsistencies before risk scoring
• Supports faster and more accurate decisions
• Strengthens fraud and misrepresentation controls

Example or Use Case

A personal property insurer receives an application with basic building information. The platform retrieves construction details, roof characteristics, wildfire exposure, and replacement-cost estimates.

Verification rules compare these records with the applicant's submitted information. Material differences trigger an underwriter referral with supporting evidence. Verified applications proceed directly to risk scoring.

This process creates a stronger assessment foundation while reducing repetitive research for underwriting teams.

3. Risk Scoring and Decisioning

Brief Overview

Risk scoring estimates the likelihood and potential severity of future insured losses. Decisioning converts that assessment into an approval, decline, conditional offer, or underwriter referral.

How Automation Supports the Process

Predictive risk scoring models analyze verified applicant data, historical losses, exposure variables, and behavioral indicators. Rules engines apply eligibility requirements, risk-appetite limits, and mandatory referral conditions.

The decision engine combines rule results with model outputs and confidence thresholds. Explainability tools identify the variables that influenced each score and recommended action.

Key Business Benefits

• Accelerates risk classification and submission prioritization
• Applies underwriting guidelines more consistently
• Improves alignment between risk appetite and decisions
• Supports explainable, traceable underwriting outcomes
• Routes complex submissions to appropriate specialists

Example or Use Case

A commercial auto insurer evaluates fleet size, driver history, vehicle usage, operating locations, and previous claims. The predictive model calculates a risk score using approved underwriting variables.

Low-risk submissions meeting every eligibility rule proceed toward automated pricing. Borderline cases enter an underwriter queue with risk drivers and supporting evidence. Prohibited risks receive a documented decline recommendation.

This approach combines analytical speed with controlled human judgment.

4. Pricing and Policy Decisions

Brief Overview

Pricing converts assessed risk into an appropriate premium, coverage structure, limit, deductible, and policy condition. Policy decisioning confirms whether the proposed terms satisfy underwriting and regulatory requirements.

How Automation Supports the Process

Rating engines apply approved pricing factors, actuarial tables, product rules, and jurisdictional requirements. Predictive models can refine risk segmentation within permitted governance boundaries.

Decision engines compare calculated terms against minimum premiums, coverage limits, authority thresholds, and profitability targets. The system records every pricing input, rule result, adjustment, and approval.

Key Business Benefits

• Produces faster and more consistent quotations
• Reduces manual rating calculations and errors
• Aligns pricing with assessed risk and appetite
• Maintains traceable decision and approval records
• Supports controlled product and market expansion

Example or Use Case

A small-business insurer receives a verified retail-store submission. The rating engine evaluates location, revenue, occupancy, prior losses, selected limits, and requested deductibles.

Eligible submissions receive calculated premiums and standardized coverage terms. Unusual exposures or pricing exceptions move to an authorized underwriter.

The workbench presents the calculated premium, applied rules, risk drivers, and required approvals. This process improves pricing consistency while preserving underwriter authority over commercial exceptions.

5. Underwriter Review and Referrals

Brief Overview

Underwriter review addresses submissions that exceed automated decision thresholds or require specialist judgment. Referrals may result from unusual exposures, missing evidence, pricing exceptions, or limited model confidence.

How Automation Supports the Process

Workflow engines route each exception according to product, complexity, location, authority level, and urgency. The underwriter receives a consolidated risk summary with supporting documents and verified data.

The workbench displays triggered rules, model outputs, confidence scores, pricing details, and referral reasons. Underwriters can approve, modify, decline, or request additional information.

Every action enters an auditable decision record. Completed reviews can also improve rules, workflows, and model-monitoring processes.

Key Business Benefits

• Prioritizes submissions requiring specialist attention
• Reduces time spent locating supporting information
• Improves referral consistency across underwriting teams
• Preserves human control over complex decisions
• Strengthens auditability and approval governance

Example or Use Case

A specialty insurer receives a submission involving multiple locations and unusual liability exposures. The platform identifies insufficient historical data and several appetite exceptions.

It routes the case to a senior underwriter with relevant evidence. The underwriter adjusts coverage terms and documents the decision before approval.

6. Policy Issuance and Workflow Completion

Brief Overview

Policy issuance converts an approved underwriting decision into a bound and active insurance contract. Workflow completion ensures that every document, approval, payment, and system record is finalized correctly.

How Automation Supports the Process

Workflow engines transfer approved terms into policy administration and billing systems. The platform generates quotations, schedules, endorsements, disclosures, and policy documents using validated data.

Automated checks confirm required approvals, signatures, payments, and regulatory notices. APIs distribute completed documents to brokers, policyholders, portals, and document repositories.

