Amazon’s director of artificial general intelligence research told an industry conference last month that reliability, not capability, is what’s blocking enterprise AI deployment. The models can do the work. They just can’t do it consistently enough to ship without a human checking every output.

For accounting firms, that gap between “can do” and “can trust” is the difference between a tool that saves time and a liability that costs your license.

Right now, AI can draft a research memo on a complex depreciation question. It can pull relevant code sections, summarize case law, and give you a starting point in 90 seconds instead of two hours. But it can’t file the return. It can’t sign the 1040. It can’t attest to the accuracy of Schedule C when the model hallucinates a deduction or misreads a threshold.

The error rate for frontier language models on structured tasks sits somewhere between 2% and 8%, depending on the complexity of the prompt and the specificity of the domain. Tax compliance work requires an error rate below 1% before you can remove the CPA from the loop. We’re not there yet.

That doesn’t mean AI has no place in your practice. It means you need to draw a bright line between the work AI can handle today and the work that still needs a licensed professional’s signature.

Where AI Breaks Down in Tax Work

The failure modes aren’t random. They cluster around three types of tasks.

First, anything that requires judgment about materiality or risk. A model can tell you the rule. It can’t tell you whether applying that rule in this specific fact pattern will survive an audit. One of the managing partners we work with described it this way: “The AI gives me the technically correct answer, but it doesn’t know my client’s audit history or the examiner’s tendencies in this district.”

Second, multi-step workflows where an error in step three invalidates everything downstream. Tax return preparation is a dependency graph. If the model miscategorizes a transaction in January, that error propagates through basis calculations, carryforwards, and estimated payments for the rest of the year. A 3% error rate per step compounds into a 15% error rate over five steps.

Third, any task where the source data is messy or ambiguous. AI models trained on clean datasets struggle when the bank feed has duplicate entries, the client’s spreadsheet mixes personal and business expenses, or the prior-year return was prepared by someone who didn’t document their assumptions. Garbage in, garbage out still applies.

These aren’t edge cases. They’re the center of the distribution for most tax engagements.

What AI Can Do Right Now

The work AI handles well today falls into three buckets: research, draft preparation, and data transformation.

Research is the obvious win. A well-prompted model can pull relevant code sections, summarize PLRs, and compare your fact pattern to analogous cases in minutes. It won’t catch every nuance, but it gives you a head start that used to require an hour of manual searching. You still read the output. You still verify the citations. But the grunt work is done.

Draft preparation is the next tier. AI can take a set of facts and produce a first-pass memo, a client letter, or a disclosure statement. The structure is there. The language is serviceable. You’re editing instead of writing from scratch. For routine correspondence, that’s a 60% time savings. For complex technical writing, it’s closer to 30%, but 30% of four hours is still meaningful.

Data transformation is where the reliability picture gets better. Taking a CSV export from a client’s accounting system and mapping it to your firm’s standard chart of accounts is a structured task with clear rules. The model either maps the account correctly or it doesn’t. You can validate the output with a reconciliation check. If the totals tie, the mapping worked. If they don’t, you catch it before anything goes downstream.

These three use cases share a common feature: they all have a human review step baked into the workflow. The AI output is never the final output. It’s an input to the next stage of work, where a CPA applies judgment, checks the math, and signs off.

That’s the pattern that works today. AI as a drafting tool, not a decision-making tool.

The Reliability Threshold for Compliance Work

The reason tax preparation and filing can’t be automated yet comes down to acceptable error rates and the cost of a mistake.

In advisory work, a small error in a projection or a scenario model is annoying but not catastrophic. You catch it in the next review cycle. You update the numbers. The client adjusts their decision. No one loses their license.

In compliance work, an error that makes it onto a filed return triggers penalties, interest, and potential malpractice claims. The IRS doesn’t care that your AI tool had a 97% accuracy rate. They care that line 43 on Schedule C is wrong and you signed the return.

The math is straightforward. A typical tax return for a small business has 200 to 400 data points that feed into the final numbers. If your AI tool has a 3% error rate, you’re looking at six to twelve mistakes per return. Even if half of those are immaterial, you’re still spending more time finding and fixing errors than you would have spent preparing the return manually.

