The Death of Accounting Software

How dynamic workflows will reshape accounting solutions

For more than three decades, larger scale accounting software (i.e. ERP solutions) has been developed around a stable premise: the process is designed first, encoded in the application and then adopted by the user. The steps to customize the menus, approval paths, reports and exception rules differ subtly from one product to another, but the underlying model is consistent. The consultant configures the client workflows in the software, as best they can, and this establishes how users interact with software, while the organization modifies their processes to work with the software.

For most small and mid-sized businesses, improving operational efficiency has traditionally meant selecting software with the closest approximation of their desired workflow and then layering additional automation tools around it. The objective has never been to create a workflow unique to the organization, but rather to configure and extend a predefined one. This architectural constraint has shaped enterprise software for decades and has become so commonplace that many organizations no longer question it.

This model has been effective because accounting requires consistency: a general ledger must remain complete; transactions must be traceable; access must be controlled; and review must be documented. Standardized applications brought discipline to processes that had previously depended on paper files, individual memory and local practice. These applications also made it possible for organizations and accounting firms to serve more clients without rebuilding their systems for every engagement.

The next evolution of these accounting solutions continues with this approach as they are unlikely to discard the foundational structure. Ledgers, controls, audit trails and reporting standards will remain essential. Rather than redesigning these applications from first principles, most software vendors are extending them by layering AI onto their existing architecture. AI assistants can answer questions, categorize transactions, summarize reports, draft communications and automate selected tasks, but they generally operate within workflows that were designed long before modern AI capabilities emerged. The underlying process remains largely unchanged. The limitation, however, is that AI remains subordinate to the workflow rather than defining it. It can accelerate individual tasks, provide recommendations and reduce manual effort, but it generally cannot reshape the process based on the circumstances of a particular client, engagement or organization. The workflow itself remains a product of software design rather than organizational knowledge.

AI is beginning to make it possible for workflow to be assembled and adjusted around the circumstances of the work, rather than being fully prescribed in advance. This is where a fundamentally different architectural model begins to emerge. Rather than treating AI as another feature embedded within predefined workflows, dynamic workflow environments place AI at the orchestration layer. Instead of asking, “What is the next step in this process?” the system asks, “Given everything I know about this client, this organization, previous decisions, current priorities and available information, what should the next step be?” The workflow is no longer static; it evolves continuously as the system learns from human decisions, changing business conditions and accumulating organizational knowledge.

That distinction is important as AI is currently treated as an additional feature within existing applications. A user can ask a question, summarize a report, draft a client message or receive a suggested account classification. These tools can be useful, but they leave the operating model largely unchanged and manually configured workflows by the implementation continue to take 2 to 6 months. The user still decides what must happen next, moves information between applications and steps, follows up on missing evidence and determines when an exception requires professional attention.

A more consequential use of AI is emerging in systems that can interpret an objective, identify the steps required, use available tools, monitor the result and revise the sequence when circumstances change. The Associated Press has described the central difference between a conventional chatbot and an AI agent as the movement from producing an answer to taking action across a multistep task. A 2025 MIT Sloan Management Review and Boston Consulting Group study, cited by the AP, characterized these systems as capable of planning, acting and learning with greater independence than prompt-and-response tools.[1] The terminology remains unsettled, and some products described as agentic provide only limited autonomy. The underlying direction, however, is clear: software is being developed to participate in the execution of work rather than only presenting information to the person executing it.

Microsoft has advanced a similar position. Reporting on Satya Nadella’s recent comments, the Financial Times noted his warning that organizations must develop and retain their own learning capabilities rather than allow the value created from their knowledge to accumulate entirely within external AI platforms.[2] His observations regarding learning systems suggest that enterprise advantage will depend on the interaction between human expertise, organizational data and systems that can learn from repeated use. In that model, the most valuable capability is not access to a general-purpose model. It is the ability to turn an organization’s experience into a continually improving and evolving operational system.

