The pilot usually isn’t the problem

A lot of accounting firm owners can point to an AI pilot that looked promising for two weeks.

The tool could read invoices. It could draft client emails. It could categorise transactions with decent accuracy. A team member showed it in a partner meeting, everyone saw the potential, and the project got described as a priority.

Then month-end arrived.

The team went back to downloading files, chasing missing statements, reworking coded transactions, checking reconciliations, and building review packs the way they always have. The AI pilot sat beside the real process instead of becoming part of it.

That pattern is behind the concern raised in the recent Council Post on why agentic AI pilots fail to reach revenue workflows. In accounting and bookkeeping, the main blocker is rarely the model.

It’s the workflow.

An agent can’t reliably complete a job that nobody has properly defined. If each manager collects onboarding documents in a different order, if coding rules live in a senior bookkeeper’s head, or if a partner’s review process changes by client, the agent has no stable operating environment.

It has instructions, perhaps. It doesn’t have a process.

For firms between $1 million and $25 million in annual revenue, that distinction matters because small pockets of unstructured work add up quickly. We usually see annual leakage in the range of $60,000 to $180,000 across unbilled rework, delayed starts, partner review time, overtime, and advisory conversations that don’t happen. This isn’t one dramatic failure. It’s a few minutes lost in hundreds of recurring tasks.

The right order is straightforward. Fix and document the workflow first. Then apply an agent where the steps, data, owner, exception path, and review point are clear.

What an accounting workflow looks like before it is ready

Most firms don’t have no process. They have processes that work because experienced people compensate for the gaps.

That works until a key employee is away, client volume grows, or the firm tries to hand the work to an AI agent.

Take a standard monthly bookkeeping engagement. On the surface, the workflow is simple: get the data, code transactions, reconcile accounts, review results, send questions, close the month.

In reality, the work often looks more like this:

• A bookkeeper checks three inboxes and a client portal for missing documents.
• Bank feeds are connected for some clients but not others.
• Transactions are coded from memory, old notes, and prior-month entries.
• A manager notices a variance but isn’t sure if it is normal until they ask the client.
• The client gets a vague email with a long list of questions.
• The answers arrive piecemeal over six days.
• Review occurs late because the team is trying to finish five other files.
• The partner sees the results after the window for a useful advisory conversation has passed.

There may be good people at every point in that chain. The workflow is still unstable.

The same issue shows up in client onboarding. One team uses a detailed checklist. Another starts by emailing a document list. A third begins setting up the chart of accounts before records arrive. Historical clean-up gets scoped informally, and nobody clearly marks the moment when the client moves from onboarding to recurring service.

That lack of definition has commercial consequences. In many firms, 20% to 30% of new clients can delay billable recurring work by a quarter because information, setup decisions, and historical data take too long to resolve. The client is paying attention during the sale, then experiences friction during their first weeks of service. That’s not a good handover.

Before you purchase another agent licence, ask one practical question: could a capable new hire follow this workflow tomorrow without relying on the person who designed it?

If the answer is no, the workflow needs work first.

Standardise the work before you automate it

Standardisation doesn’t mean forcing every client into the same chart of accounts or ignoring professional judgement. It means identifying what repeats and defining where judgement must enter.

For each recurring workflow, document six things.

1. The trigger

What starts the work, and when?

For a month-end close, it might be the third business day after month end, when bank, AP, AR, payroll, and payment feeds should be available. For onboarding, it might be the signed engagement letter and first payment.

Avoid triggers like “when the manager is ready” or “after the client sends everything.” Those are conditions, not operating triggers.

2. The required inputs

List the systems, documents, and data fields required to complete the work.

For a close, that may include bank feeds, credit card feeds, payroll journals, accounts receivable reports, accounts payable reports, loan statements, prior-period reconciliations, and client documents. Specify where each item is expected to appear and who is responsible when it doesn’t.

This is where many agent pilots break. The team asks the agent to reconcile transactions without a dependable way to know whether all source data has arrived.

3. The standard decision rules

Document the common decisions that happen repeatedly.

How do you treat owner draws? What is the approval threshold for unusual expenses? Which vendor descriptions should be flagged rather than auto-coded? When does a transaction need supporting evidence? What variance requires a client question?

A rule does not need to be complex. “Flag any expense over $2,500 with no receipt attached” is useful. “Use judgement” is not something an agent or a junior team member can consistently apply.

4. The exception path

The strongest workflows make exceptions visible.

Set out what happens when a feed is unavailable, a client misses the document deadline, a reconciliation difference exceeds tolerance, or a coding rule conflicts with the current financial picture. Name the escalation owner and the response timeframe.

Without this, agents either stop too early or make a confident assumption that shouldn’t be made.

5. The review point

Partners often say they need to review everything because quality matters. That’s understandable, but it doesn’t tell the team what the review is for.

A useful review standard specifies the work that must be checked, the tolerances, and the items the reviewer should focus on. For example, the reviewer may approve all balance sheet reconciliations, investigate changes over a defined threshold, and review every new coding rule created that month.

That gives an agent a clean handoff. It can prepare the work and the evidence. The professional makes the decisions requiring accountability.

6. The output and client handoff

Define the finished product. Is it a closed file, a partner-ready close pack, a client question list, or a monthly management report with three advisory observations?

When the output is vague, teams keep working because nobody is sure the work is done.

You can find useful process design examples in our operations resources, but don’t make documentation an academic project. Pick the highest-volume workflow, map it on one page, test it with the people who do the work, and improve it in the next cycle.

Where AI agents fit once the process is clear

Once your workflow is stable, an AI agent becomes more than a chat interface. It can coordinate work across systems, follow rules, prepare outputs, and escalate exceptions to the right person.

