The real question is not, can an AI agent do it?

For an accounting or bookkeeping firm, the better question is this: what will it cost to run the workflow for every client, every month, at the standard your firm is accountable for?

That distinction matters.

The current conversation around agentic AI often assumes that more volume means lower unit cost. In software, that is often true. Build a feature once and sell it to another 1,000 users, and the marginal cost is close to zero.

Multi-step AI workflows don’t always work that way.

A workflow that collects client documents, reads invoices, matches bank transactions, identifies exceptions, drafts journal entries, and asks a reviewer for sign-off can incur cost at each step. It may call several models. It may access multiple systems. It may create follow-up tasks when source documents are missing or confidence is low. The harder the client file, the more work the agent does.

That concern was central to Gartner’s view, reported by Computer Weekly, that agentic AI may not gain the expected economies of scale. For accounting firms, it is a practical warning, not an academic one.

Don’t approve an agent rollout based only on a demo that processed one clean bank feed and five neatly labelled invoices. Price the full workflow per client. Include AI usage, systems integration, exception handling, partner review, and the human work created when the workflow gets stuck.

If you get that number right before rollout, AI can protect margin and create room for advisory work. If you get it wrong, you can automate a low-value task and quietly create a new cost centre.

Why accounting workflows can get more expensive at scale

Accounting firms are not processing identical transactions for identical businesses.

One client has a clean Xero file, bank feeds connected, a stable chart of accounts, and a finance manager who responds within hours. Another has three bank accounts, a card provider with unreliable exports, a part-time office manager, invoices sent through text messages, and payroll adjustments made after the pay run.

The second client is where agent economics change.

An AI workflow has to do more than read data. It has to determine what is missing, request information, interpret a reply, compare it with the ledger, decide if the result meets a confidence threshold, and escalate the right item to a person. A simple workflow can turn into 12 or 15 actions before it reaches a reviewer.

That does not mean firms should avoid agents. It means they should stop treating every client as a flat monthly AI cost.

The cost drivers usually fall into five areas:

1. Volume of inputs. More invoices, transactions, entities, payroll lines, and inbox messages mean more work. This is obvious, but firms often count transactions and miss documents, messages, and follow-up requests.

2. Data quality. Poor source data creates repeated checking. A workflow may need to reread documents, search prior periods, or ask clients to clarify an item.

3. Workflow branching. A clean reconciliation may take one pass. A reconciliation with duplicate payments, unclear merchant names, intercompany transfers, or stale deposits creates extra model calls and human review.

4. System access. Agents working across Xero, QuickBooks, a document portal, payroll, CRM, email, and practice management software can have usage costs and integration overhead at every handoff.

5. Review standards. A draft journal entry is not a booked entry. A partner or manager still needs enough evidence to approve it. Firms that ignore review time will overstate the savings.

This is why a firm can increase its client count while seeing AI workflow cost per client stay flat or even rise. As the workflow expands into more complex work, the number of decisions and exception paths can grow faster than the base transaction volume.

The point is not to chase the lowest token bill. Token cost is only one line in the model. The bigger cost is often a poorly designed workflow that generates a queue of vague exceptions for senior staff.

For a clearer view of where these workflows fit, see Omni Ops. The goal is to design operations around defined outcomes and escalation rules, not hand over an uncontrolled process to a chatbot.

Map the work before you automate it

Take month-end close. It looks like one service line on an invoice. In reality, it is a chain of small tasks performed by different people over several days.

A bookkeeper checks whether feeds are connected. They chase missing statements. They code transactions. They match payments. They investigate variances. They post accruals, prepayments, payroll journals, and loan movements. A manager reviews unusual balances. A partner decides what needs to be raised with the client.

At year-end, that workload compounds. We commonly see 30% to 50% of staff capacity compressed into roughly four weeks around key reporting deadlines. That is where burnout, rushed review, and margin pressure show up.

A Month-End Close Agent can help, but only if it is designed around the real sequence of work.

The Month-End Close Agent (Omni ops) pulls bank, AP, AR, and payroll feeds. It reconciles transactions against rules and source documents, flags variances outside set thresholds, drafts journal entries, and prepares a partner-ready close pack.