This view is consistent with established research on process management. Thomas H. Davenport and Thomas C. Redman argued in Harvard Business Review that AI and process management should be developed together. Their position matters because isolated AI tools often improve individual activities without improving the end-to-end process in which those activities sit.[3] A faster classification, summary or response does not necessarily improve a workflow if information remains incomplete, responsibilities remain unclear or exceptions continue to circulate between people and systems. Process performance depends on the design of the whole sequence, including the points at which judgment, verification and escalation occur.

This is where accounting provides a useful test. Accounting work is commonly described as a sequence of technical tasks: import transactions, classify them, obtain documents, reconcile accounts, prepare entries, review the file and issue a report. In practice, the difficulty lies in the context connecting those tasks. A transaction description may be incomplete, or the same vendor may represent different expenses for different clients, or a receipt may explain only part of a payment, or a recurring balance may be normal in one business and a warning sign in another. The accounting treatment may depend on information obtained in a meeting, an agreement stored elsewhere or a decision made during the prior year.

Experienced accounting professionals manage this complexity by drawing on knowledge that is distributed across the firm. They remember how the client operates, which documents are usually late, which accounts are sensitive, which estimates require partner involvement and which explanations have previously been accepted. This knowledge determines the real workflow, even when the software displays the same screens for every client whether we are talking about a large business or the local business around the corner. The formal process may be standardized, but the workflows are continuously adapted by the people performing them.

Dynamic workflows in accounting

An autonomous AI accounting solution would make more of that operating knowledge available to the system. It would not remove the ledger or weaken the control framework. It would use the ledger, documents, communications, prior decisions and assigned responsibilities to determine how each matter should proceed. The workflow would remain governed, but it would no longer need to be identical in every circumstance.

Consider a monthly bookkeeping and close process. In a traditional application, imported transactions enter a queue and are processed according to a standard set of steps. Rules may automatically classify known items, and machine learning may suggest classifications for the remainder. The user is still responsible for identifying missing information, contacting the client, deciding whether the answer is sufficient, completing the reconciliation and escalating unusual matters.

In an autonomous AI accounting solution with dynamic workflows, the system begins with the objective and the control requirements. It determines which information is already available, which transactions can be resolved from approved rules or prior evidence and which items require additional context. It may request a receipt from the designated contact, compare the response with the transaction and the client’s historical treatment, and route the item to a reviewer when the confidence level or financial significance falls outside an approved threshold. If the reviewer corrects the proposed treatment, the system records not only the final account but also the reason, the supporting evidence and the circumstances in which the decision applies.

Consider the payroll process for an organization. During the monthly payroll run, the system compares the payroll register with the active employee master file and identifies two employee records that did not exist in the previous pay period. Rather than simply flagging the additions, the workflow recognizes that new employees introduce a series of dependent activities extending beyond payroll. It retrieves the onboarding status from the organization’s human resources workflow to confirm that these employees were intentionally hired, verifies whether the required tax forms and employment documentation have been completed, and determines whether payroll has sufficient information to calculate statutory deductions accurately. If documentation is still outstanding, the workflow can recommend the use of approved default withholding assumptions, notify the appropriate HR representative that the missing information is required, and allow payroll processing to continue within the organization’s established policies rather than delaying these employees pay.

During the monthly payroll run, the system identified another employee included in the payroll run that when compared with the employee security access list, identified their access had been disabled. Rather than assuming an error has occurred, it recognizes an inconsistency between the organization’s security list and payroll run. The workflow confirms the employee’s HR status and compares it to recent security events to determine next steps which might be routing this exception to the appropriate manager if the information is inconsistent. The objective is not simply to identify an exception but to establish whether the underlying business processes remain aligned before payroll is finalized.

In both situations, the payroll controls have not changed. The organization still requires authorization of new employees, completion of statutory payroll documentation, segregation of duties, and appropriate review of employment status. What changes is the applied workflow. Instead of following a rigid sequence of predefined steps, the workflow adapts to the evidence presented by each situation, coordinating information across payroll, human resources, identity management and compliance processes to achieve the same control objectives more efficiently.

