AI tax automation on Azure that actually ships

Most AI tax stories are still slideware. Wolters Kluwer has quietly turned one of them into a working, audited product that is already reading W‑2s, 1099s and K‑1s, then pushing clean data into tax returns with human review on the edge cases.

For Microsoft-centric tax and finance leaders, this is more than a press release. It is a concrete, Azure-based blueprint you can adapt for your own tax, accounts payable, onboarding or KYC workflows using Power Platform, Azure AI and tools like n8n.

In this deep dive we will unpack what CCH Axcess Expert AI and CCH Axcess Scan are actually doing, then map the same pattern to a practical architecture you can implement in your own estate.

From AI hype to near zero-touch tax returns: why this case study matters

Most firms have already dabbled with AI in tax and accounting. A Copilot here, a chatbot there, perhaps a proof of concept that summarises PDFs. The seasonal bottleneck of tax document intake, though, usually still looks depressingly manual.

Wolters Kluwer is tackling that bottleneck head on inside CCH Axcess. Using what it calls Expert AI, the platform now:

Ingests client documents such as W‑2s, 1099s and K‑1s
Classifies them by form and type
Extracts key fields
Writes the resulting structured data into CCH Axcess Tax for preparers to review

All of this is built on Azure AI services, with professionals still reviewing low confidence items and signing off returns. The press coverage talks openly about a "zero touch tax return" experience for straightforward cases, with humans firmly in control of judgment calls.

For leaders in the UK or multi-region groups, you can read "W‑2s, 1099s and K‑1s" as shorthand for your own mix of P60s, P45s, dividend vouchers, self assessment pages and partnership schedules. The shape of the problem is the same: high volume, repeatable documents where 80 to 90 per cent of the work is extraction and checking.

Why this matters:

Agentic AI on Azure is already handling real tax workloads
The pattern is reusable outside of US tax
Governance, auditability and human judgement have been designed in from the start

What Wolters Kluwer announced at Microsoft Ignite 2025

At Microsoft Ignite 2025, Wolters Kluwer appeared on a session about AI agent architectures, pitfalls and real business impact. The centrepiece was CCH Axcess Scan, powered by Azure AI and Expert AI.

Key public points from the announcement and related releases:

CCH Axcess Scan automates ingestion and processing of client tax documents, including complex forms such as K‑1s, W‑2s and 1099s, pushing structured data into CCH Axcess Tax for review-ready returns.
Expert AI is not a bolt-on widget. It is embedded into the cloud-native CCH Axcess platform as a "digital core" that connects tax, audit, workflow, practice management and client collaboration.
The same Expert AI capabilities are being rolled out across modules such as CCH Axcess Audit, Client Collaboration, Intelligence, Workflow and Advisor.
Everything is explicitly positioned as human-in-the-loop, with explainable, auditable outputs and clear governance.

From an architectural point of view, this is a fully agentic AI pattern implemented on Azure, not just some prompts wired to a chat UI.

Expert AI and CCH Axcess in plain English: built in, not bolt-on AI

It is worth dwelling on the "built in, not bolted on" language, because it is exactly where many internal AI initiatives go wrong.

In simple terms, CCH Axcess Expert AI is:

A layer of AI capabilities (classification, extraction, generation, routing) that sits inside the platform, not next to it.
Fed by a unified data model covering tax, audit, workflow and firm management.
Exposed in specific contexts such as "scan this document into a return", "summarise this contract for audit assertions" or "predict which documents to request from this client".

Contrast that with a typical bolt-on pattern in many firms:

Documents sit in disconnected file shares or ad hoc SharePoint libraries.
A generic "AI bot" or third party tool reads from those locations with limited context.
Results arrive in an email, Excel file or another disconnected interface.
Someone then manually rekeys or reconciles the outputs into the core system.

Wolters Kluwer has instead let the platform own the workflow. AI shows up as a native feature inside that flow. For your own organisation, that is the core lesson: the most effective AI tax automation will live inside, or be tightly coupled to, your system of record and your workflow engine.

How CCH Axcess Scan works: from ingestion to review-ready tax returns

Wolters Kluwer has not published every internal implementation detail, but combining their solution overview with Azure AI patterns, the high level CCH Axcess Scan workflow looks like this:

Document intake: Clients upload tax documents (W‑2s, 1099s, K‑1s and others) via portals such as CCH Axcess Client Collaboration or through firm-side scanning workflows. Files land in a secure cloud repository.
Classification and extraction: Azure AI services, combined with Wolters Kluwer's own models and templates, classify each page or document type, then extract key fields such as payer, recipient, amounts, tax withheld and tax year.
Validation and enrichment: The extracted data is validated for missing values, format issues and basic consistency. For example, totals and subtotals can be cross-checked and identifiers such as National Insurance numbers validated.
Data posting into tax software: Cleaned and validated data is then posted directly into CCH Axcess Tax, mapped to the correct schedules and fields.
Human review of exceptions: Anything low confidence is routed to preparers in a review interface that shows the source image side by side with extracted structured data and the AI's confidence scores.
Final sign-off: A tax professional reviews the overall return, addresses flagged items and signs off. The AI's role is documented with an audit trail of what was suggested and what was accepted or overridden.

For many firms, this sequence is very close to the intake and prep process you already operate, just with the bulk typing work removed.

Under the hood: Azure AI services, unified data core and agentic workflows

Although Wolters Kluwer understandably does not publish code, their public materials give us enough to infer a reference architecture that you can mirror.

