AI agents don’t get cheaper just because you run more of them

A recent Gartner view, reported by Computer Weekly, cuts against a common assumption in the market. Agentic AI doesn’t necessarily produce the same economies of scale that firms expect from standard software.

That matters for accounting and bookkeeping firms.

A workflow agent can look cheap at the start. It reads a handful of documents, drafts a client email, or checks a bank reconciliation. Then you add more client entities, more source systems, longer transaction histories, validation steps, and exceptions. The agent has to reason through more context. It may call multiple tools. It may run a second review after an initial answer. Costs can rise faster than the number of jobs completed.

The operational risk isn’t just an unexpected AI invoice. It’s deploying an agent into client work with no clear boundary around what it can do, what it must escalate, and how much processing is acceptable before a human takes over.

For a $1M to $25M firm, this is where AI plans either become useful or become another technology cost centre. The firms getting value aren’t asking an agent to autonomously “do the books.” They’re breaking work into defined stages, metering usage, and setting approval checkpoints around accounting judgment.

The prize is still significant. In accounting and bookkeeping, we commonly see an annual leakage band of $60K to $180K across rework, partner review, delayed close cycles, fragmented client requests, and work that should have been billed as advisory. The goal is to remove that leakage without creating an agent system that quietly consumes margin.

Where multi-step agent costs build up

A standard software workflow often has a predictable cost. You pay per user, per month, or per transaction. An AI agent is different because its cost is tied to the work it needs to complete.

Every step can consume tokens and processing time. In a simple workflow, that may be trivial. In an accounting workflow, it can compound quickly.

Take a month-end close for a client with multiple bank accounts, payroll, receivables, payables, recurring journals, and a few messy coding exceptions. An agent may need to:

1. Pull data from the ledger, bank feed, AP system, payroll platform, and document store.
2. Read transaction descriptions and supporting documents.
3. Compare records to prior periods and client-specific rules.
4. Flag exceptions and retrieve more context.
5. Draft proposed journals.
6. Create a close summary.
7. Run a validation pass before presenting the work to a manager.

That is not one AI request. It’s a chain of requests, tool calls, checks, and retries.

The same issue appears during onboarding. An agent collecting historical data may repeatedly ask for missing bank statements, map account names to a new chart of accounts, identify opening balance gaps, and prepare a clean opening trial balance. If the client sends five partial document batches, the workflow must know when to continue, when to stop, and when to escalate.

Without controls, people often respond to weak agent output by adding more prompts, more checks, and more model calls. The agent becomes more expensive while still not resolving the actual issue, which is normally a missing document, a policy decision, or a human judgment call.

That is why scale needs governance. More clients do not automatically mean lower unit cost.

The manual work worth targeting first

Accounting firms don’t need an agent for every task. Start where work is repetitive, rules are visible, inputs are available, and a qualified person can approve the output quickly.

The three agent patterns we build in Omni ops are a strong fit for that approach.

Month-End Close Agent

The Month-End Close Agent pulls bank, AP, AR, and payroll feeds. It reconciles transactions, flags variances, drafts journal entries, and prepares a partner-ready close pack.

The useful part is not that it replaces the senior accountant. It reduces the work that happens before professional judgment.

Think about the current process. A bookkeeper chases missing documents, reviews unmatched items, compares account movements to prior months, copies explanations into a workpaper, and asks a manager whether a variance needs a journal. The manager then spends time finding what changed rather than deciding what it means.

In many firms, 30% to 50% of staff time is concentrated in four weeks of the year around month-end and year-end pressure. That isn’t only a capacity problem. It becomes a margin problem when experienced people spend late evenings rebuilding schedules and responding to preventable questions.

A well-bounded agent can assemble the evidence, identify exceptions, draft the narrative, and leave the final review with the person who owns the client relationship.

Client Onboarding Agent

The Client Onboarding Agent collects documents from new clients through a guided workflow, sets up the chart of accounts, and produces a clean opening trial balance.

