>

AI Agents

How EY Can Transform Assurance and Tax Consulting with Agentic AI

StackAI

AI Agents for the Enterprise

StackAI

AI Agents for the Enterprise

How EY Can Transform Assurance and Tax Consulting with Agentic AI

Agentic AI in assurance and tax consulting is quickly moving from an interesting experiment to a serious delivery advantage. For firms like EY, the opportunity isn’t simply faster drafting or better chatbots. It’s a new way to run engagements: AI agents that can plan work, retrieve evidence, execute repeatable steps across systems, and package outputs into review-ready artifacts with guardrails.


That promise comes with real constraints. Assurance and tax work lives under stricter expectations than most corporate workflows: evidence must be traceable, judgments must be defensible, and independence must be protected. Done well, agentic AI can raise quality and consistency while reducing cycle time. Done poorly, it can create opaque workpapers, data leakage risk, and a reviewer burden that wipes out any efficiency gains.


This guide breaks down what agentic AI in assurance and tax consulting looks like in practice, where it fits in EY’s delivery model, the highest-impact use cases, the governance required for regulated work, and a realistic roadmap from pilot to scale.


What “Agentic AI” Means in Assurance and Tax (and Why It Matters)

Definition (plain English)

Agentic AI refers to AI systems that can plan, take actions, and coordinate tasks across tools and workflows within defined constraints. Instead of answering a single prompt, an agent can execute a multi-step process such as: gather documents, extract key fields, reconcile totals, flag exceptions, draft a workpaper section, and route it for approval.


In the context of agentic AI in assurance and tax consulting, the key word is coordinate. The agent isn’t just generating text; it’s orchestrating work across engagement artifacts and enterprise systems.


To make that clearer, here’s how agentic AI differs from two common predecessors:


RPA (traditional automation)

Rule-based and deterministic. Great for stable, structured processes like moving data between systems. Weak when inputs are messy, documents vary, or judgment is needed.


GenAI copilots

Prompt-driven assistants. Strong for summarizing, drafting, and explaining. Typically require constant user direction and don’t reliably execute end-to-end workflows.


Agentic workflows

Goal-driven execution across steps. Agents can call tools, retrieve information, apply rules, and propose outputs for review. The best implementations include explicit controls: permissions, approval gates, and audit trails.


The practical takeaway: agentic AI in assurance and tax consulting is less about autonomy and more about structured delegation. It takes repeatable parts of complex professional work and makes them faster, more consistent, and easier to review.


Why now (market and delivery pressures)

Several forces are converging:


  • Increasing volume and complexity

  • Margin pressure and talent constraints

  • Client expectations for near-real-time insights

  • Higher scrutiny on defensibility


Agentic AI fits this moment because it can compress the time between “request” and “review-ready artifact” while strengthening structure and consistency—if it’s deployed with the right safeguards.


Where Agentic AI Fits in EY’s Assurance & Tax Delivery Model

The “agent orchestration” concept for professional services

A practical way to think about agentic AI in assurance and tax consulting is a team model:


Orchestrator agent

Coordinates the workflow, tracks engagement status, routes tasks, and enforces policies (what can be done, by whom, and with what approvals).


Specialist agents

Execute focused tasks such as:


  • Evidence agent: collects and validates PBC support

  • Testing agent: proposes test steps, prepares samples, documents results

  • Research agent: retrieves authoritative guidance and summarizes it

  • Drafting agent: produces structured workpaper content with traceability

  • Reconciliation agent: performs tie-outs and variance explanations

  • Packaging agent: assembles binder-ready deliverables with versioning


This mirrors how high-performing teams already operate. The difference is that agents can handle repetitive steps across tools: document management systems, ERPs, GRC platforms, workpaper tools, ticketing, and e-signature.


A key principle for EY-scale delivery is avoiding monolithic “do everything” agents. The most reliable programs break work into targeted, reusable workflows with clear inputs and outputs. That structure surfaces feasibility constraints early: messy data sources, integration needs, and compliance requirements.


Guardrails needed for regulated work

Agentic AI in assurance and tax consulting must be designed for review, not just output. Guardrails make the difference between useful automation and an unmanageable risk.


