Planning an AI agent isn’t the same as deploying one

The numbers around agentic AI are getting a lot of attention. A report covered by The Tribune found that while almost every company is planning some form of agentic AI, only 9% to 14% have put agents into real production.

That gap matters more in accounting and bookkeeping than in most industries.

Your firm doesn’t need another experiment that produces a clever summary of a P&L. It needs work removed from the team’s calendar. It needs fewer nights spent chasing missing statements, fewer manual reconciliations, cleaner review files, and more partner time for conversations clients will actually pay for.

For a firm in the $1 million to $25 million revenue range, that difference has real financial weight. We regularly see annual operational leakage in the $60K to $180K range across firms of this size. It doesn’t usually show up as a single obvious line item. It shows up in overtime during close, write-downs on fixed-fee work, senior reviewers doing junior-level checking, delayed onboarding, and advisory meetings that never get scheduled.

The reason so many AI agent projects stall isn’t usually that the model is weak. It’s that the firm starts with a broad ambition like “automate bookkeeping” or “use AI for advisory.” Those ambitions are too vague to build, govern, or measure.

A production agent needs a defined job. It needs approved data sources, a handoff point, exception rules, and a number that tells you if it is working.

For most accounting and bookkeeping firms, month-end close is the right place to begin.

The accounting work that creates the deployment gap

Think about a normal close cycle for a managed accounting client.

A bookkeeper signs into the banking platform, checks whether feeds have refreshed, downloads a statement when the feed fails, and scans for uncategorised transactions. They look for AP bills in the accounting platform, compare open receivables against the prior month, check payroll postings, and send a client another reminder for the documents that still haven’t arrived.

Then come the reconciliations.

Some are routine. Some have timing differences. Some include unclear transfers, duplicate transactions, or merchant names that don’t match the chart of accounts. The bookkeeper makes calls, creates draft entries, leaves notes, and escalates the items that could change the management accounts.

A senior accountant or manager reviews the file. They may find an unreconciled balance, a payroll accrual that hasn’t reversed, an expense coded to the wrong account, or a variance large enough to require a client question. The partner sees the close pack after the review chain is complete, often when there is little time left to turn the numbers into a useful client conversation.

None of that is one task. It is a sequence of data gathering, checking, judgment, routing, drafting, and approval.

That is exactly why generic AI pilots fail. A staff member may use a chatbot to draft an email or suggest journal wording, but the underlying close process has not changed. The team still has to find the data, decide what is missing, move requests through email, and assemble the review pack.

The work remains fragmented.

At year-end, the problem compounds. Many firms see 30% to 50% of staff capacity squeezed into roughly four weeks of peak work. Even if the figure differs by service mix, the pattern is familiar. The highest-cost people end up managing exceptions and client follow-up because there is no reliable operating system around the process.

The deployment question is not, “Can AI reconcile a bank account?”

The better question is, “Can an agent take a defined set of close actions, document what it did, flag what it cannot resolve, and deliver a reviewer-ready output under our controls?”

That is a production use case.

Start with one workflow and a visible scorecard

The firms that get past the planning stage pick a workflow with four characteristics.

First, it occurs frequently enough to matter. Month-end close gives you 12 opportunities a year per recurring client to improve the process.

Second, the workflow has a known starting point and a clear finish. For example, the close begins when source feeds and documents are available. It finishes when a manager receives a close pack with reconciliations, open exceptions, draft journals, and material variance notes.

Third, the work contains repeatable decisions. This doesn’t mean every decision is automatic. It means the firm can describe what should happen in common situations and when a human must take over.

Fourth, success can be measured without argument.

For a Month-End Close Agent, your scorecard might include:

• Close cycle time from source-data availability to reviewer-ready pack
• Number of client follow-ups required to collect missing documents
• Percentage of reconciliations completed without manual data gathering
• Number and dollar value of exceptions escalated to a reviewer
• Review hours per client close
• Days between close completion and the client advisory conversation

These measures are more useful than asking staff if they “like AI.” They tell you whether the process changed.

A practical first target might be a narrow client group. Pick 15 to 30 monthly bookkeeping clients using the same accounting platform, similar bank feeds, and a relatively stable chart of accounts. Don’t begin with your most complex multi-entity client. Start with a repeatable segment where you can learn without putting a major account at risk.

This is also where the AI audit for accounting and bookkeeping helps. It identifies where the firm is losing time, which workflows are suitable for an agent, and what controls are needed before you switch anything on.

What a Month-End Close Agent does in practice

The Month-End Close Agent in Omni ops is not a replacement for professional judgment. It is a structured worker that handles the repeatable parts of the close and routes exceptions to the right person.

Here is what that can look like from start to finish.

1. It checks source readiness

At the start of the close window, the agent checks the bank, AP, AR, payroll, and accounting feeds. It confirms the period being closed and records which feeds are current.

If a bank feed is delayed or a payroll report is missing, the agent doesn’t carry on quietly and create a bad file. It logs the gap and triggers the appropriate request. For a client, that could be a guided message asking for a statement or report. For an internal team member, it could be a task with the client name, period, and exact item required.

This is a small point, but it changes how the team works. Instead of someone discovering a missing document halfway through a reconciliation, the missing input is made visible at the beginning.

2. It prepares routine reconciliations

The agent pulls available transactions and compares them with the ledger, AP, AR, and payroll records. It identifies familiar transaction patterns, proposed matches, and items that don’t fit the rules.

It can prepare reconciliations for review and classify exceptions. A transfer that follows an established pattern may be marked as a proposed match. An unusual payment to a new supplier might be held for review. A large variance from the prior month can be tagged for an explanation.