The system records timestamps, policy versions, decision histories, and outstanding requirements. Failed validations enter an exception queue before policy binding.

Key Business Benefits

• Reduces delays between approval and policy issuance
• Prevents repeated data entry across core systems
• Improves document accuracy and version control
• Strengthens regulatory and operational audit trails
• Provides faster confirmation to brokers and customers

Example or Use Case

A commercial property submission receives final approval with adjusted limits and deductibles. The platform transfers those terms into the policy administration system.

It generates the required documents and verifies payment completion. After validation, the policy binds automatically and reaches the broker through the connected portal.

Architecture of an Automated Underwriting System

Underwriting automation architecture separates data processing, decision logic, AI services, and user workflows. This modular structure improves scalability, governance, integration, and system maintenance.

i. Data and Integration Layer

This layer connects submission channels, core platforms, external databases, and document repositories. APIs, event streams, and data pipelines standardize information for downstream processing, often built on ACORD data standards.

ii. Rules and Decisioning Layer

Business rules enforce eligibility, appetite, pricing, referral, and authority requirements. Decision engines combine these rules with model outputs to approve, decline, or refer submissions.

iii. AI and Analytics Layer

This layer supports predictive scoring, computer vision, document intelligence, and generative assistance. Model registries, monitoring services, and explainability controls govern production performance.

iv. Underwriting Workflow and User Layer

Workflow orchestration assigns tasks, manages exceptions, and records approvals. Underwriter workbenches present risk summaries, evidence, recommendations, and decision histories.

Together, these four layers form the underwriting automation architecture that creates a controlled information flow. Data enters through validated integrations and moves into rules and AI services. Decision results then reach workflows, underwriters, and policy administration systems.

Loose coupling also allows insurers to upgrade individual capabilities without replacing the complete platform. This modular design operationalizes the rules-plus-ML hybrid underwriting model introduced earlier, letting business rules and machine learning models each do what they do best.

Building the 4-layer underwriting reference architecture on top of Guidewire PolicyCenter, Duck Creek Policy, or Sapiens?

AI Models Used in Insurance Underwriting

AI insurance underwriting uses several model types because underwriting involves different data and decisions. No single model can process documents, inspect properties, estimate losses, and assist underwriters effectively.

Four model categories support modern underwriting platforms:

Each model operates within defined rules, confidence thresholds, and approval controls. Model outputs should support underwriting decisions without replacing mandatory business logic or human authority.

Production deployments also require data-quality checks, version control, explainability, bias testing, and performance monitoring. These controls help insurers detect model drift and document how automated recommendations were produced.

The following sections explain each model's operation, business value, and practical underwriting application.

1. Predictive AI for Risk Scoring

Brief Overview

Predictive AI estimates future insurance outcomes using historical data and approved risk variables. Common outputs include loss probability, claim frequency, claim severity, and expected loss.

How It Works

This predictive risk scoring process analyzes patterns across policy, claims, exposure, customer, and external datasets. Algorithms assign scores or probability estimates to new submissions with comparable characteristics.

Rules engines interpret these outputs against underwriting appetite and referral thresholds. Explainability methods identify variables that materially influenced each prediction.

Business Value

• Improves risk segmentation across large submission volumes
• Supports consistent and evidence-based underwriting decisions
• Identifies high-risk or unusual applications earlier
• Strengthens submission prioritization and portfolio management
• Helps align pricing decisions with expected exposure

Example or Use Case

A commercial fleet insurer analyzes vehicle types, operating regions, driver histories, mileage, and previous claims. The predictive model estimates expected loss frequency and severity for each submission.

Low-risk fleets proceed toward automated pricing when every eligibility rule is satisfied. Higher-risk fleets enter specialist review with identified risk drivers and supporting data.

Underwriters retain final authority over exceptions, overrides, and commercially sensitive decisions. This control prevents predictive scores from becoming unsupported automatic conclusions.

2. Computer Vision for Risk Assessment

Brief Overview

Computer vision property underwriting analyzes images, videos, aerial imagery, and inspection records for visible risk characteristics. It converts visual evidence into structured attributes that underwriting systems can evaluate.

How It Works

Image models detect property conditions, construction features, roof characteristics, vehicle damage, equipment types, and safety hazards, drawing on property risk analytics platforms such as Zesty.ai and Cape Analytics. Geospatial imagery can also identify surrounding environmental exposures.

The system assigns confidence scores to every detected attribute. Low-confidence findings and material inconsistencies move to human inspectors or underwriters for verification.