Core guardrails typically include:


  • Human-in-the-loop approvals

  • Permissioning and least privilege

  • Immutable audit trail

  • Evidence provenance

  • Separation of duties and independence considerations


In short: agentic AI in assurance and tax consulting should operate like a controlled workflow engine, not a free-form assistant.


A practical maturity model (crawl → walk → run)

A realistic maturity model helps teams scale without overreaching.


Tier 1: Assistive drafting and summarization (low autonomy)

Examples: summarize contracts, draft walkthrough narratives, create meeting minutes, generate first-pass memos.


Risk profile: lower, but still requires privacy controls and reviewer guidance.


Tier 2: Semi-autonomous workflow execution with approvals (moderate autonomy)

Examples: PBC tracking, evidence extraction, tie-outs, sample selection suggestions, exception write-ups routed for review.


Risk profile: medium, because actions span tools and outputs feed regulated workpapers.


Tier 3: Continuous assurance and continuous tax monitoring (high governance)

Examples: near-real-time control monitoring alerts, continuous transaction triage, ongoing tax exposure monitoring.


Risk profile: higher, because the system affects timing, coverage, and potentially client decisions.


The best programs treat Tier 2 as the main event. That’s where agentic AI in assurance and tax consulting delivers measurable time savings without betting the engagement on full autonomy.


High-Impact Agentic AI Use Cases in Assurance (Audit & Controls)

Assurance is a natural fit because audit work combines repeatable procedures with heavy documentation and traceability demands. Agentic AI can compress cycle time while improving consistency in engagement artifacts.


Evidence collection and PBC acceleration

One of the most immediate wins for agentic AI in assurance and tax consulting is reducing PBC friction.


What an evidence agent can do:


  • Generate and tailor PBC lists

  • Send requests and track status

  • Validate completeness and consistency

  • Extract key data


The impact isn’t just time saved. It also reduces late-stage scramble, improves documentation discipline, and allows teams to identify issues earlier.


Risk assessment and planning support

Planning is a high-leverage stage where better organization pays off downstream. Here, agentic AI should support the team, not replace judgment.


Planning agents can:


  • Review prior-year workpapers and findings

  • Scan meeting minutes, contracts, and key documents

  • Draft risk narratives with rationale


When done right, this improves consistency across engagements and reduces the time managers spend hunting for context.


Controls testing and SOX automation

SOX and controls testing involves repeatable steps, structured documentation, and frequent evidence handling—ideal for agentic workflows.


Examples include:


  • Draft walkthrough documentation

  • Propose test steps aligned to control objectives

  • Prepare sample lists and testing packets

  • Exception workflow support


This is where SOX controls testing automation becomes less about speed and more about standardization and defensibility.


Substantive testing augmentation

Substantive testing includes recalculations, tie-outs, variance analysis, and transaction testing—often time-consuming but partly mechanical.


Agentic workflows can:


  • Automate recalculations and tie-outs

  • Generate variance explanations for review

  • Triage transaction testing with anomaly signals


The key is to keep humans in control of conclusions while letting the agent handle repetitive computation and packaging.


Workpaper drafting with traceability

Drafting doesn’t have to mean “black box text.” For agentic AI in assurance and tax consulting, the goal is structured drafting that helps reviewers.


A drafting agent can:


  • Create standardized workpaper sections

  • Embed source references

  • Maintain versioning and reviewer checklists


A practical litmus test: if a reviewer can’t quickly trace the output back to evidence, the drafting workflow isn’t ready for regulated work.


High-Impact Agentic AI Use Cases in Tax Consulting

Tax work spans compliance, advisory, and controversy support. Across these areas, the opportunity for tax compliance automation and AI in tax advisory is especially strong when workflows are data-heavy and documentation-intensive.


Tax compliance workflow orchestration

Compliance is full of repeated steps: collecting inputs, reconciling, drafting workpapers, and packaging for filing readiness.


An agentic compliance workflow can:


  • Collect data from ERP and prior filings

  • Flag missing inputs and unusual movements

  • Generate first-pass workpapers

  • Route questions and tasks


This doesn’t eliminate the need for experienced review. It reduces the mechanical burden so teams focus on judgment and client advisory.


