A recent VentureBeat piece made a point that’s easy to miss in all the AI hype. When an agent gets something wrong, it’s rarely because it hallucinated a number out of thin air. It’s because nobody told it where its job ends and a human’s job begins. The agent did exactly what it was built to do. It just wasn’t supposed to do that much.

For a bookkeeping or tax firm, that’s not an abstract risk. That’s a journal entry posted without review. A client transaction categorized wrong three months in a row before anyone notices. A tax position taken on autopilot that should have gone to a partner first. The technology to automate a huge share of compliance work already exists and works well. The failure mode isn’t the AI being bad at its job. It’s firms skipping the step where they decide what the job actually includes.

The real question isn’t “does it work”

Every firm owner we talk to eventually asks the same thing: is this AI thing actually accurate enough for my clients’ books? Fair question, and the honest answer is yes, for a large chunk of the work. Bank feed reconciliation, recurring vendor categorization, payroll journal entries, accrual reversals — this is pattern-matching work that a well-built agent handles with more consistency than a tired staff member during week three of close.

But accuracy was never really the risk. Authority was. A model can categorize a transaction with 98% confidence and still be the wrong entity to make that call, because the transaction touches a related-party loan, an unusual write-off, or a client’s tax election that needs a human signature. The agent isn’t wrong about the number. It’s wrong about whose decision it was to make.

This is why the firms getting real value out of AI right now aren’t the ones who deployed the fastest. They’re the ones who sat down first and wrote out, transaction type by transaction type, what the agent can post on its own and what has to stop at a reviewer’s desk. That list is boring to write and it’s the single most important document in the whole rollout.

Where this shows up in your calendar

Three pressure points make this urgent for firms in the $1M-$25M range, and you’ve probably felt all three this year.

Month-end and year-end crunch. For a lot of firms, 30-50% of total staff hours get burned in about four weeks of the year. Everyone’s doing the same reconciliation work at the same time, under the same deadline, which is exactly the environment where a rushed staff member either over-relies on last month’s coding or skips a review step to hit the date. That’s not a training problem. That’s a workload design problem, and it’s the same problem that makes unclear AI authority dangerous — tired humans and unbounded agents make the same kind of mistake.

Client onboarding drag. New client setup, document collection, chart-of-accounts build, historical cleanup, this routinely stretches into weeks. We typically see 20-30% of new clients pushed a full quarter before they generate real billable work, and some churn before they ever get there. Onboarding is also where authority questions get murkiest, because the agent is working with a client’s history it’s never seen before and has no baseline to sanity-check against.

Advisory time crowded out. Compliance work fills the calendar first because it has hard deadlines. Advisory conversations, the ones billing at 2-3x the compliance rate, get pushed to “next quarter” and often just don’t happen. Every hour a partner spends re-checking an agent’s categorization is an hour that isn’t going toward a client conversation that actually grows the relationship.

None of these get solved by adding more AI. They get solved by defining boundaries clearly enough that AI can absorb the repetitive 80% and route the judgment-heavy 20% to the right person, every time, without a partner having to babysit the process.

What a properly bounded agent actually looks like

This isn’t theoretical. Here’s how it plays out with the systems we build inside firms.

Our Month-End Close Agent pulls the bank, AP, AR, and payroll feeds, reconciles them, and flags variances. It drafts the journal entries. Where it has clear authority — recurring vendor codes, matched bank transactions, standard accruals that follow last month’s pattern — it posts them. Where a variance falls outside a defined threshold, or a transaction type has never been coded before, it stops and routes to a reviewer with the entry drafted and the reasoning attached. The partner isn’t starting from zero. They’re approving or correcting, which is a five-minute task instead of a two-hour one. The close pack that lands on the partner’s desk is ready to review, not ready to build from scratch.

Our Client Onboarding Agent runs the document collection workflow, builds the chart of accounts, and produces an opening trial balance. Its authority is scoped tightly to setup and structure. It doesn’t make judgment calls about historical entries that look unusual. Those get flagged for the onboarding lead to review before the client’s books go live. That one boundary is often the difference between an onboarding that takes ten days and one that drags a client through a quarter of delay.

Our Advisory Insights Agent reads each client’s monthly numbers and surfaces three things worth discussing, then drafts the partner’s talking points before the meeting. Its authority is explicitly limited to surfacing and drafting. It never sends anything to a client directly and it never represents its output as final advice. The partner reads it, edits it, and owns what gets said in the room. That boundary isn’t a limitation. It’s what makes the output trustworthy enough for a partner to actually use it.

