Stop expecting scale to make agent costs disappear

There is a tempting story around agentic AI. Buy the tools, connect enough client data, put agents across the practice, and unit costs will drop as volume rises.

That story doesn’t hold up well inside an accounting or bookkeeping firm.

The Gartner view reported by Computer Weekly is a useful warning. Autonomous agents don’t automatically gain the same economies of scale as older software. Each workflow can carry its own data issues, exception paths, model usage, review burden, and integration costs. In accounting, those costs become very real at month-end, year-end, and during a messy onboarding.

An agent that can reconcile a clean bank feed for one client is not necessarily cheap or reliable when it meets 40 clients with different bank connections, chart structures, approval rules, payroll systems, and document habits.

For a firm doing $1 million to $25 million in annual revenue, that doesn’t mean avoid AI. It means stop treating AI agents as a broad deployment project.

Start with one bounded task. Define what a completed job is. Measure the full cost to complete it. Keep a human approval gate where judgement, materiality, or client risk enters the picture. Then decide if you should expand.

This is the discipline behind Omni ops. It is not about pushing more work through a chatbot. It is about designing a repeatable operating workflow where the firm knows what the agent does, what the reviewer does, and what each completed task actually costs.

The margin problem is usually hiding in the handoffs

Most accounting firms don’t lose margin because one senior accountant spends an hour on a difficult technical issue. That work is visible, scoped, and usually priced accordingly.

Margin leaks in the accumulated small tasks around it:

• Chasing a missing statement from a client
• Downloading files from an unstructured portal
• Renaming and filing documents
• Matching a transaction that the rules engine could not classify
• Asking a manager where to code an unusual payment
• Rebuilding a reconciliation after a feed disconnects
• Drafting a journal entry, then waiting for approval
• Reformatting a close pack for the partner meeting
• Pulling numbers together before an advisory call

None of these jobs is unusual. The problem is that they arrive in dozens of variations across hundreds of clients.

At month-end and year-end, firms commonly see 30% to 50% of annual staff effort compressed into roughly four weeks. Teams work longer hours, review quality gets harder to maintain, and the planned advisory meetings move into next month. That is expensive because advisory work often bills at two to three times the rate of compliance work.

For accounting and bookkeeping firms, we usually see annual margin leakage in the $60K to $180K range once repeat handling, avoidable review cycles, and missed advisory capacity are added up. Your number could be lower or higher. The point is to find it from your own workflow data, not accept a generic software ROI claim.

The AI audit for accounting and bookkeeping is designed to identify that work at the task level. You don’t need a transformation deck. You need a clear view of which jobs recur, where the exceptions sit, and where a controlled agent can earn its place.

Price the completed job, not the promise of automation

An agent workflow should be priced and managed around a completed task.

Take bank reconciliation. A completed job might mean:

1. Bank and transaction feeds are available.
2. Transactions are matched or categorised against an approved policy.
3. Unmatched items are identified with a reason.
4. Proposed adjustments are drafted where required.
5. A reviewer approves, rejects, or routes exceptions.
6. The reconciliation status and audit trail are recorded.

That is a measurable unit of work. “AI helps with reconciliations” is not.

Once the task is defined, you can calculate the cost properly:

• Agent execution cost, including model or token use
• Integration and data retrieval cost
• Staff time spent preparing the inputs
• Reviewer time for accepted work and exceptions
• Rework created by incorrect classifications
• Escalation time when the agent cannot proceed
• Ongoing workflow maintenance

This is where broad agent rollouts often disappoint. A low apparent cost per interaction can look attractive until the team spends 12 minutes reviewing the output, correcting two exceptions, and documenting why a client-specific rule applies. You have not reduced the cost of the job. You have moved it.

The target is not zero review. Accounting work has professional judgement, client context, and compliance implications. The target is a lower, predictable cost per completed job with a clear human control point.