Tax research and position support

Tax research is one of the most valuable and risky areas for AI in tax advisory. It’s valuable because it can accelerate time-to-insight. It’s risky because sources must be authoritative, current, and appropriately licensed.


A research agent can:


  • Retrieve and summarize guidance

  • Draft memos with structured sections

  • Enforce jurisdiction and date sensitivity

  • Support reviewer validation


The most successful teams treat research agents as “accelerated first pass” tools, not final-answer engines.


Global mobility and cross-border coordination

Cross-border work adds coordination complexity: multiple countries, multiple deadlines, and inconsistent data formats.


Agentic workflows can help by:


  • Managing task routing and checklists

  • Explaining assumptions and documenting them

  • Reducing miscommunication


This is less about clever AI and more about disciplined orchestration across distributed teams.


Indirect tax (VAT/GST/sales tax) reviews

Indirect tax often requires matching products, rates, jurisdictions, and transaction data—at scale.


Agents can:


  • Detect rate or mapping issues

  • Prepare exception cases

  • Support documentation for audits


In this area, agentic AI helps teams focus on resolution and strategy instead of data wrangling.


Controversy and audit defense readiness

When controversy hits, speed and documentation discipline matter. Agentic AI in assurance and tax consulting can improve audit defense readiness long before a dispute occurs.


A controversy-support workflow can:


  • Build an audit-ready evidence binder

  • Track communications and deadlines

  • Identify missing support proactively


The result is a calmer, more defensible response posture when scrutiny increases.


Architecture: How EY Could Implement Agentic AI Safely (Reference Blueprint)

The architecture for agentic AI in assurance and tax consulting should prioritize security, traceability, and controllable workflows. The goal is a system that can execute multi-step work while proving what it did and why.


Core components

A practical reference blueprint includes:


  • Secure model layer

  • Retrieval layer (RAG)

  • Tool connectors

  • Workflow engine

  • Data boundaries and security controls


Key controls include:


  • Client data isolation

  • Encryption and key management

  • Logging and monitoring

  • Leakage prevention

  • PII handling


In professional services, trust is built as much on controls as on capability.


Agent control mechanisms

Agentic workflows should never be “open-ended.” The system needs explicit control mechanisms.


Common mechanisms include:


  • Policy engine

  • Approval gates

  • Rate limits and kill switch

  • Anomaly monitoring


This is where agentic AI becomes enterprise-ready: not because it’s smarter, but because it’s governable.


Model risk management and validation

Model risk management (MRM) matters in this context because errors can lead to wrong conclusions, flawed documentation, or compliance issues.


A practical approach includes:


  • Task-based accuracy thresholds

  • Benchmarking and test suites

  • Red-teaming

  • Ongoing monitoring


The objective is not perfection; it’s controlled reliability with transparent limitations.


Governance, Quality, and Independence: What Must Be True for EY

Technology alone won’t make agentic AI in assurance and tax consulting successful. Governance determines whether teams trust the outputs, reviewers can defend them, and leadership can scale adoption.


Quality management and documentation defensibility

Defensibility comes from reproducibility. Outputs should be explainable and tied back to evidence.


Strong programs implement:


  • Evidence traceability

  • Standardized workflows as controlled artifacts

  • Reviewer guidance


This reduces reviewer burden and keeps quality consistent across teams.


Regulatory and ethical considerations (high-level, non-legal advice)

Agentic AI in assurance and tax consulting operates under professional standards and client confidentiality expectations. Governance should address:


  • Documentation expectations

  • Privacy principles

  • IP and licensing


A program that scales without these considerations will eventually hit a trust wall.


Independence and conflict management

Independence is both a policy and system design issue.


Practical safeguards include:


  • Engagement boundary enforcement

  • Role-based access and approvals

  • Environmental separation


Independence shouldn’t be left to “best efforts.” It should be embedded in the workflow.


Change management and talent enablement

The fastest way to stall adoption is to deploy tools without changing how teams work. Agentic AI introduces new responsibilities.


