Month-End Close Automation With AI Agents, Step by Step

By Jude Lee · · Workflow

Two accountants reviewing trial balance and reconciliation screens during month-end close in a modern firm office

The gap between wanting to automate the close and actually doing it

There is no shortage of interest in automation in this profession, and a much shorter list of firms running something in production. Our read on why: “automate the close” is not one project. It’s fifteen small ones, most of which involve professional judgment, and a few of which are pure clerical drudgery that has been waiting years for a machine. Sorting those apart is the actual work, and it’s unglamorous enough that it rarely gets done before a tool gets bought.

An “agent” is a specific thing, and it isn’t a chatbot

Honest definition: an AI agent is a model given a goal, a set of tools it can call (read the GL, query a bank feed, write a draft memo, open a task), and permission to take multiple steps toward that goal without a human prompting each one. A chatbot answers; an agent acts.

That distinction matters in a close because most close pain is multi-step. “Pull the trial balance, compare to prior month, flag any account moving more than the threshold, pull the underlying transactions for the top five variances, and draft commentary” is five tool calls and a write-up. A chat interface makes you do each step. An agent does the sequence and hands you the output for review.

It is also where the risk lives. An agent that can write to your ledger is a different governance object than one that can only read it.

Our editorial position: the safest agentic close is one where the agent has read access to everything and write access to nothing except a draft document.

Split the close into three buckets before you automate anything

Before evaluating a single tool, take your existing close checklist for one representative client and mark every line as mechanical, judgment-assisted, or judgment-only.

Good candidates for an agent to draft
  • Categorizing recurring, high-confidence transactions against prior-period treatment
  • Producing the bank reconciliation exception list (unmatched items, stale outstanding checks, duplicate-looking entries)
  • Checking that recurring accruals, prepaids, and depreciation schedules actually posted, and at the expected amounts
  • Cross-footing the trial balance and flagging accounts with no activity that normally have activity
  • Drafting first-pass flux commentary tied to specific transactions
  • Assembling the open-items list and chasing client responses on document requests
Keep with a human
  • Materiality thresholds and what counts as an explainable variance
  • Any estimate: reserves, allowances, percentage-of-completion, accrual sizing
  • Approving and posting journal entries
  • Revenue recognition judgment calls under the applicable FASB standards
  • Anything that goes to the client as a conclusion rather than a question
  • The final sign-off, always attributed to a named person

The design principle we’d argue for: an agent should surface issues, not silently resolve them. An agent that says “these 14 items look wrong, here’s why” is enormously useful. An agent that quietly reclassifies 14 items is a review nightmare you won’t discover until an audit or a tax return doesn’t tie.

How an AI assistant actually reaches your general ledger

This is where MCP comes in. The Model Context Protocol is an open standard for connecting an AI assistant to external data and tools through a defined server — a consistent way to say “this assistant may call these specific functions on this specific system, with these credentials.” It is a protocol, not a product, which means it isn’t a lock-in decision the way choosing a platform is.

In practice, three things matter more than the protocol itself:

  1. Scope. Give the connection read-only access to the entities the workflow needs — trial balance, chart of accounts, transactions for a date range — and nothing else. Not payroll detail, not client PII you don’t need for the close.
  2. Per-client isolation. One connection per client file, so an agent working on Client A structurally cannot read Client B.
  3. Audit logging. Every call the agent makes should be recorded: which account, which date range, which user session initiated it. If you can’t answer “what did the agent look at on the 4th?” you don’t have a controlled workflow.

A skill is how you stop re-explaining your close every month

A skill is a packaged, reusable set of instructions that teaches an assistant to perform one job your firm’s way, every time. For the close, the obvious candidates are:

The practical value of a skill is consistency across staff. A senior and a first-year get the same output structure, which makes review faster — which is where a lot of the real close time goes.

Standing this up without betting the firm on it

  1. Pick one client and one month you've already closed

    Run the agent against a closed period where you know the right answer. Compare its exception list to what your team actually found. This is your only honest accuracy test.

  2. Start read-only, output to a document

    No write access to the ledger. The agent’s deliverable is a draft memo or a spreadsheet of proposed items, nothing posted.

