Agentic AI does not automatically get cheaper at scale

There’s a belief forming around AI that accounting firms need to challenge before it damages margin.

The belief is simple. Build an autonomous agent once, run it across 50 or 500 clients, and the cost per job will fall sharply.

That can happen with narrowly defined automation. It is far less certain with agentic workflows.

A workflow that reads documents, determines what matters, requests missing information, checks source records, posts a draft journal, and explains exceptions is not one prompt repeated 500 times. It is a chain of decisions. Every decision can trigger more model calls, more tool use, more context loaded into the workflow, and more human review.

That is the issue raised in reporting on Gartner’s view of agentic AI economics. Multi-step reasoning does not necessarily benefit from the same economies of scale people expect from traditional software. You can have more volume and still find that an agent consumes more resources per job when the underlying cases are messy.

For an accounting or bookkeeping firm doing $1 million to $25 million in revenue, that matters. Your margin is shaped by the work that arrives incomplete, late, out of sequence, or with unclear client context. AI sees that same mess unless you design the workflow around it.

The right question is not, “Can we deploy an autonomous close agent across the whole client base?”

Ask this instead.

“What does one completed, reviewed job actually cost us when the agent handles it?”

That includes token use. It includes integration calls. It includes the team member who reviews outputs. It includes exception handling. It includes the senior reviewer who gets pulled in when a decision affects a client’s financial statements.

Before you expand agentic AI, price one narrow workflow. In most firms, document intake or bank reconciliation is a better starting point than an end-to-end autonomous close.

The hidden cost inside an agent job

A staff member completing a reconciliation has a visible cost. You can estimate their time, apply a loaded cost rate, and compare it to what the client pays.

An AI agent has a less obvious cost structure because its work is distributed across steps.

Take a basic bank reconciliation. The apparent job is to match bank transactions to ledger entries. In practice, the workflow might involve:

• Pulling transactions from the bank feed and accounting platform
• Filtering duplicates, pending items, and incomplete descriptions
• Comparing transactions to existing ledger activity
• Looking at prior coding for recurring vendors
• Checking supporting documents in a client portal or shared drive
• Identifying transactions that need client clarification
• Suggesting coding and reconciliation status
• Producing an exception list for a bookkeeper
• Escalating unusual items for review

A well-built agent can reduce handling time here. But it may also make 20, 50, or 100 model and tool calls in the process, depending on how it has been designed. A simple recurring vendor match is cheap. A transaction with an ambiguous merchant name, a missing receipt, a personal expense risk, and a prior-period adjustment is not.

Then there is review.

If the agent gives your team 95 plausible answers and five weak answers, the economic result depends on whether the reviewer can spot the five weak ones in two minutes or has to re-check every answer. Firms often measure the agent’s completion rate and ignore the review burden. That is how an apparently productive deployment becomes a margin leak.

The real cost per completed job is closer to this:

AI runtime cost + integration cost + review time + exception time + rework cost

You don’t need perfect cost attribution in week one. You do need a disciplined baseline. If you can’t estimate those five components on a small sample, you can’t responsibly forecast the cost of deploying the workflow across your client base.

This is one reason the AI audit for accounting and bookkeeping starts with workflow economics, not generic AI capability. The point is to find where an agent can produce a measurable improvement without creating a new layer of operational work.

Start with one narrow accounting workflow

The best pilot has three qualities.

First, it has enough volume to show you patterns. Ten unusual jobs will not tell you much. A sample of 50 to 100 reconciliation tasks, intake packets, or AP coding batches usually starts to reveal the true mix of clean work and exception-heavy work.

Second, it has a defined finish line. “Help with month-end” is not a finish line. “Produce a reviewed list of unmatched transactions with proposed coding” is.

Third, it has clear human ownership. Someone needs to decide what the agent can complete, what it can draft, and what must remain under review.

For many firms, document intake is the cleanest place to start.