The key word is coordinate. An agent isn’t there to replace accounting judgement. It is there to remove the repeated administrative handling that prevents your people from applying that judgement.

The Month-End Close Agent

The Month-End Close Agent in Omni ops pulls bank, AP, AR, and payroll feeds. It checks whether expected inputs are available, reconciles accounts, flags variances, drafts journal entries, and prepares a partner-ready close pack.

A well-designed implementation might run like this:

1. On the agreed close date, the agent checks data availability for each client.
2. It creates an exception list for missing feeds, statements, and client documents.
3. It applies approved coding rules to routine transactions and marks uncertain items for review.
4. It completes preliminary reconciliations against defined tolerances.
5. It compares balances and trends against the prior month and budget where available.
6. It drafts questions in plain language, grouped by the person who can answer them.
7. It assembles a close pack that shows completed work, unresolved items, proposed journals, and notable variances.
8. The manager or partner reviews the exceptions and approves the final decisions.

Notice what the agent is not doing. It isn’t guessing how your firm handles payroll corrections. It isn’t creating a new revenue recognition policy. It isn’t silently posting high-risk transactions because the workflow failed to identify a review point.

That is why process definition comes first.

For a visual worksheet to map those steps, use the Month-End AI Close Map for Accounting Firms. The direct downloadable close map gives your team a practical way to identify inputs, decision rules, exceptions, and reviewer handoffs before you configure anything.

If you’re unsure which workflow should be first, Book a 60-min Omni Audit. We will work through the workflow, the bottlenecks, and the most realistic point for an agent to enter.

Fix onboarding before it becomes a quiet churn risk

Client onboarding is an ideal place to apply this approach because it is repetitive, commercially important, and often fragmented across sales, admin, bookkeeping, and partners.

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.

That sounds simple until you look at the process underneath it.

A new client may need bank access, accounting platform access, payroll records, loan statements, prior financial reports, tax registrations, entity details, contact permissions, and answers to a dozen setup questions. If the firm sends one generic email with a long attachment list, clients delay. When they delay, staff follow up manually. Then setup begins with incomplete information, and the first monthly close becomes a clean-up job.

A stronger workflow starts with a client-ready checklist broken into stages. The client sees only what they need to do next. The internal team can see what has been received, what is missing, and what blocks the account from moving forward.

The agent can then:

• Send guided requests based on the engagement scope.
• Check submitted documents against the onboarding checklist.
• Identify missing or unreadable files.
• Extract core information for the setup team to verify.
• Draft the proposed chart of accounts from the client type and service model.
• Build an opening trial balance workpaper.
• Route unusual items, historical clean-up issues, and missing balances to a named owner.
• Create a clear handoff to the recurring bookkeeping team.

The result isn’t just faster setup. It is a better first experience for the client and a cleaner starting position for the delivery team.

For broader ideas on where this work sits across the firm, look through our AI insights library and the practical material in our guides. The principle remains the same in each case. Build a reliable workflow, then build the agent around it.

Protect advisory time by making the close usable

Compliance work expands to fill the available calendar. That is one reason advisory often remains a stated goal rather than a reliable revenue line.

The economics are clear. Advisory billable rates are commonly two to three times the rate of baseline compliance work. Yet many firms finish the close so late, or with so much unresolved detail, that the partner never gets to the client conversation that would create value.

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 can flag a margin movement, a cash conversion concern, a payroll increase, a customer concentration pattern, or a working capital issue. It can provide the supporting numbers and frame questions for the client.

It should not deliver generic commentary like “expenses increased this month.” Your workflow needs to define the client context, materiality thresholds, and the types of observations that matter for each segment.

For example, a trades client may need a prompt around job margin and overdue debtors. A professional services client may need a view of utilisation, contractor spend, and cash runway. The agent prepares the evidence. The partner turns it into advice.

That is the point of using AI in an accounting firm. Not to generate more activity, but to get the right people into higher-value conversations earlier.

Start with the workflow that has a financial owner

Don’t start with the loudest AI feature. Start with a recurring workflow where someone owns the commercial result.

For most firms, that will be month-end close, onboarding, or the preparation of advisory reviews. Each has clear volume, visible delays, and a measurable impact on delivery margin or growth.

Measure a small set of before-and-after indicators:

• Days from month end to completed close.
• Number of client chases per file.
• Manager and partner review hours.
• Rework after first review.
• Days from signed engagement to first billable recurring month.
• Number of advisory meetings completed with a prepared insight pack.

You don’t need perfect data to begin. You need a baseline good enough to tell whether the agent is reducing effort and improving throughput.

Then run one workflow through a complete cycle. Keep the scope narrow. Use actual files, actual exceptions, and actual review standards. If the process changes every week, stop configuring and fix the process.

You can see Omni for accounting and bookkeeping to understand how we assess this work. We don’t start with a deck of tool recommendations. We start with how your firm gets work done now.

Use the audit to find the real constraint

A 60-minute Omni Audit produces three useful outputs: a view of where leakage is occurring, a prioritised workflow opportunity, and a practical next-step plan for implementation. No deck. No vague innovation discussion.

For an accounting firm, that could mean identifying that month-end is delayed by missing-input follow-up, not reconciliation itself. Or that onboarding is losing margin because historical clean-up isn’t separated from standard setup. Or that advisory isn’t blocked by a lack of insight, but by close packs arriving too late for partners to use them.

Those distinctions save a lot of wasted technology spend.

AI agents can do meaningful work in accounting and bookkeeping firms. But they need a clear runway. Standardise the recurring workflow, define the exceptions, preserve professional review, and then let the agent handle the coordination that drains your team’s time.

If you’re ready to identify the first workflow worth fixing, Book a 60-min Omni Audit. You can also review the AI audit for accounting and bookkeeping before the call.