That sounds straightforward until you price each stage.

For one client, a typical monthly sequence might include:

• Checking that each connected feed has updated and identifying gaps.
• Reading 40 supplier invoices and matching them to bills or payments.
• Reviewing 250 bank transactions, separating high-confidence matches from ambiguous items.
• Comparing AR and AP balances to the ledger.
• Checking payroll journals against the payroll system.
• Identifying material movements against the prior month.
• Drafting five proposed journals with links to evidence.
• Creating a close pack and a short exception list for reviewer approval.

Now multiply that by 80 clients. Then ask what happens if 25 of those clients have incomplete documents, broken feeds, or inconsistent coding.

That is the number to model. Not the cost of one prompt.

Use a per-client cost model that includes review

The simplest way to evaluate an agent workflow is to calculate a contribution view for each client segment.

Start with this formula:

Monthly agent cost per client = AI usage + integration usage + workflow maintenance + human exception time + reviewer time

Then compare it against:

Monthly margin gain = staff hours removed or redeployed × loaded hourly value, minus monthly agent cost

The loaded hourly value is important. Don’t use the lowest payroll rate. Include the true cost of delivery, including supervision and rework. If an experienced manager spends 20 minutes untangling an agent exception, that is not cheap capacity.

Use three client bands rather than one average.

Clean and connected clients

These clients have reliable feeds, disciplined document submission, and stable transaction patterns. They are the right first group for a Month-End Close Agent.

The agent can handle more routine work with fewer branches. Reviewer time should fall because the close pack is structured, evidence is linked, and exceptions are ranked by materiality.

This is where firms usually see the strongest early return.

Standard clients with recurring exceptions

These clients are not broken, but they need regular follow-up. They may have missing receipts, owner transactions, inconsistent invoice timing, or unusual payment references.

Here, agent cost can still make sense. The workflow needs strict limits. It should resolve routine items, send defined document requests, and escalate quickly when confidence drops.

Don’t let the agent repeatedly investigate an item that a bookkeeper can resolve in two minutes with context. Set a cap on retry loops and use a clear human handoff.

Complex or messy clients

These clients may have multiple entities, poor records, frequent late changes, inventory issues, related-party movements, or an owner who sends a pile of documents at the end of the month.

They might benefit from an agent, but they should not be priced like clean clients.

This is often where firms discover an uncomfortable truth. Their fixed-fee package is absorbing an amount of exception work that has nothing to do with ordinary bookkeeping. AI makes that visible. It does not automatically make it profitable.

For these clients, you may need a higher service tier, an onboarding remediation project, a document discipline requirement, or a human-led process until the file is stable.

That work is exactly what the AI audit for accounting and bookkeeping is built to uncover. We map the workflow, identify the exception drivers, and calculate where an agent should stop rather than continue.

Don’t confuse automation with autonomous judgment

The most useful accounting agents do not replace professional judgment. They create a better first pass and a better review pack.

A well-designed workflow should have clear boundaries:

• The agent can retrieve, classify, compare, calculate, and draft.
• It can apply approved rules to repeatable scenarios.
• It can ask clients for missing information using agreed templates.
• It can rank exceptions based on materiality and confidence.
• It cannot post high-risk entries without an approved review path.
• It should not invent an explanation for an unexplained variance.
• It should stop when the available evidence is insufficient.

Those boundaries protect quality, but they also protect economics.

If every uncertain item triggers three more research loops, the AI cost rises and the final answer may still need human review. A firm is better off setting a threshold such as, “If confidence is below the agreed level or the variance exceeds the materiality limit, create an exception with the evidence gathered so far.”

That gives the reviewer a useful decision. It doesn’t give them an AI-generated mystery to solve.

The same principle applies to onboarding.

The Client Onboarding Agent (Omni ops) collects documents from new clients through a guided workflow, sets up the chart of accounts, and produces a clean opening trial balance. It can send reminders, check whether the required items are complete, classify source documents, and surface gaps before the engagement gets stuck.