In accounting, autonomy must be bounded by authority, materiality, data access and professional responsibility. A system may gather evidence, prepare a recommendation and coordinate review without being authorized to post an entry, release a payment or issue financial information. High-risk actions should require explicit approval, and the basis for automated decisions must be retained. Google’s 2026 announcement of a proactive AI assistant reflected this distinction: the system was designed to continue routine work independently while requesting permission before higher-stakes actions.[4] The principle is directly relevant to accounting. A useful system should be able to advance the work without obscuring who remains accountable for the result.

The design also requires a separation between deterministic controls and probabilistic reasoning. Bank balances, approval limits, user permissions and mathematical relationships should not depend on a language model’s judgment. They should be enforced through reliable rules and system controls. AI is more appropriately used where the work depends on interpretation, context, prioritization or the selection of the next step. A sound architecture combines both: fixed controls establish the boundaries, while adaptive reasoning helps determine how the work should proceed within them.

This approach is receiving attention in current process research. A 2026 paper on agentic business process management systems describes a shift from design-driven process management toward systems that can sense process conditions, reason about improvement opportunities and act to maintain performance.[5] Research on AI-native enterprise resource planning has similarly examined systems that interpret user intent and assemble workflows across specialized agents rather than relying entirely on static, rule-based sequences.[6] These studies do not establish that mature, autonomous enterprise systems are already available. They do show that workflow composition, learning and orchestration have become serious areas of research rather than speculative extensions of chat interfaces.

The market evidence is also more measured than the promotional language suggests. The Financial Times has reported that early adopters are finding practical uses for agents in activities such as candidate screening, marketing follow-up and cybersecurity investigation, while widespread deployment remains limited.[7] It has also reported that the cost of running complex agents is causing organizations to impose usage limits and examine whether each automated task produces sufficient value.[8] These constraints reinforce the importance of workflow design. An agent should not be used where a rule, lookup or established software function can complete the work more reliably and at lower cost. The model should be reserved for the portions of the process that genuinely require interpretation or adaptation.

For accounting firms, the strongest systems will be built around an orchestration layer that coordinates workflows across specialized agents, business rules, organizational knowledge and professional oversight. Rather than sending every transaction directly to a large language model, the orchestration layer continuously determines the most appropriate path based on the information available and the objective being achieved. Deterministic rules resolve routine transactions where the outcome is known. Historical treatments and prior decisions identify recurring patterns and client-specific practices. Source documents, accounting records and external evidence are assembled to establish context and support conclusions. Specialized agents perform targeted activities such as reconciliations, payroll analysis, document collection or variance investigation, with the orchestration layer coordinating their activities and managing the flow of information between them. Only when sufficient evidence cannot be established, conflicting information exists or professional judgement is required does the workflow escalate the matter to an accountant for review. Professionals are no longer consumed by repetitive processing but instead focus their attention on the relatively small number of transactions and events that genuinely require experience, skepticism and judgement. This structure is not only more economical; it is more consistent with professional accountability.

The orchestration layer also optimizes the use of AI itself. Not every task requires a large language model, and not every decision benefits from probabilistic reasoning. The orchestration layer selects the most appropriate capability for the work at hand, invoking deterministic logic where certainty exists, specialized agents where domain expertise is required, and large language models only when contextual reasoning adds measurable value. AI becomes one component of the workflow rather than the workflow itself.

Impact on the Accounting Professionals

The effect on software competition may be substantial. Traditional accounting software competes through feature depth, integration, usability and the size of their installed base. Those factors will continue to matter. Adaptive solutions introduce another source of value: the quality of the learning environment created around the application. Two firms using the same ledger could achieve different results because one has developed clearer decision rules, better evidence capture, stronger review feedback and a more complete record of how its professionals resolve exceptions.

Over time, that record becomes a form of organizational capital. Review notes no longer disappear inside completed files. Corrections become reusable guidance. Client-specific knowledge becomes available at the point of work. New employees can be shown not only what was decided, but why the matter required attention and when the same reasoning should be applied again. The system does not replace professional development; it gives the organization a more consistent way to preserve and distribute what its professionals have learned.

