AI agents shake up accounting firms' bookkeeping workflows

AI agents are starting to automate bookkeeping and tax work at accounting firms, prompting firms to rethink workflows, oversight and the role of human staff.

A new category of software is emerging in professional services through tools such as OpenClaw and Claude Cowork. Unlike conventional software that waits for step-by-step instructions, these systems can carry out defined tasks with some autonomy across documents, files and applications.

In accounting, that matters because much of the work remains repetitive and rule-based. Transaction categorisation, reconciliations, and document processing have traditionally consumed a large amount of junior staff time, making them clear targets for automation.

Two models

OpenClaw reflects an open-source approach. It runs locally on a user's machine, connects with messaging platforms and lets users issue commands in natural language, often through a chat interface. The software then executes tasks across files and applications, emphasising flexibility and user control.

Claude Cowork takes a more managed approach. It operates within a desktop application environment, where users point it to folders or documents and assign a task. The agent reads the material, processes it and produces an output with limited manual intervention, relying on proprietary models and built-in governance features.

This distinction reflects a broader divide in business software. Open systems tend to favour adaptability and control for technically confident users, while managed systems are built around standardisation and oversight.

Core tasks

Both tools target the same operational problem. Accounting firms handle large volumes of structured work that can be described through rules, templates or repeatable steps.

OpenClaw has shown speed in handling structured workloads, including processing hundreds of transactions within minutes and populating tax workpapers from multi-page documents. It appears most effective when tasks can be performed via scripts or application programming interfaces.

Claude Cowork is more focused on document-led processes. Trials have shown it can extract data from PDFs, organise files, assemble draft reports, and process thousands of transactions into structured outputs within minutes.

Even so, the output is not yet reliable enough to eliminate the need for human review. Results are often compared with junior-level work: strong on routine items but still weak in edge cases and areas requiring professional judgment.

That is changing how time is spent inside firms. Work that once required hours of manual handling can now be completed much faster, leaving accountants to review, validate and correct what the systems produce.

Changing workflows

Adopting AI agents is not just a software purchase. Firms must redesign processes, define where human checks belong and decide which tasks can safely move from staff to machines.

OpenClaw's modular skills system lets users build customised workflows for tasks such as bookkeeping and tax preparation. That may appeal to smaller practices or technically capable teams that want to shape their own templates and routines.

Claude Cowork instead relies on predefined workflows and integrations for common use cases. That reduces setup, but also limits the level of customisation available to users.

The result is a different adoption path for different types of firms. Smaller practices and early adopters may favour the flexibility of open-source software, while larger organisations may be more comfortable with standardised systems that align with established control frameworks.

Early adoption

The use of AI agents in accounting is still in its early stages, but interest is spreading. Individual practitioners and smaller firms have led much of the experimentation, with shared demonstrations showing the systems handling bookkeeping and tax preparation tasks at scale.

Larger firms have moved more cautiously, using pilot programmes in controlled environments and limited use cases. These tests examine how agents interact with existing systems and whether their outputs can be supervised to meet professional and regulatory requirements.

Client expectations are adding pressure as well. As corporate customers become more aware of AI tools, they are increasingly likely to expect some of the efficiency gains to show up in service delivery.

Training is therefore becoming part of the transition. Accountants are learning to define tasks clearly for AI systems and to review outputs effectively, shifting the required skill set from execution to supervision.

Risk and control

The same features that make AI agents useful also create new operational risks. Tools that can read, write and modify files across systems raise concerns about data integrity, confidentiality and unintended actions.

OpenClaw's local execution model keeps data on the user's machine, reducing exposure to outside systems but placing more weight on internal security practices. Its open-source structure may also create risk if firms rely on unverified components.

Claude Cowork's use of cloud-based models introduces separate questions about data transfer and storage. Firms, therefore, need to assess how sensitive information is handled beyond their own environment.

Across both models, prompt injection remains a concern. Hidden instructions within documents or other inputs can alter an agent's behaviour, creating risks of data leakage, faulty processing or unauthorised actions.

To manage this, firms are adopting safeguards such as restricted permissions, audit logs, sandbox environments and human approval for important actions. Working on copied data rather than live systems is also becoming standard practice.

Governance is developing alongside the technology. In effect, firms are treating AI agents much like junior staff: every output must be checked, and responsibility remains with qualified professionals.

The broader impact may be economic as much as technical. As routine processing becomes less labour-intensive, firms can redirect staff time towards analysis, advisory work and client interaction, while new specialist roles emerge to manage AI systems, design workflows, and oversee results.

For now, full integration into core accounting work remains gradual as firms continue to test reliability, control, and fit. Human oversight remains central because current systems still lack the judgment needed for complex decisions and compliance.