A document intake workflow worth pricing

A new bookkeeping client might send bank statements, payroll reports, AP aging, prior financials, tax returns, loan documents, and access credentials over a two-week period. Some documents arrive through a portal. Some arrive by email. Others come as photos, exports, or files with unclear names.

Your team then chases missing information, classifies documents, identifies account history, maps a chart of accounts, and works through historical clean-up before any recurring work becomes stable.

That onboarding drag is costly. We often see firms lose weeks before billable monthly work settles in. A meaningful share of new clients may delay productive work by a quarter because the inputs are incomplete or the clean-up is larger than expected.

The Client Onboarding Agent in Omni ops can collect documents through a guided workflow, identify what has been received, request what is still missing, prepare a chart-of-accounts starting point, and produce a clean opening trial balance for review.

But don’t price this as “one agent per new client.”

Price the steps:

1. How many documents does the average client provide?
2. How many need OCR, classification, or data extraction?
3. How often does the agent need to ask a follow-up question?
4. How many follow-ups are resolved without staff intervention?
5. How long does a bookkeeper spend reviewing the proposed chart mapping?
6. How many cases need partner input because the business model is unusual?
7. How many opening balances require historical investigation?

A straightforward professional services client may take a fraction of the AI and review effort required for a contractor with multiple bank accounts, financed equipment, subcontractors, owner draws, and two years of uncategorised activity. Treating both as the same “onboarding run” will give you false economics.

Use a cost card before you automate volume

Build a simple cost card for the pilot workflow. Don’t make it complicated. The purpose is to compare actual delivery cost with your current method and with the fee you can sustain.

Track each job across five fields.

Cost area	What to measure
AI runtime	Model calls, tokens, tool calls, document processing, and storage used per job
Staff review	Minutes spent checking, approving, correcting, or escalating outputs
Exception work	Minutes spent on missing data, client questions, unsupported integrations, and edge cases
Quality	Error rate, rework rate, and the type of errors found
Throughput	Time from job received to job completed and reviewed

Then segment the results. Do not average everything into one number.

Split clean jobs from messy jobs. Split recurring clients from new clients. Split high-document-volume clients from low-volume clients. You may find that 70% of work is efficient and 30% is expensive enough to justify a different service tier, stricter intake rules, or human-led handling.

That finding is valuable. It tells you where the agent should be deployed, where it needs stronger controls, and where the client relationship needs to change.

If the agent saves 20 minutes on a clean reconciliation but creates 15 minutes of review on every case, the design needs work. If it saves 45 minutes on a batch of routine bank transactions and only requires five minutes of review, you have a workflow worth extending.

This is where Omni ops is useful as an operating model, not just a software concept. The work needs triggers, source systems, approval points, exception queues, and a clear completion standard. Without those, the agent is a clever assistant that creates unpredictable handoffs.

Review cost is the number most firms miss

Token pricing gets attention because it is easy to see on a bill. Review cost is often larger.

Think about month-end. Your team is under pressure to complete bank reconciliations, AP and AR checks, payroll posting, accruals, variance analysis, and reporting packs. In many firms, 30% to 50% of staff workload is concentrated in a handful of month-end and year-end weeks.

Now introduce an agent that drafts reconciliations and journal entries.

The Month-End Close Agent can pull bank, AP, AR, and payroll feeds, reconcile accounts, flag variances, draft journal entries, and prepare a partner-ready close pack. That can remove a large amount of repetitive preparation work.

The constraint is not just whether it can create a draft. The constraint is the review design.

A reviewer should not need to reperform the close to trust the draft. They need:

• Source links for material numbers
• A visible explanation of the matching or coding logic
• Confidence scores or clear exception labels
• A defined threshold for automatic treatment
• An approval queue for entries that affect material accounts
• A record of what was changed after the agent’s recommendation

If the agent gives clear evidence and routes only genuine ambiguity to the reviewer, it can improve margin. If it produces fluent explanations without traceable source data, your team still has to do the real accounting work underneath it.