  3. Write the skill from your existing checklist

    Don’t invent a new close process. Encode the one your best reviewer already runs, including thresholds.

  4. Instrument the review step

    Track, for a few cycles, how many agent-proposed items your reviewer accepts, edits, or rejects. Rejection rate by category is the number everything else depends on: it tells you which parts to keep automating, which to pull back, and what time saving you can honestly claim.

  5. Expand by workflow, not by client count

    Get flux commentary right across ten clients before adding categorization. Widening one proven step is much safer than running five half-working ones.

Where this breaks

Agents produce confident, well-written explanations for variances they don’t actually understand — a plausible narrative attached to the wrong driver is worse than no narrative, because it survives a quick review. They also degrade on messy source data: inconsistent vendor naming, personal expenses run through the business, mid-period chart of accounts changes.

And sometimes the right answer isn’t AI at all. If your recurring categorization problem is 40 transactions a month from the same 12 vendors, a deterministic rule in your accounting software will beat an LLM on cost, speed, and auditability. Rules are better than models whenever the logic is genuinely fixed.

Buying the feature versus building the workflow

The major ledger and mid-market finance platforms are shipping AI assistants into products you already pay for; check your vendor’s current release notes rather than a secondhand summary. Off-the-shelf features have a real advantage: the data stays within your existing vendor agreement, the vendor owns the security posture, and there’s nothing for you to maintain. Confirm the vendor’s subprocessor list before you assume the data stays put — most in-product AI routes inference to a third-party model provider, and that provider belongs in your due diligence too.

Where in-product AI falls short for firms is standardization across clients on different platforms. If you close 60 clients across QuickBooks Online, Xero and Zoho Books, in-product AI gives you three different behaviors and no shared workpaper standard. That’s the case for a custom build: one skill, one review format, one audit log, applied across ledgers. It carries a real recurring cost, though — a custom MCP server sits on top of vendor APIs that change without asking you, and someone at the firm has to own breakage, re-auth, and scope changes on an ongoing basis. Name that person before you start; if it’s your one technical partner during busy season, that’s an argument against. If you’re a single-platform shop, the in-product features may get you most of the way there for a fraction of the effort — a legitimate answer, not a cop-out.

What it’s actually worth — your numbers, not a benchmark

Don’t accept a percentage you can’t reproduce. Build the model yourself:

C × H × R
Clients closed per month × hours per close × loaded hourly rate = current close cost
Illustrative formula — plug in your own numbers
Your Step 4 accept rate
Derive any time-reduction estimate from the accept/edit/reject rates you measured in your own pilot — it is an output of measurement, not an input you assume
Model input you generate
+ review time
Add the reviewer minutes the agent's output creates or removes — this often decides the outcome
Model input

Then add the parts firms forget. Recovered senior hours only pay off if they’re reallocated to advisory or additional client capacity — otherwise you’ve bought slack, not margin. Count avoided rework: a categorization error found in month two costs far less than one found at tax time. And count the close-calendar compression itself; getting statements to clients three days earlier has commercial value even when the hours are unchanged.

The governance layer you need regardless

Everything above involves taxpayer and client financial data. The IRS’s Publication 4557, Safeguarding Taxpayer Data, sets out written information security plan obligations for tax professionals, and any AI vendor or connection touching that data belongs in your WISP and your vendor due diligence. The AICPA’s professional standards on confidentiality apply to an AI tool exactly as they apply to a subcontractor.

Practical minimums before a single client file goes near an agent: confirm the provider’s data retention and training policy in writing, check whether your engagement letters and client consents cover third-party processing, keep per-client access isolated, and retain logs. Where a specific arrangement has real compliance stakes — cross-border data handling, audit clients, anything touching an attest engagement — confirm the treatment with qualified counsel or your firm’s risk partner and your state board of accountancy before you scale it.

The firms that get this right won’t be the ones that automated the most steps. They’ll be the ones who can explain, to a reviewer or a regulator, exactly what the machine did and exactly which human agreed with it.

Not sure where to start?

Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.

Get a free automation audit

Related guides