AI Agents: What They Are & What They Aren't

AI Agents: What They Are, What They Aren’t, and Why the Difference Will Cost You

Of all the terms currently circulating in the accounting technology market, none is more over-used, more misunderstood, or more commercially exploited than “AI agents.”

It appears in vendor presentations, platform release notes, conference keynotes, and software brochures at a rate that has now become almost meaningless. Every platform that announced “AI” last year is announcing “AI agents” this year. Every tool that offered smart suggestions is now offering autonomous workflows. The language has escalated. The underlying capability, in most cases, has not kept pace.

This is not a minor terminology dispute. The gap between what “AI agents” means in the marketing materials and what it means in practice, in real workflows, under real operational conditions, for real accounting firms, is significant enough to determine whether your investment in this technology transforms your practice or merely decorates it.

I have spent over a decade building and deploying AI and automation in enterprise environments, financial services, regulated industries where the consequences of a poorly designed autonomous system are measured in regulatory sanctions, not just inefficiency. I know what genuine agency looks like. I know what it requires. And I know how far most of what is currently sold under that label falls short.

This article is an attempt to give accounting professionals the vocabulary and the framework to tell the difference.

What an AI Agent Actually Is

Let us start with a precise definition, because precision is the only thing that protects you from being sold something you don’t want.

An AI agent is a system that can:

Perceive, receive and interpret inputs from its environment, whether that is a data feed, a document, a trigger event, or a natural language instruction.

Reason, evaluate what it has perceived against a goal, consider multiple possible actions, and determine the most appropriate next step. Not follow a script. Reason.

Act, execute that action, which may involve using external tools, querying other systems, writing data, sending communications, or initiating other processes.

Adapt, observe the outcome of its action, update its understanding of the situation, and determine what to do next, including handling exceptions and unexpected states it has not been specifically programmed to address.

The critical word in that list is “reason.” The difference between an AI agent and a sophisticated piece of automation is not the number of steps it can perform or the speed at which it performs them. It is whether the system is reasoning, evaluating context, weighing options, making decisions, or whether it is executing a predefined sequence of instructions that happens to be triggered automatically.

A workflow that fires when a transaction is received, categorises it using historical patterns, and posts it to the correct account is automation. Fast, useful, valuable automation, but automation.

An agent receives the same transaction, recognises that it does not match any known pattern, considers the available options, queries the client’s historical data for context, determines that it most closely resembles a category from six months ago that was manually adjusted, applies that category with a confidence flag, and notifies the relevant accountant with a summary of its reasoning, so the accountant can confirm, override, or let it stand.

The outputs of these two systems may look similar in many cases. The architecture underlying them is fundamentally different. And that architectural difference is what determines the ceiling.

The Four Levels of Agent Capability

Wolters Kluwer, in their research on the future of accounting firms, offers a useful taxonomy that maps agent capability onto a spectrum. It is worth understanding because it gives you a common language for evaluating what vendors are actually offering.

Taskers, automate low-value, repetitive tasks. Document classification. Data extraction. Scheduled reminders. These are the simplest agents, and the most widely deployed. Most of what the market currently calls “AI agents” in accounting is operating at this level.

Automators, run entire defined processes end-to-end. Flowing categorised transactions into trial balances. Preparing a VAT return from structured data inputs. Generating a standard management report from a live data feed. Still rule-governed, but operating across a complete workflow rather than a single task.

Collaborators, provide intelligent guidance within complex workflows. Routing decisions. Escalation logic. Contextual suggestions based on what has happened at earlier stages of the process. This is where genuine reasoning starts to appear, the agent is not just executing, it is interpreting.

Orchestrators, coordinate multiple agents to deliver a complete outcome. Moving a tax return from client data intake through categorisation, anomaly detection, draft preparation, exception flagging, and professional review, with the appropriate agent handling each stage and the orchestrator managing the handoffs, the dependencies, and the decision points throughout.

Most accounting software currently operates at Tasker level. Some of the more advanced specialist platforms are reaching Automator. Collaborators are rare. Orchestrators, genuine multi-agent systems that coordinate an end-to-end workflow with reasoning at the handoff points, are the frontier. They exist in some but few enterprise environments. They seem to exist, to me at least, in YouTuber videos claiming to have the secret to creating AI Agents. But, they are beginning to emerge, even in accounting. But, they are not yet widespread.

When a vendor tells you they have “AI agents”, the first question is: which level? The answer will tell you a great deal.

What Most "AI Agents" in Accounting Actually Are

In the interest of being concrete, let me describe what the current market landscape looks like, not from a position of cynicism, but from one of clarity.