A useful first dashboard is simple:

Measure	What it tells you
Tasks started	Demand entering the workflow
Tasks completed without escalation	How often the agent works within its boundary
Average reviewer minutes per task	The true human cost
Exception rate	Where the workflow needs redesign
Rework rate	Whether quality is holding
Cost per approved completion	The number that matters
Time released for advisory work	Commercial value created

Track those numbers for 30 to 60 days on a small client set. Don’t extrapolate from a polished demo.

Start with a bounded month-end task

Month-end is an obvious starting point because the process repeats, the volume is high, and the pain is visible. But “automate month-end” is still far too broad.

Break it into small units.

A firm might begin with bank-feed readiness and reconciliation preparation for a group of clients with consistent bookkeeping processes. The agent pulls approved bank, AP, AR, and payroll feeds. It checks data freshness, identifies missing sources, matches standard transactions using the firm’s categorisation rules, and creates an exception list.

The human bookkeeper still handles unclear coding, unusual transactions, material variances, and final sign-off.

That is the operating model for the Month-End Close Agent in Omni ops. It pulls bank, AP, AR, and payroll feeds, reconciles, flags variances, drafts journal entries, and prepares a partner-ready close pack. The important word is “prepares.” It does not become the partner of record. It gets the work to a point where professional review is faster and more focused.

A sensible first pilot could look like this:

Choose 10 to 20 suitable clients

Pick clients with stable source systems and a reasonably consistent chart of accounts. Avoid the difficult cases first. Do not use a pilot to prove the agent can survive chaos. Use it to prove the workflow has an economic case when the conditions are controlled.

Define the exception rules in writing

Set clear boundaries before the agent begins. For example:

• Escalate transactions above a defined materiality threshold
• Escalate new suppliers without a prior coding pattern
• Escalate payroll variances above an agreed percentage or dollar amount
• Never post a journal entry without reviewer approval
• Stop if a bank feed is more than a defined number of days behind
• Never change chart-of-accounts structure without an authorised owner

This is how you protect the firm. You are telling the agent where its authority ends.

Set the approval gate

A reviewer should see a compact queue, not a narrative dump. They need the proposed treatment, the supporting evidence, the confidence or rule basis, and a clear action: approve, edit, or escalate.

The approval gate needs to be part of the workflow design, not a manual safety check added after the fact. If the reviewer has to open six systems to validate every decision, the pilot is poorly designed.

Measure completed work, not agent activity

An agent can run thousands of steps and still create no value. Your measure is approved reconciliations completed, close packs prepared, or exception cases resolved within the agreed time.

If the completed-job cost falls and review time stays manageable, expand to the next client cohort. If exceptions climb, narrow the task or improve the source data. Don’t respond by giving the agent broader access.

If you want a working session on where that boundary should sit in your firm, Book a 60-min Omni Audit. We will map the task, the approvals, and the economics in the context of your actual client mix.

Onboarding is another strong candidate, with tighter controls

Client onboarding is often presented as an administrative issue. It is a commercial issue.

The client has signed, but billable work does not begin because the team is chasing access, bank feeds, historical statements, payroll files, prior-year reports, and chart-of-accounts answers. In many firms, 20% to 30% of new clients delay billable work by a quarter due to an incomplete or poorly managed onboarding process.

The Client Onboarding Agent creates a more disciplined path. It collects documents through a guided workflow, identifies what is missing, captures source-system details, helps set up the chart of accounts, and produces a clean opening trial balance for review.

Again, the right implementation is not autonomous end to end.

The agent can send reminders based on missing items. It can interpret uploaded documents against a checklist. It can flag a mismatch between the prior trial balance and supplied bank data. It can draft an opening balance structure based on the firm’s approved template.

It should not make final decisions about unusual historical balances, tax-sensitive treatment, or a client’s chart design without the right person approving it.

The best unit price for onboarding might be “onboarding pack completed and approved” rather than “number of documents processed.” A completed pack includes the required documents, system access, chart setup, opening-balance work, exception record, and partner or manager approval.

That definition makes your client promise sharper too. You can tell a new client exactly what you need, what happens next, and what will delay go-live.