New or expanded roles often include:


  • Agent workflow designer

  • AI QA reviewer

  • Model risk lead


Training should focus on a practical skill: reviewing AI-assisted work. Engagement teams need to know what’s safe to rely on, what requires re-performance, and how to document review.


KPIs and Value Realization: How EY Should Measure Success

To prove value, agentic AI in assurance and tax consulting needs measurable outcomes. The best KPIs track efficiency, quality, and risk together.


Efficiency metrics (time, cycle time, throughput)

Useful metrics include:


  • PBC cycle time reduction

  • Workpaper drafting time

  • Testing coverage per hour


Cycle-time metrics are often the easiest to measure and resonate with both engagement leaders and clients.


Quality and risk metrics

Efficiency without quality creates downstream costs. Track:


  • Rework rate

  • Review notes per section

  • Exception detection rate

  • Documentation completeness score

  • Audit trail integrity


These metrics ensure assurance automation doesn’t turn into “faster chaos.”


Client outcomes

Ultimately, value should show up in client experience:


  • Faster close and filing readiness

  • Better visibility into risks and exposures

  • Improved audit readiness


These outcomes help justify investment beyond internal productivity.


A simple ROI model (example framework)

A practical ROI model can be built around:


  1. Baseline hours by workflow

  2. Assisted hours after agent deployment

  3. Redeployed capacity

  4. Program costs


The most credible business case for agentic AI in assurance and tax consulting combines time saved with risk reduction and better engagement quality.


Implementation Roadmap for EY (90 Days to 12 Months)

Scaling agentic AI requires sequencing. The most successful teams start with a small set of high-volume workflows, prove value, then expand with governance.


Phase 1 (0–90 days): Select workflows and set guardrails

Focus on 1–2 workflows that are high-volume and low-judgment, such as:


  • PBC tracking and evidence completeness checks

  • Reconciliations and tie-outs

  • First-pass drafting for standardized workpapers


In this phase:


  • Define policies and roles

  • Set approval gates

  • Stand up a secure sandbox

  • Build a measurable baseline


This phase is about proving that agentic AI in assurance and tax consulting can work safely, not about building the most advanced agent.


Phase 2 (3–6 months): Expand to multi-step agents

Once Tier 1 and early Tier 2 workflows are stable:


  • Integrate with workpaper and document systems

  • Create reusable playbooks

  • Establish monitoring and QA routines


This is where productivity gains become repeatable instead of one-off.


Phase 3 (6–12 months): Scale and standardize

To scale across practices:


  • Build a center of excellence

  • Enable cross-engagement analytics within policy

  • Pilot continuous monitoring carefully


Scaling isn’t about deploying more models. It’s about deploying repeatable workflows with predictable risk.


Common pitfalls and how to avoid them

  • Over-automation without controls

  • Shadow AI tools

  • Poor data hygiene

  • No reviewer training or accountability


Avoiding these pitfalls is often the difference between a pilot that looks good and a program that scales.


Conclusion: The Practical Future of Agentic AI at EY

Agentic AI in assurance and tax consulting has a clear path to real impact: faster evidence handling, more consistent documentation, stronger traceability, and better orchestration across tools and teams. For EY, the opportunity is an operating model upgrade—moving from isolated copilots to governed workflows that produce review-ready deliverables.


The teams that win won’t be the ones with the most autonomy. They’ll be the ones with the most disciplined controls: human-in-the-loop approvals, least-privilege access, evidence provenance, immutable audit trails, and strong model risk management.


The most practical next steps are straightforward:


  • Run a 2–4 week pilot on 1–2 high-volume workflows with measurable KPIs

  • Build a one-page control checklist for agent behavior, data boundaries, and approvals

  • Identify the top 10 workflows by volume and risk tier, then sequence rollout accordingly


To see how enterprise-grade agentic workflows can be built with strong governance and security controls, book a StackAI demo: https://www.stack-ai.com/demo

StackAI

AI Agents for the Enterprise


Table of Contents

Make your organization smarter with AI.

Deploy custom AI Assistants, Chatbots, and Workflow Automations to make your company 10x more efficient.