In each case, the design question came first: what can this agent decide alone, and what does it have to escalate. The build came second. Firms that skip straight to the build end up with an agent that’s technically accurate and organizationally reckless, which is the exact trap the VentureBeat piece is describing.

Building the authority map for your firm

If you’re thinking about deploying agents into your close process or your tax prep workflow, here’s the exercise worth doing before you write a single prompt or sign a single vendor contract.

Start by listing your transaction categories, not your software features. Recurring AP, one-off AP, payroll, intercompany, accruals, related-party transactions, unusual write-offs, tax elections. For each category, answer three questions. How often does it repeat in a predictable pattern? What’s the dollar threshold above which a mistake actually hurts? And who currently makes the call — a staff bookkeeper, a senior, or a partner?

Categories that are high-repetition, low-threshold-risk, and currently handled by junior staff are your first candidates for full agent authority. Categories that are low-repetition, high-dollar, or already routed to a partner should stay routed to a human, with the agent doing the prep work and drafting but never the posting.

Write this down as an actual document, not a shared understanding. “The team knows what needs review” is exactly the kind of assumption that breaks down during the crunch weeks, when the team doing the reviewing changes week to week. A written authority map is what lets you onboard a new senior in a week instead of a month, because the boundaries live in a document instead of in one person’s head.

This is also the exact worksheet we built into the Month-End AI Close Map for Accounting Firms, a practical checklist for mapping your close process transaction by transaction before you hand any of it to an agent. It’s not a sales deck. It’s the same exercise we run with firms in the first session of an audit, laid out so you can start it yourself. You can grab the direct version here if you want to work through it this week.

The dollar reality

For a firm doing $1M-$25M in revenue, unclear authority boundaries aren’t a compliance footnote. They cost real money in three places at once, and industry ranges put the total leakage for firms this size somewhere between $60,000 and $180,000 a year.

Some of that shows up as the overtime and temp staffing firms pay to survive month-end crunch. Some of it shows up as the billable work that never gets invoiced while a new client sits half-onboarded for a quarter. And a good chunk of it is invisible until you calculate it directly: the advisory revenue that never gets billed because partners are stuck reviewing routine entries that a properly bounded agent could have handled without ever needing a second look.

Firms that skip the authority-mapping step and deploy AI anyway tend to swing to the opposite extreme after the first bad experience. One over-coded entry gets missed, a partner panics, and the firm pulls the agent back to reviewing everything, which erases most of the time savings that justified the investment in the first place. Getting the boundaries right the first time is what keeps the ROI intact.

Where the Omni Audit fits

We built the Omni Audit specifically for this decision point. It’s a 60-minute session, no deck, and it ends with three concrete outputs: a map of where your firm’s manual hours are actually going, a short list of which of those hours are safe to hand to an agent with clear authority limits, and a rough dollar estimate of what that shift is worth to your firm specifically.

We’re not selling a generic AI platform. We’re scoping the exact same authority questions this article walks through, against your actual close process and your actual client mix. If you want to see how this looks for a firm your size, see Omni for accounting and bookkeeping or read through our broader guides on how the ops layer handles handoffs between agents and staff.

The conversation usually starts with the Month-End Close Agent because it’s the fastest place to prove the model works with real boundaries in place. From there it extends into onboarding and advisory work, once the firm has confidence in how escalation actually behaves in practice, not just on paper. You can read more on how we scope this across the ops layer or browse recent write-ups in our insights section on how other firms have sequenced the rollout.

Book the audit before you book the vendor

The mistake we see most often isn’t firms moving too slowly on AI. It’s firms signing a vendor contract before they’ve written down where the authority limits go, then discovering the gaps during their busiest month of the year. That’s the expensive way to learn this lesson.

The cheaper way is a 60-minute conversation before anything gets built. Book a 60-min Omni Audit and bring your close checklist. We’ll walk through it together and tell you exactly where an agent can carry real authority today, and where it should be drafting for a reviewer instead.

If you’re still early in figuring out what this looks like for your firm, start with the set-authority-limits-before-deploying-agents audit page or the blog for more detail on how other firms have drawn these lines. But the fastest way to get a real answer, specific to your books and your team, is still the direct one. Book my Omni Audit and we’ll map the boundaries before you deploy anything.