The conversational interface. A chat window, often positioned prominently in the platform UI, that allows users to ask questions in natural language and receive responses drawn from the platform’s data. This is a useful feature. It is generative AI applied to data retrieval. It is not an agent. It does not act. It responds.

The smart suggestion. A feature that observes your behaviour and offers a recommendation, categorise this transaction as X, match this payment to Y, flag this anomaly for review. You still decide. You still click. The system has suggested; it has not acted. This is pattern recognition, not agency.

The scheduled workflow. A predefined sequence of steps that fires on a trigger, a new document arrives, the workflow runs: extract data, categorise, post, notify. This is automation. Valuable, important automation. But the steps are fixed. The sequence is predetermined. If something unexpected happens, the workflow stalls, errors, or misfiles, because it was not designed to handle the unexpected.

The “agent” that requires approval at every step. A system described as an agent that presents each action for human confirmation before proceeding. There is nothing wrong with this as a governance model, human-in-the-loop is often exactly the right design choice, particularly in a regulated profession. But a system that cannot act without prompting at every stage is not autonomous. It is a very well-organised suggestion engine.

None of these is without value. The issue is not that these things exist, it is that they are being sold under the label of “AI agents”, which creates false expectations about what they can deliver and, more dangerously, a false sense of having arrived at the frontier when you have not yet left the starting point.

Where Genuine Agency Is Emerging

To be fair to the market, the foundations of genuine agentic capability is beginning to appear, and it is worth acknowledging. There are several examples of those pushing the frontiers and pursuing agentic accounting AI.

BILL, in the US market, has deployed a W-9 Agent that autonomously collects and validates vendor tax forms, and a Reconciliation Agent that codes expense transactions through to reconciliation without manual intervention. These are not conversational interfaces or smart suggestions. They are systems that take a defined task and complete it, end-to-end, within governed parameters.

In February 2026, Pilot announced what it described as the world’s first fully autonomous AI accountant for small businesses, a system that manages the entire bookkeeping process, from client onboarding through to monthly close, without human intervention in the routine workflow. Whether the claim fully holds up under operational scrutiny is a separate question; what matters is that the architectural ambition is genuine, and the direction of travel is clear.

Sage’s MTD AI Agent, announced at Accountex Manchester in September 2025, coordinates the quarterly MTD workflow end-to-end, client segmentation, document chasing, deadline management, exception flagging, across a defined process. This is Orchestrator-level thinking applied to a specific, well-scoped accounting workflow.

These are meaningful developments. They represent the early arrival of genuine agency in accounting software. But they are also the exception, not the rule. For every product operating at this level, there are dozens being marketed with the same language that are operating at Tasker level or below.

The accounting firm that can tell the difference is the one that will make the right investment.

The Five Questions That Cut Through the Noise

In every industry I have worked in, the ability to interrogate a vendor’s AI claims has been one of the most valuable skills a senior decision-maker can develop. Here are the five questions I would ask any vendor positioning their product as an AI agent for accounting.

One: “What decisions does the system make autonomously, and what triggers human review?” A genuine agent has defined decision boundaries. It knows what it can handle independently, what it should flag for human oversight, and how it escalates when it encounters something outside its parameters. If the vendor cannot answer this specifically, if the answer is “it handles everything automatically” with no mention of governance or escalation, that is a red flag in either direction. Either the system is not truly autonomous, or it is autonomous without adequate control.

Two: “How does it handle a situation it hasn’t been trained on?” This is the test of genuine reasoning versus sophisticated pattern matching. A scripted automation fails, errors, or misfires. A reasoning agent evaluates the situation, considers what it knows, makes a decision, even if that decision is to escalate to a human, and explains its reasoning. Ask for a live demonstration with an edge case. Watch what happens.

Three: “What tools and external systems can it interact with natively?” A true agent is not confined to one platform. It can query external systems, retrieve data from multiple sources, trigger actions across tools, and orchestrate across your technology stack. An agent that can only operate within the boundaries of its own platform is not an orchestrator, it is a workflow automation with a conversational interface. One major obstacle to this, is the lack of mature agentic connection protocol frameworks, this is going to take time unfortunately.

Four: “Can you show me the audit trail of a decision the agent made?” In a regulated profession, every action that has compliance implications needs to be traceable. A production-ready agent generates a record of what it perceived, what it decided, why, and what it did, not just what the output was. If the vendor cannot show you this, the system is not ready for live accounting work.