The agent should never be allowed to invent certainty. Its role is to prepare, compare, and surface evidence. Your team defines the thresholds. A $200 coding difference for one client may be immaterial. A $200 difference in a high-volume, low-margin account could matter. That is firm policy, not an AI decision.

3. It drafts journals and variance notes

When the close rules call for accruals, prepayments, depreciation, payroll adjustments, or recurring entries, the agent can draft the journal entry with supporting detail.

“Draft” is the operative word. The journal should include a clear source, rationale, account mapping, period, and confidence or exception status. A designated reviewer approves it before posting when your control framework requires approval.

The agent also compares current results against prior periods and expected ranges. It may flag that revenue is down 18% month on month, that wages as a percentage of revenue have shifted, or that receivables are aging faster than usual.

That doesn’t replace the manager’s interpretation. It gives the manager a starting point with the underlying numbers already assembled.

4. It creates a partner-ready close pack

Once routine checks are complete, the agent creates a structured close pack. It includes reconciliations, proposed journals, unresolved items, material variances, client questions, and an audit trail of actions taken.

The manager doesn’t receive a pile of exported reports. They receive a work queue.

They can approve low-risk items, challenge exceptions, add professional context, and focus their time where it matters. The partner receives a cleaner view of what changed and what should be discussed with the client.

That is how the agent becomes operational. It has a defined input, a defined scope, a review pathway, and a defined output.

If you want to see how the operating layer fits around these workflows, look at Omni ops. The point isn’t to add another dashboard. It is to make work visible, assigned, and controlled across the team.

The controls that make an agent usable in a firm

A common reason leaders hesitate is valid. Financial data is sensitive, client files require care, and an incorrectly posted entry can cause real problems.

The answer is not to avoid agents. It is to build the controls before deployment.

Set permissions by role. A close agent may read specified ledgers, banking feeds, AP, AR, and payroll data. It should not have unrestricted access to every client system just because that is technically easier.

Create action tiers. For example:

• The agent can collect data and prepare work without approval
• The agent can send approved reminder templates under defined conditions
• The agent can draft journals but cannot post them
• A manager must approve all entries above a materiality threshold
• The partner must approve exceptions involving tax treatment, related parties, unusual revenue recognition, or a significant client issue

Keep an audit trail. Your team should be able to see which sources were used, which rule or instruction was applied, what the agent proposed, and who approved the final action.

Run parallel close cycles before relying on the agent. For the first one or two months, compare the agent-supported output with your usual process. Track where it catches issues, where it creates noise, and where your instructions need refinement.

This is not bureaucracy. It is how a firm moves from a demo to reliable production work.

You can Book a 60-min Omni Audit if you want to map that first workflow properly. In 60 minutes, we focus on three outputs: the leakage points in your current process, the agent workflow with the strongest business case, and the practical deployment path. No deck. No generic AI roadmap.

Don’t ignore onboarding and advisory

Month-end close is often the best first deployment, but it should connect to the rest of the client lifecycle.

The Client Onboarding Agent can collect documents through a guided workflow, track what is missing, set up the chart of accounts from approved templates, and produce a clean opening trial balance for review.

This matters because onboarding delays often poison the economics of a new relationship. We see firms where 20% to 30% of new clients delay billable work by a quarter because documents arrive late, historical records need cleaning, or setup moves between several staff members without a clear owner.

An onboarding agent won’t turn a disorganised client into a perfect client. It will make the missing information visible, prompt consistently, and stop the team from rebuilding the same checklist in email for every engagement.

Then there is advisory.

The Advisory Insights Agent reads a client’s monthly numbers, identifies three points worth discussing, and drafts partner talking points before the meeting. It might highlight margin movement, cash conversion pressure, overdue receivables, rising payroll cost, or a change in monthly revenue trend.

The partner still owns the conversation. That is the value clients pay for. But when compliance work consumes the calendar, those conversations are often skipped. Advisory billing rates are commonly two to three times compliance rates, depending on the firm and engagement. Freeing even a modest block of partner and manager capacity can change the mix of your revenue.

You can see how this client-facing layer works through Omni advisory. It is not about sending automated financial advice without oversight. It is about making sure the right insights reach a qualified adviser before the client meeting.

Use the close map before you choose technology

You do not need to buy a large software stack to begin. You do need a clear map of the work.

The Month-End AI Close Map for Accounting Firms is a practical worksheet for doing that. It helps you list the close stages, systems, recurring inputs, manual checks, exception types, approval points, and measures that define success.

If you prefer to work directly from the file, you can access the direct close map worksheet.

Work through the map with a manager who knows the reality of the close, not just the ideal process written in a procedure document. Ask where staff are copying data, where client chasers happen, where work waits for review, and where senior people are routinely doing work that should be prepared earlier.

You may also find useful implementation thinking in our AI guides. The important thing is to keep the first decision grounded in workflow economics, not product features.

Put production ahead of experimentation

The 9% to 14% production figure should not make you pessimistic about agentic AI. It should make you more disciplined.

Most firms will have access to similar models and similar software. The difference will come from who turns a high-friction workflow into an operating process with real ownership, clean inputs, clear controls, and measurable outcomes.

For an accounting firm, that could mean a Month-End Close Agent that collects the right data, prepares reconciliations, drafts journals, flags exceptions, and delivers a reviewer-ready close pack. It could then lead to a Client Onboarding Agent that gets new work started faster, and an Advisory Insights Agent that protects time for higher-value conversations.

Start with a workflow you can measure. Prove the outcome on a controlled client group. Improve the rules. Then expand.

If you want an outside view of where the $60K to $180K leakage is sitting in your firm, see Omni for accounting and bookkeeping. When you’re ready to turn that view into an action plan, Book a 60-min Omni Audit.