Business Value

• Accelerates property and asset inspections
• Reduces dependence on manual image review
• Improves consistency across visual assessments
• Identifies hazards before pricing and policy issuance
• Supports more current risk information

Example or Use Case

A property insurer receives exterior photographs and aerial imagery for a building. Computer vision property underwriting identifies roof condition, building materials, vegetation proximity, and visible structural concerns.

The platform compares these attributes with submitted application data. Matching information continues into risk scoring, while significant differences generate a referral.

An underwriter reviews the original images, detected attributes, and confidence levels. This approach improves assessment speed while preserving verification for uncertain or material findings.

3. NLP and Document Intelligence

Brief Overview

Natural language processing converts unstructured insurance documents into classified, searchable, and decision-ready information. It supports submissions containing emails, forms, loss runs, reports, schedules, and policy records.

How It Works

NLP and intelligent document processing classify files before extracting relevant fields, clauses, and entities. Optical character recognition converts scanned pages and images into machine-readable text, using services such as AWS Textract, Azure Document Intelligence, or Google Cloud Document AI.

Validation services compare extracted data across documents and connected systems. Confidence thresholds determine whether information proceeds automatically or requires manual verification.

Business Value

• Reduces manual document review and data entry
• Accelerates submission intake and triage
• Detects missing or conflicting information earlier
• Standardizes data across different document formats
• Improves downstream scoring and decision accuracy

Example or Use Case

A commercial insurer receives a broker email containing application forms, loss runs, and location schedules. The platform separates each attachment and identifies its document type.

NLP submission ingestion extracts insured names, locations, coverage requests, and historical claims. Validation rules compare these details across every submitted file.

Complete records continue into enrichment and scoring workflows. Missing fields or low-confidence values enter a verification queue with direct references to their source documents.

4. Generative AI for Underwriter Assistance

Brief Overview

GenAI underwriter assistance summarizes submissions, retrieves evidence, and prepares review materials for underwriters. It assists decision-making but should not independently approve, decline, or price complex risks.

How It Works

A language model receives controlled access to approved underwriting data, documents, guidelines, and prior decisions. Retrieval systems ground responses in authorized sources rather than unrestricted model knowledge.

GenAI underwriter assistance can create risk summaries, compare documents, explain referral triggers, and draft broker questions. Guardrails restrict unsupported conclusions, sensitive data exposure, and unauthorized actions.

Business Value

• Reduces time spent reviewing lengthy submissions
• Provides faster access to relevant underwriting evidence
• Improves consistency across risk summaries
• Accelerates broker and internal communications
• Preserves underwriter attention for complex analysis

Example or Use Case

A specialty underwriter receives a submission containing policies, engineering reports, and historical claims. The assistant produces a structured summary with cited source references.

It highlights missing information and explains triggered underwriting rules. The underwriter verifies the evidence before making any decision.

Prompt logs, retrieved sources, outputs, and user actions remain available for monitoring and audits. This human-controlled design improves productivity without transferring underwriting authority to the model.

Business Benefits and ROI of Underwriting Automation

Underwriting automation ROI measures financial improvements produced by faster decisions, stronger risk selection, and reduced manual effort. For P&C underwriting AI initiatives, this calculation is especially important given high submission volumes and thin margins. Insurers should compare measurable outcomes against implementation, integration, governance, and operating costs.

i. Faster Underwriting Decisions

Automated intake, enrichment, and scoring reduce delays between submission receipt and quotation. Faster responses can improve broker satisfaction and conversion opportunities.

ii. Improved Underwriter Productivity

Automation removes repetitive document review, data entry, and routine verification. Underwriters can focus on complex risks, portfolio management, and broker relationships.

iii. Greater Pricing Consistency

Centralized rules and approved models apply pricing inputs uniformly, consistent with actuarial standards published by the Casualty Actuarial Society. This consistency reduces unsupported variations across teams, products, and distribution channels.

iv. Reduced Operational Costs

Straight-through workflows lower manual processing requirements for eligible submissions. Savings depend on automation coverage, exception volumes, infrastructure, and ongoing model-management costs.

v. Better Risk Selection

Predictive insights and enriched data help insurers identify attractive, borderline, and prohibited risks earlier. Stronger selection can improve portfolio quality and underwriting discipline.

Insurers should track quote turnaround time, referral rates, handling time, automation rates, override frequency, and loss performance. These metrics connect technical delivery with measurable business impact.

Straight-Through Underwriting

i. What It Is

Straight-through underwriting (STU) processes eligible submissions without manual intervention. The system collects data, validates requirements, scores risk, calculates pricing, and issues decisions automatically.

Where It Works Best

STU performs best with standardized products, reliable data, stable rules, and predictable risk profiles. Personal lines and lower-complexity commercial products often provide suitable starting points for commercial underwriting automation.