For ideas on where operational AI fits across a service business, spend some time with the Omni platform. The key is to build from the operating work outward, not from a tool feature inward.

Review cost is the number most firms miss

Owners often ask how much an agent costs. The better question is, “How many human minutes does this completed task still require?”

A workflow that costs a few dollars to run may be uneconomic if every output needs ten minutes of senior review. On the other hand, an agent that needs a short bookkeeping review but clears the first pass on a large share of routine tasks may create a strong margin result.

Review cost rises when:

• Input data is inconsistent
• Client-specific rules are stored in people’s heads
• The output does not show supporting evidence
• Exceptions are not classified well
• Reviewers cannot correct the agent’s path efficiently
• The task boundary is too broad

This is why agentic AI does not behave like a simple licence model. More volume can expose more exceptions. More clients can mean more unique rules. More connected systems can create more failure points.

The answer is not to abandon scale. It is to earn scale through standardisation.

Document the rules. Standardise the inputs. Segment clients by workflow complexity. Use separate task types for clean, routine jobs and high-judgement work. Keep an exception library so the team can see which cases recur and decide whether they need a new rule, a different client process, or permanent human handling.

The Omni advisory approach is useful here because it connects operational workflow data to the firm’s commercial decisions. If review time is taking too long, you may need to change the process, adjust client fees, or move a client into a higher-touch service tier. AI should reveal those decisions, not hide them.

Turn released capacity into advisory revenue

Reducing bookkeeping handling time matters, but it is only half the opportunity. The real commercial gain comes when you use released capacity deliberately.

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. It gives the partner a starting brief, not a substitute for advice.

For example, it may identify that gross margin has moved for three consecutive months, debtor days are drifting beyond the client’s normal range, or payroll costs are rising faster than revenue. The partner applies judgement, checks the context, and leads the client conversation.

Without this kind of support, advisory is the work that gets pushed out by compliance deadlines. A partner intends to call 15 clients after month-end. The close takes two extra days. The calls become emails, then they disappear.

Price agent work by completed jobs and you can see how much capacity is genuinely being released. If the Month-End Close Agent reduces routine preparation and the reviewer queue is under control, protect those recovered hours. Allocate them to advisory calls, cash-flow reviews, pricing conversations, or client retention work.

Don’t assume the team will naturally find that time. Put it in the operating plan.

Use a practical close map before you buy more tools

If you want a worksheet to use with your team, the Month-End AI Close Map for Accounting Firms breaks the close into task stages, input requirements, exceptions, approvals, and measures. You can also access the direct close map download.

Use it in a 45-minute meeting with your bookkeeping lead, a manager, and one partner. Pick one close process. Mark every handoff. Circle the decisions that require professional judgement. Then identify the first task that has clear inputs, a repeatable output, and a manageable exception path.

That exercise often reveals that the best first use case is smaller than the firm expected. That is good news. Smaller pilots are easier to measure and easier to stop if the economics are wrong.

You can find more practical operating material in our AI resources and guides, but don’t mistake research for a deployment plan. Your firm needs its own task definitions and measures.

What a 60-minute Omni Audit gives you

A useful AI conversation should end with decisions, not a list of tools to investigate.

In a 60-minute Omni Audit, we work through three outputs:

1. A ranked workflow shortlist
   We identify the month-end, onboarding, or advisory tasks with the clearest near-term economic case.

2. A task-level agent design
   You see the inputs, actions, exception paths, human approval gates, and the definition of a completed job.

3. A value and measurement plan
   We estimate where the $60K to $180K leakage band may sit in your firm and what you need to track before scaling.

There is no deck to admire and no requirement to begin with a firm-wide AI program. The point is to leave with a bounded pilot that a partner can own.

You can see Omni for accounting and bookkeeping first if you want the vertical view. When you are ready to identify the first completed task worth automating, Book a 60-min Omni Audit.

The firms that get value from agents will not be the ones that hand broad client workflows to autonomous software first. They will be the ones that know their task costs, control their exception paths, and expand only after the completed-job economics hold up.