Onboarding is a good example of where uncontrolled agent use can go wrong. The agent may be able to request documents and classify what arrives. But it shouldn’t keep making assumptions about a missing prior-year reconciliation or an unclear owner loan balance.

The workflow needs a checkpoint. Once the agent sees a defined condition, such as incomplete source records after two requests or a balance that doesn’t tie to supplied evidence, it creates an escalation task. It doesn’t keep reasoning indefinitely.

That protects cost and quality. It also improves the client experience. Industry ranges suggest 20% to 30% of new clients delay billable work by a quarter when onboarding drags through document collection and historical clean-up. A structured workflow gets the client to a clear next action faster.

Advisory Insights Agent

The Advisory Insights Agent reads each client’s monthly numbers, surfaces three things to talk about, and drafts the partner’s talking points before the meeting.

This is where the commercial upside is often missed. A firm can automate parts of compliance work, then fill the released time with more low-value clean-up. The better move is to use that time for client conversations.

Advisory work can often bill at two to three times the rate of compliance work. Yet it is usually the first thing crowded out when the close isn’t complete until the final days of the month.

An advisory agent should not invent a recommendation from a bare ledger extract. It should be given a controlled brief. Compare actuals to budget where available. Highlight cash conversion changes. Identify material expense movement. Surface overdue receivables. State the data gaps. Then let the partner apply judgment and client context.

For more examples of where this fits, our AI advisory work focuses on turning operational data into a preparation process that supports better conversations.

Meter tokens like you meter labour

If you would track chargeable hours, contractor spend, and software licences, you need to track agent usage too.

This doesn’t mean asking your team to become AI engineers. It means setting commercial guardrails before an agent touches a broad client portfolio.

Start with a cost baseline for one workflow. Pick a representative close, onboarding case, or advisory preparation cycle. Measure the total inputs, number of model calls, tool calls, validation runs, exceptions, and human review time.

Then create a simple cost model.

• Set a target cost per completed close pack, onboarding file, or advisory brief.
• Split standard cases from exception cases.
• Cap the number of retries the agent can make.
• Define the maximum documents or transaction lines processed in one pass.
• Track the cost of validation separately from the cost of drafting.
• Review cost per client entity, not only the total monthly AI bill.

The distinction between standard and exception work is important. A straightforward bookkeeping client may have clean feeds, recurring transactions, and a stable chart of accounts. An agent can complete much of the preparation cheaply and predictably.

A client with several entities, inconsistent source records, payroll corrections, foreign currency transactions, and late documents is different. That workflow should move to a human checkpoint sooner. You don’t want an agent repeatedly trying to resolve a problem that needs a five-minute call with the client or a manager decision.

This is a practical operating model, not a restriction on innovation. The agents that earn trust are the ones that know when to stop.

Build approval checkpoints into the workflow

Approval checkpoints are where firms retain accountability. They aren’t a sign that the agent failed.

In client accounting work, there are decisions that carry risk. A proposed journal may affect tax treatment. A chart-of-accounts mapping may change reporting consistency. An advisory insight may need commercial context before it is discussed with an owner.

Set checkpoints around those decisions.

A sensible month-end workflow could use four:

1. Data completeness check
   The agent confirms that required feeds and documents are present. Missing inputs create a task, not an assumption.

2. Exception threshold
   The agent can clear routine matching issues within defined tolerance rules. Material variances, unusual account movements, or uncategorised items move to a reviewer.

3. Journal approval
   The agent can draft a journal entry with its evidence trail. A qualified accountant approves any journal before posting.

4. Client-ready release
   A manager or partner approves the close pack and any narrative that will be shared with the client.

The same structure works for onboarding. The agent can gather, classify, and prepare. A human approves opening balances, mapping decisions, and any unresolved historical issue before the system goes live.

This is also why agent deployment should begin with a narrow workflow. Don’t connect a new agent to every ledger and ask it to reconcile every client. Start with one client segment, one close stage, and one approval owner. Learn where exceptions occur. Then expand.