Complex, unusual, or high-value submissions require specialist assessment rather than complete automation.

ii. Business Benefits

• Shortens quotation and policy issuance timelines
• Reduces manual work for routine submissions
• Improves decision consistency across channels
• Expands processing capacity without proportional staffing
• Allows underwriters to focus on complex risks

iii. Human Oversight and Exceptions

Insurers should define confidence thresholds, referral triggers, and underwriting authority limits. Missing data, conflicting evidence, unusual exposures, and model uncertainty should stop automated processing.

Underwriters must review referred cases with supporting evidence and decision explanations. Teams should monitor automation rates, overrides, declines, pricing outcomes, and downstream loss performance.

Effective STU automates suitable risks while preserving human control over material exceptions. This balance supports operational speed without weakening accountability, governance, or underwriting discipline.

Challenges of Automated Underwriting

Automated underwriting can scale decisions, but weak foundations create operational and regulatory risks. Insurers must address five connected challenges before expanding production use.

i. Data Quality and Integration

Incomplete, outdated, or inconsistent data weakens rules and model outputs. Validation, lineage, reconciliation, and monitoring controls must operate across every connected source.

ii. Legacy Systems

Older policy platforms may lack modern APIs and real-time processing capabilities. Insurers often need integration layers, event-driven services, or phased application modernization.

iii. Model Bias and Explainability

Biased training data can produce unfair or discriminatory outcomes. Explainability, fairness testing, representative datasets, and documented review processes support responsible decisions.

iv. Compliance Requirements

Insurers must maintain governance throughout each AI system's lifecycle, as outlined in the NAIC Model Bulletin on the use of AI in insurance. Controls should cover development, validation, deployment, monitoring, documentation, and third-party models. Frameworks such as the NAIC AI Model Bulletin sit alongside state-specific rules, such as Colorado's SB21-169, New York's Circular Letter No. 7, and oversight from California regulators.

v. Human Oversight

Insurance regulators expect clear authority limits and escalation rules for automated decisions. Underwriters must review uncertain, complex, or commercially sensitive cases with adequate supporting evidence.

Successful implementation depends on more than model accuracy. It requires reliable architecture, governed data, measurable controls, and accountable human decision-making.

Insurers should address these requirements during system design. Adding governance after deployment increases remediation costs and limits operational confidence.

The Future of Insurance Underwriting Automation

Future underwriting platforms will combine real-time data, governed AI, workflow orchestration, and human expertise. The objective will shift from isolated task automation toward connected decision systems.

i. AI-Assisted Underwriting

AI assistants will summarize submissions, retrieve evidence, identify missing information, and recommend next actions. Underwriters will verify outputs and retain authority over material decisions.

ii. More Real-Time Risk Decisions

Connected devices, geospatial services, telematics, and external databases will provide updated exposure information. Event-driven architecture will process relevant changes without waiting for scheduled reviews.

iii. Increased Straight-Through Underwriting

Insurers will expand automated processing across suitable personal and commercial products. Growth will depend on reliable data, stable decision rules, and controlled exception management.

v. Human-AI Collaboration

Automation will handle repetitive analysis, while underwriters manage judgment-intensive risks and portfolio strategy. Workbenches will combine model explanations, documents, rules, and external evidence within one interface.

Future systems will also require continuous model monitoring, bias testing, audit trails, and access controls. These capabilities will support regulatory accountability as automation becomes more influential.

Insurers should therefore build modular platforms instead of adopting disconnected tools. Flexible architecture allows teams to introduce new models and data services without redesigning complete underwriting workflows.

Conclusion

Insurance underwriting automation in 2026 is becoming a connected decision capability rather than a standalone efficiency tool. It coordinates submission intake, enrichment, scoring, pricing, referrals, and policy issuance across one governed workflow.

The underlying architecture determines whether these capabilities can scale reliably. Data integrations establish trustworthy inputs. Rules engines enforce underwriting policies. AI models identify patterns and support assessment. Workflow systems preserve human authority and operational accountability.

Business value should remain measurable throughout implementation. Insurers must track decision speed, handling effort, referral volumes, overrides, pricing consistency, and portfolio outcomes. These measures reveal whether automation improves underwriting performance or merely moves existing inefficiencies into new systems.

Governance must also begin during system design. Explainability, fairness testing, access controls, audit trails, and model monitoring cannot remain post-deployment additions.

Insurers should begin with a clearly defined product, reliable data, and measurable decision workflow. They can then expand automation after validating operational, financial, and risk outcomes.

A phased approach creates stronger foundations for scalable, responsible, and commercially effective underwriting transformation.