Five: “What happens when the agent makes a mistake?” Every autonomous system operating at scale will occasionally produce an incorrect output. The question is not whether errors occur, it is how the system detects them, how they are corrected, how the learning from the correction is applied, and what the governance process is for catching errors before they reach a client or a filing. If the answer to this question is vague, the system is not enterprise-ready.

The Agency Paradox and Why Most People Aren't Actually Ready for It

Here is something I have observed repeatedly in real deployments, and something the market almost never talks about honestly.

Accounting firms ask for AI agents. They are shown what an agent can do, the multi-step workflow, the autonomous decision-making, the end-to-end execution without human prompting at every stage. They are initially impressed. They want it.

And then, as the discussion progresses, something shifts.

The agent makes a decision they didn’t anticipate. It handles an edge case in a way that surprises them. It acts, which is, after all, exactly what it was asked to do, and the response is not confidence. It is discomfort. Suddenly, the conversations change. Can we add a review step here? Can we require approval before it does that? Can we limit what it’s allowed to decide?

What started as a request for genuine agency becomes, through a series of entirely understandable interventions, something much closer to a supervised workflow with a very capable assistant at the centre of it. The agent is still there. But it has been progressively constrained, hedged, and supervised until its autonomy, the thing that made it an agent rather than an automation, has largely been removed.

I do not say this as a criticism of the firms involved. At Bots For That, we encounter this dynamic regularly. We find ourselves adapting our implementations, inserting more human-in-the-loop interactions, adding oversight at points where the system was designed to act independently, not because the agent is wrong, but because the humans working alongside it are not yet ready to trust it fully. Or give it the time and data to become the agent it can, just like hiring a new person.

This is the real frontier of AI in accounting. Not the technology. The trust.

And trust in autonomous systems is not simple. It is not just a question of demonstrating that the agent gets things right, though that matters. It is a question of demonstrating that the system is understandable, predictable, and controllable in a world that is dynamically changing, where the risks are growing and not fully mapped, and where the professional and regulatory consequences of an error are real.

True agency means giving an autonomous system the authority to determine how, when, why, and what, not providing it with a set of rules and decision trees dressed up as agency. The moment you specify every decision the agent should make, you have not built an agent. You have built a workflow with better marketing. Only, one which now has a higher cost than older, simpler methods.

The profession will get to genuine agency. But it will get there gradually, as trust is demonstrated and earned through track record, transparency, and, critically, the ability to explain what the agent decided and why, even when the human asking the question does not fully understand the underlying technology.

In the meantime, the honest observation is this: if you are implementing an AI agent and finding yourself adding approval steps, limiting its decision authority, and requiring human sign-off at every significant point, you are not implementing an agent. You are implementing the first stage of a trust-building process that, if managed correctly, will eventually get you there.

That is not a failure. It is a realistic starting point. Know what you are doing and why, and build toward the real thing with a clear view of what it requires.

Why This Matters for Your Firm, Right Now

The reason this precision matters is not academic. It has direct commercial consequences.

Accounting firms that invest in genuinely agentic systems, systems that reason, adapt, and orchestrate across complete workflows, will, over a three-to-five year period, develop a structural operational advantage that firms running Tasker-level automation will be very unlikely to close. The compounding effect of genuine autonomy, the hours returned to the team, the capacity redirected to advisory work, the error rates falling as the system learns, is qualitatively different from the efficiency gains of smart suggestions and scheduled workflows.

Equally, firms that invest in “AI agents” that are actually sophisticated automation scripts will find, twelve to eighteen months in, that the promised transformation has not materialised. The workflow is faster. The categorisation is more consistent. But the capacity has not been freed up in the way that was expected, because the humans are still in the loop at every significant decision point, confirming, correcting, and approving, just with a better interface around the process.

The Wolters Kluwer Future Ready Accountant data makes the underlying point clearly: firms with highly integrated technology stacks, the prerequisite for genuine agentic capability, are significantly more likely to be experiencing revenue growth than those operating in more fragmented environments. The integration is not the cause of the growth. But it is the foundation on which the agentic capability that drives growth is built.

True agentic AI is arriving in accounting. The firms that invest in the real thing, built on genuine AI infrastructure, by companies that have deployed agents at scale, in regulated environments, under real operational pressure, will be in a categorically different position to those who bought the marketing.

Know what you’re buying. The difference is not a matter of terminology. It is a matter of where your firm will be in five years.

| Daniel Lawrence is the Founder of Bots For That and creator of the beanieverse platform, a suite of AI-powered tools for the accounting and bookkeeping sector. With over a decade of experience deploying enterprise automation and AI in financial services and other highly regulated industries, he writes about AI transformation in accounting from the outside in. |