There are also limits that should be recognized. Historical decisions may contain errors or reflect circumstances that no longer apply. A workflow that learns without disciplined review can institutionalize poor practice. Client information may be incomplete, permissions may be excessive and model outputs may be difficult to explain. Adaptive workflows therefore require governance over what can be learned, who can approve a new pattern, how long decisions remain valid and how the system responds when evidence conflicts with prior treatment.

These requirements make the role of the accountant more important, not less. The profession will be responsible for defining the boundaries within which systems operate, determining the evidence required for different decisions, establishing escalation thresholds and reviewing whether the workflow remains appropriate. Technical accounting knowledge will remain essential, but it will increasingly be expressed through the design and supervision of the environment in which the work is performed.

The shift will also change implementation. Installing an adaptive system cannot be limited to migrating data and configuring a chart of accounts. The organization must identify its objectives, controls, sources of evidence, decision rights and exception criteria. It must decide which knowledge should be standardized across the firm and which should remain specific to a client or engagement. It must create a feedback process through which corrections improve future work without becoming unexamined precedent.

This is a more demanding undertaking than adding an AI assistant to an existing product. It is also more likely to produce durable value. Personal productivity tools can save time for individual employees, but they rarely change how responsibility, evidence and decisions move across the organization. A dynamic workflow can improve the process itself because it connects the work performed by different people and systems to a common objective.

Final thoughts

For decades, accounting software has required organizations to adapt their processes to fit the software, and this approach has served accounting professionals as it delivers consistency, strong controls and reliable financial reporting results. Those principles are not changing but what is changing is how work is organized around those controls or the AI driven accounting workflows.

As organizations evolve, new legislation is introduced or business processes change, the workflow no longer needs to wait for the next software release or an upgrade project of next release. The operating environment can adapt by changing the information it captures, the evidence it requests and the sequence in which work is performed, while the underlying accounting controls remain intact.

That, in my view, is the real opportunity created by autonomous AI solutions. AI is an operating environment that understands the objective, coordinates the work required to achieve it and continuously improves as the organization learns.

Author

References

[1] Matt O’Brien, ‘What does agentic AI mean? Tech’s newest buzzword is a mix of marketing fluff and real promise,’ Associated Press, November 18, 2025.

[2] Financial Times, ‘Microsoft’s early AI lead has become a test of faith,’ July 10, 2026. See also Satya Nadella’s essay, ‘A Frontier Without an Ecosystem Is Not Stable,’ discussed in subsequent reporting on organizational learning systems and the ownership of enterprise knowledge.

[3] Thomas H. Davenport and Thomas C. Redman, ‘How to Marry Process Management and AI,’ Harvard Business Review, January-February 2025.

[4] Michael Liedtke, ‘Google announces slew of AI advances, including a personal AI assistant coming soon,’ Associated Press, May 19, 2026.

[5] Marlon Dumas, Fredrik Milani and David Chapela-Campa, ‘Agentic Business Process Management Systems,’ arXiv, January 25, 2026. Position paper based on the 2025 Workshop on AI for Business Process Management.

[6] Hongyang Yang et al., ‘FinRobot: Generative Business Process AI Agents for Enterprise Resource Planning in Finance,’ arXiv, June 2, 2025.

[7] Financial Times, ‘Lessons from the agentic AI trailblazers,’ May 2026.

[8] Financial Times, ‘Businesses face up to budget-busting AI bills,’ June 30, 2026; and ‘We created a monster: companies rein in AI usage as costs strain budgets,’ June 19, 2026.

Source links

Associated Press – agentic AI explained

Financial Times – Microsoft’s early AI lead

Harvard Business Review – How to Marry Process Management and AI

Associated Press – Google proactive AI assistant

Agentic Business Process Management Systems

FinRobot

Financial Times – Lessons from agentic AI trailblazers

Financial Times – Businesses face budget-busting AI bills