But onboarding is one area where firms should aggressively measure exception cost. Industry experience suggests that around 20% to 30% of new clients can delay billable work by a quarter when records and documents are incomplete. An agent can reduce the chasing. It cannot make a disengaged client provide a missing prior-year reconciliation.

Price onboarding as its own workflow. Don’t bury it inside a monthly fee and assume the agent makes the work free.

Build a pilot that produces real unit economics

A sensible pilot does not start with every client and every process.

Start with 15 to 25 clients in one service category. Choose files that are representative, not just the cleanest ones. Include enough variation to expose the branches you will face at scale.

For each client, record:

• Number of documents and transactions processed.
• Number of workflow steps completed.
• Number and type of exceptions raised.
• AI and integration usage.
• Human minutes spent resolving exceptions.
• Reviewer minutes to approve the close pack.
• Rework required after review.
• Days from period end to completed close.

Run the pilot across at least two monthly cycles. One month is rarely enough because client behaviour, late documents, and bank feed reliability vary.

At the end, segment the findings. You want to know which client characteristics predict cost. Is it transaction volume? Number of disconnected systems? Document timeliness? Entity complexity? A weak chart of accounts? Owner-managed spending?

That analysis gives you a commercial decision, not just a technology result. You can decide which clients belong in the automated service tier, what level of review is required, and where a surcharge or remediation project is justified.

If you want to work through that model with someone who understands both operations and delivery margin, Book a 60-min Omni Audit. It is a working session, not a slide deck.

The bigger payoff is advisory capacity

Most accounting firm owners do not want AI simply to process more low-margin compliance work. They want their best people out of the month-end firefight and into client conversations that improve retention and revenue.

That is where the Advisory Insights Agent (Omni ops) comes in. It reads each client’s monthly numbers, surfaces three things worth discussing, and drafts partner talking points before the meeting.

The economics are different here.

Advisory work often commands two to three times the billable rate of routine compliance work. That does not mean every AI-generated insight becomes a paid advisory engagement. It means the value of recovering even a few hours of partner or manager time can be materially higher than the direct cost saved on coding transactions.

The sequence matters. First, use agents to stabilise document collection and close. Then use the cleaner, earlier data to create better advisory conversations. If you jump straight to AI commentary while the underlying ledger is unreliable, you create risk and lose trust.

Our Omni Advisory work focuses on that progression. The aim is to connect operational automation to a client-facing decision process, not produce generic commentary that a partner has to rewrite.

Put a dollar range around the leakage

For accounting and bookkeeping firms in the USD $1 million to $25 million range, the annual leakage from manual handoffs, late documents, repeated review, and lost advisory time can sit in the $60,000 to $180,000 range.

That is not a universal benchmark. A firm with highly standardised clients may sit below it. A firm carrying legacy clients with poor record discipline may exceed it.

The important point is where to look.

Calculate the annual cost of staff time spent chasing documents. Add rework after close review. Add the margin impact of delivery delays. Add the advisory meetings that did not happen because managers were clearing exception queues. Then compare that total with the cost of a controlled agent workflow, including review.

You’ll have a much better investment case than “we need an AI strategy.”

If you need a practical way to map the close process internally, the Month-End AI Close Map for Accounting Firms is a useful worksheet. You can also download the direct close map for a step-by-step view of inputs, controls, exceptions, and reviewer decisions.

Treat pricing as a design decision

The firms that win with agentic AI will not be the ones that deploy the most agents. They will be the ones that know the cost and margin profile of each workflow by client type.

They will build an agent that handles routine work well. They will stop it when context is missing. They will give reviewers concise evidence. They will use the data from the workflow to reset service tiers, onboarding requirements, and client expectations.

That is a more durable model than assuming volume will solve the cost problem.

Start by looking at one workflow, preferably month-end close or client onboarding. Measure every human handoff, every follow-up, every exception, and every approval. Then decide where an agent can produce a reliable result at a cost that makes sense for that client.

You can see Omni for accounting and bookkeeping to understand the process we use to identify those opportunities. Or, if you are ready to put numbers around the opportunity in your own firm, Book my Omni Audit. In 60 minutes, we will identify the workflow, quantify the likely leakage, and define the first controlled agent use case.