If you need a practical starting point, download the Month-End AI Close Map for Accounting Firms. It is a worksheet for mapping the close sequence, the data needed at each stage, the escalation rules, and the sign-off points before you automate anything.

You can also access the direct worksheet here: Month-End AI Close Map download.

Don’t confuse a longer workflow with a better answer

The temptation with AI is to add another layer of reasoning whenever output isn’t quite right. Ask the model to check itself. Ask another model to critique it. Retrieve more files. Run a deeper analysis.

Sometimes that is justified. Often it is a sign that the workflow has no clear decision rule.

A stronger design asks three questions at every stage:

• What evidence must be present for the agent to continue?
• What result can it produce without human approval?
• What condition requires a person to take over?

For example, the Month-End Close Agent can draft an explanation of why advertising expense rose 18% against the prior month. It can point to supplier invoices and campaign activity. But if the client has no current budget and the spend is outside the normal pattern, the agent shouldn’t manufacture a conclusion. It should flag the issue, attach the evidence, and prepare the manager’s question.

That is lower cost than repeated reasoning. It is also safer.

You can see how these workflow principles connect across the broader Omni platform. The point is not to build a chat interface for your team. It is to create controlled operating workflows around the work your firm already does.

Put the dollar reality in front of the partners

A leakage band of $60K to $180K isn’t usually found in one dramatic failure. It appears in small recurring losses.

A manager spends 25 minutes checking whether a reconciliation explanation is complete. A partner reviews the same exceptions twice because the evidence wasn’t prepared. A new client waits three weeks for a document list that should have gone out on day one. A monthly review meeting becomes a compliance update because nobody had time to prepare insights.

These delays consume skilled capacity. They also reduce the number of advisory conversations your firm can have.

The right agent workflow should show financial impact in clear terms:

• Hours removed from routine preparation.
• Reduction in close-cycle rework.
• Time from client onboarding to first billable work.
• Number of client-ready advisory briefs prepared.
• Agent cost per completed workflow.
• Human review time per exception.

Don’t lead with a vague promise that AI will save hundreds of hours. Build a workflow that proves it in one service line. If it reduces five hours of monthly close preparation across a defined client group, calculate the capacity released. If it moves a portion of that capacity into advisory meetings, calculate the associated revenue opportunity. If it costs more to run than the time it removes, change the workflow before scaling it.

This discipline gives partners a better basis for deciding where to invest. You can find more practical operating material in our AI insights library, especially if you’re deciding where agents fit alongside your existing systems.

Use an audit to set the controls before scale

Most firms don’t need a slide deck full of generic AI ideas. They need one clear view of the workflows that are worth automating, the points where a human must approve work, and the commercial limit for agent usage.

That is what we cover in the AI audit for accounting and bookkeeping.

In 60 minutes, we work through three outputs:

1. A shortlist of workflows where agent support can remove preparation and follow-up work.
2. A control map covering token budgets, exception rules, data access, and approval checkpoints.
3. A practical commercial view of the time leakage and margin opportunity in your firm.

There is no deck for the sake of it. You leave with a view of what to test first, who should own the decision, and what needs to be measured before the workflow reaches more clients.

If month-end is straining your team, onboarding is delaying revenue, or advisory work keeps being pushed aside, Book a 60-min Omni Audit. We will map the work, identify the costly loops, and put approval points where they belong.

Control is what makes agent scale profitable

Accounting firms can gain real capacity from AI agents. The work is structured enough to support them, and the pressure on close cycles makes the opportunity hard to ignore.

But don’t assume that adding more clients to an agent will steadily reduce unit cost. Longer workflows, more validation, more exceptions, and larger data sets can push costs up. The firms that protect margin will meter usage, cap unnecessary retries, and make human approval a designed part of the workflow.

Start with a narrow workflow. Measure the unit cost. Escalate exceptions early. Keep final accounting judgment with the right person. Then extend what works.

To assess that approach against your own client mix and operating model, see Omni for accounting and bookkeeping or Book my Omni Audit.