At a high level, the moving parts look like this:

Cloud-native digital core: CCH Axcess is a unified, cloud-native platform that holds tax, audit, workflow, document management and client data in a single architecture. This "digital core" is what Expert AI plugs into.
Azure AI services: CCH Axcess Scan uses Azure AI services for document understanding and other AI tasks. Other Expert AI modules also draw on large language models and generative capabilities for research, summarisation and advisory support.
Agentic orchestration: Expert AI is described as "agentic ready". In practice, that means AI agents can act across workflows, reading documents, updating workpapers, triggering tasks and surfacing insights, all while being constrained by business rules and permissions.
Open APIs: The platform exposes APIs so that other systems, including potentially Microsoft 365 workflows, cloud ERPs or low-code apps, can interact with the same core.

If you are a Microsoft 365 or Azure shop, you already have access to most of these primitives. The gap is usually an application layer that gives you a unified workflow and data model to orchestrate them. That is where Power Platform, custom apps or tools like n8n come in.

Human-in-the-loop by design

A defining feature of Wolters Kluwer's approach is that human judgment is not an afterthought. Across the Expert AI materials and press releases, several recurring design elements show up: explainable outputs, audit trails, and deliberate semi-autonomy.

In practical terms, your own workflows should mirror this with patterns such as:

Confidence thresholding: If a document extraction or classification is above a defined confidence threshold, it can auto post. Everything else should be queued for human review.
Exception queues: Build queues for low confidence or high risk items, with clear prioritisation rules, ownership and ageing.
Side by side review UIs: Use Power Apps or internal tools to show source documents next to extracted data. Let reviewers accept, amend or reject suggestions explicitly.
Structured sign-off: Require a named professional to sign off batches, with clear segregation of duties for sensitive areas such as tax positions.

These controls matter just as much for UK tax, VAT and FCA regulated processes as they do in US tax workflows.

Governance and trust: audit trails, explainability and regulatory comfort

For regulated functions such as tax, audit and KYC, governance is not optional. Wolters Kluwer foregrounds this by emphasising audit trails, privacy by design, and expert-curated content.

When you adapt this pattern into your own Azure estate, consider governance elements as first class requirements:

Data residency and sovereignty: Understand where your Azure AI instances and storage accounts reside. Make sure this aligns with HMRC, IRS, FCA or other regulatory expectations.
Model input and output logging: Log prompts, documents and AI outputs in a way that is retrievable for internal audit. Keep this secure and access controlled.
Policy-based access: Use Azure AD, conditional access and role based access control to ensure that only appropriate users can see sensitive documents.
Documented policies: Capture how and where AI is used in your financial processes. Regulators and clients both increasingly expect this clarity.

Beyond tax: applying the same pattern

The Expert AI pattern is not unique to tax documents. Once you have AI-driven document classification, extraction and workflow routing in place, the same building blocks work across several other processes.

Zero touch accounts payable: Supplier invoices arrive via email or portal, are classified and parsed by an Azure AI model, then matched against purchase orders. Exceptions appear in an AP queue for analysts to resolve.
AI assisted client onboarding: For accounting and audit engagements, an AI model can generate prepared-by-client lists based on industry and entity type, then track uploaded documents against that list.
KYC and regulatory routing: Identity documents, proof of address, consent forms and disclosures can be classified, key data points extracted and expiry dates checked. AI can then route these to the correct customer record.

Mapping Wolters Kluwer's architecture to your stack

You do not need CCH Axcess to adopt this pattern. A Microsoft-centric SME or mid market group can assemble a similar architecture using components you likely already own, plus some glue.

Component	Your Stack Implementation
Document Repository	SharePoint, OneDrive for Business, or Azure Blob Storage.
AI Document Analysis	Azure AI Document Intelligence (standard or custom models).
Language Layer	Azure OpenAI or Microsoft Copilot Studio.
Workflow Engine	Power Automate or n8n.
Review Interface	Power Apps, Dataverse model-driven apps, or lightweight web apps.

The goal is not to replicate CCH Axcess feature for feature. It is to reuse the same architectural ideas: unified data, embedded AI, governed workflows and human control.

Implementation blueprint: stand up a pilot in 4 to 8 weeks

With a focused scope, it is realistic to stand up a first pilot of AI document automation in a one or two month window. A pragmatic sequence:

Pick a narrow, high volume use case: E.g., P60/P45 intake or invoice processing for a single unit.
Define "good enough" success metrics: Aim for auditability and 70% automation, not perfection.
Build the document pipeline: Trigger on file upload -> Classify/Extract -> Write to DB.
Create a review UI: A simple interface for staff corrections.
Lock in governance: Define approvals and audit protocols.
Iterate on model quality: Use human corrections to retrain models.

Common pitfalls in AI agent architectures

⚠️ Watch out for:

Disconnected "AI labs": Experiments that don't integrate with systems of record won't scale.
Over generic prompts: Without context or templates, results will be inconsistent.
Insufficient governance: Never let AI write to core systems without validation.
Ignoring users' daily tools: If the review UI is awkward, adoption will fail.

Closing the loop: measuring impact and scaling firm wide

Once an AI document pipeline is live, you should measure it like any other operational capability. Track hours saved, error rates (AI vs Manual), cycle times, and user satisfaction.

As you prove value in one workflow, you can extend to adjacent ones: other tax forms, other business units, or new document types such as contracts and board minutes for audit. At each step, maintain the same discipline around governance and human oversight.

Need help designing your architecture?

If you would like a sounding board or a partner as you design or scale these patterns, nocodecreative.io's consultants regularly help firms stitch together Azure AI, Power Platform and tools like n8n into governed, production ready automation.

💡

Contact Our Expert AI Consultants