This is also why autonomous deployment should come later. A practical first stage is an agent that prepares, proposes, and escalates. It does not post material journals without controls. It does not resolve missing documentation through guesswork. It does not decide tax treatment from incomplete facts.

A firm owner should be pleased with that restraint. Accuracy, audit trail, and client confidence matter more than an impressive demo.

If you want a practical worksheet for mapping these handoffs, download the Month-End AI Close Map for Accounting Firms. The direct version is available here: download the close map. Use it to list the close steps, systems, evidence required, approval points, and exception paths before you ask an agent to run them.

Price the workflow against your margin opportunity

The annual leakage band for accounting and bookkeeping firms is often in the range of $60,000 to $180,000. That is not all labour that can be removed. It is a mix of avoidable rework, unbilled clean-up, delayed onboarding, partner review time, missed capacity, and low-value admin that crowds out better work.

The opportunity is not simply to cut hours.

A firm that creates capacity can use it in several ways:

• Complete month-end work earlier and reduce overtime
• Take on a controlled number of additional recurring clients
• Improve onboarding speed and reduce early client frustration
• Move managers into review and client-facing work
• Create room for advisory conversations

The last point matters. Advisory work commonly commands two to three times the rate of routine compliance work. Yet advisory conversations are regularly pushed aside because the team is still trying to get the books closed.

The Advisory Insights Agent is designed for that next layer. It reads each client’s monthly numbers, identifies three relevant issues to discuss, and drafts partner talking points before the meeting. It should not invent financial advice or replace professional judgement. It should help your team arrive prepared, with the basic analysis already assembled.

That agent is only valuable if the close data is timely and dependable. Which brings us back to workflow pricing. If the Month-End Close Agent creates a costly review bottleneck, it does not create the capacity the Advisory Insights Agent needs.

You can Book a 60-min Omni Audit when you want to test the economics on a real workflow rather than work from assumptions.

What a sensible pilot looks like

A sensible pilot is not an announcement that the firm is “going agentic.” It is a controlled operating test.

Pick one workflow, one client segment, and one team owner. Run it for a defined period, often four to six weeks for a recurring monthly workflow. Keep the current process available as a fallback.

For each job, record:

• Inputs received and data quality
• Number of agent actions and tool calls
• Agent output produced
• Review minutes
• Corrections made
• Exceptions escalated
• Final completion time
• Any client-facing impact

Use that record to make decisions that are hard to see in a demo.

You may find the workflow is ready for broader use with one change, such as requiring clients to submit statements by a certain date. You may find that the agent is viable only for clients using a specific accounting platform. You may decide that lower-tier clients need a different service scope because the exception burden is permanently high.

All three are good outcomes. The pilot has done its job by showing you the actual delivery model.

For practical implementation thinking, the Omni platform is built around this idea of connected work rather than standalone prompts. Your agent needs access to the right data, a defined set of actions, and rules for when a person takes over. The technology matters, but the workflow design determines whether it pays for itself.

Don’t scale uncertainty

There is no prize for deploying an autonomous agent across every client before you understand cost per completed job.

Accounting firms already know how dangerous unpriced work can be. A fixed-fee client with late records, unclear ownership decisions, and historical clean-up can consume the margin from several clean accounts. Agentic AI can reproduce that pattern at speed if you don’t set boundaries.

Start with a narrow workflow. Measure runtime and review effort. Segment the exceptions. Improve the process before you expand volume.

Then you can make a confident decision about where autonomy makes sense, where drafting is enough, and where experienced accounting judgement should remain firmly in the loop.

See Omni for accounting and bookkeeping if you want to identify the workflow with the fastest measurable return. You can also Book my Omni Audit. In 60 minutes, we’ll identify the manual workflow worth testing, map the likely cost drivers, and define the three outputs you need to make a deployment decision. No deck.