The reliability threshold for removing human review sits somewhere below 1%. At that error rate, you might see one mistake every three or four returns, and the time saved by automation outweighs the time spent on quality control. We’re not there yet with general-purpose language models, and we won’t be for another 18 to 24 months at the current pace of improvement.

Domain-specific models trained exclusively on tax code and return data will get there faster, but those tools are still in development. The models you can access today through API or SaaS platforms are general-purpose. They’re good at language. They’re not yet good enough at tax.

What This Means for Your Practice Today

If you’re running a firm that does $2M to $8M in annual revenue, you’re probably carrying 300 to 800 compliance clients and trying to build an advisory practice on top of that base. The compliance work pays the rent. The advisory work drives margin. But the compliance work eats all the time.

AI won’t let you automate the compliance work yet, but it can compress the time you spend on the low-value parts of that work. Research, drafting, data prep, and client communication are all candidates for AI assistance today, as long as you keep the review step in place.

The firms we work with typically see 20% to 35% time savings on research-heavy engagements when they start using AI for memo drafting and code lookups. That’s not enough to cut headcount, but it’s enough to free up 10 to 15 hours per week for a senior associate. Those hours can go toward advisory calls, practice development, or just reducing the weekend work during busy season.

For partners, the bigger opportunity is in client communication. Drafting the year-end tax planning letter, summarizing the return for a client who doesn’t read tax forms, and explaining a complex adjustment in plain language are all tasks where AI can produce a solid first draft. You’re still the one who reviews it and adds the personal touch, but you’re not starting from a blank page.

One accounting firm owner in our network described the shift this way: “I used to spend two hours writing a detailed planning memo for a client. Now I spend 20 minutes prompting the AI, 30 minutes editing the output, and 10 minutes adding my specific recommendations. The client gets the same quality memo, and I get 60 minutes back.”

That’s the pattern that works today. AI handles the scaffolding. You handle the judgment.

If you want to see where AI can fit into your month-end close process without touching the final sign-off, we built a worksheet that maps the typical close workflow for accounting firms and flags the steps where AI can assist. You can grab the Month-End AI Close Map and use it to identify the 10 to 15 hours per month that are currently spent on data prep and reconciliation grunt work.

The Agents We Build for Accounting Firms

When we build AI systems for accounting practices through the AI audit for accounting and bookkeeping, we don’t start with tax preparation. We start with the work that sits upstream of the return, where the reliability threshold is lower and the risk of a mistake is containable.

The Client Onboarding Agent is a good example. New client intake is a multi-week process that involves document collection, chart-of-accounts setup, and historical clean-up. Most firms lose 20% to 30% of new clients during onboarding because the process drags and the client gets frustrated.

The agent we build handles the document request workflow, validates that the files are complete, maps the client’s existing accounts to your firm’s standard chart, and produces a clean opening trial balance. It doesn’t make judgment calls about how to classify ambiguous accounts. It flags those for your review. But it handles the 80% of accounts that are straightforward, and it does it in hours instead of weeks.

The Month-End Close Agent is another high-value target. Close work is predictable but tedious. Pull the bank feeds. Reconcile AP and AR. Check for uncleared transactions. Draft the standard journal entries. Prepare the close pack for partner review.

An AI agent can handle all of that except the final sign-off. It pulls the data, runs the reconciliations, flags variances that exceed your threshold, and drafts the entries. You review the pack, adjust anything that needs judgment, and close the month. The time savings range from 40% to 60%, depending on how standardized your close process is.

The Advisory Insights Agent is where things get interesting for margin. Most firms bill compliance work at $150 to $200 per hour and advisory work at $300 to $500 per hour. The problem is that compliance work crowds out the advisory calendar. You finish the returns in April, and by the time you have bandwidth for advisory calls, it’s June and the client has already made their decisions.

The agent we build reads each client’s monthly financials, surfaces three things worth discussing, and drafts the talking points for the partner. It doesn’t make recommendations. It identifies the questions you should be asking. That 15-minute prep used to take an hour, and it often didn’t happen because the hour wasn’t available.

When you compress prep time from 60 minutes to 15, you can fit two more advisory calls into the week. Over a quarter, that’s 24 additional conversations. If half of those turn into advisory engagements, you’ve added $40K to $80K in high-margin revenue without hiring anyone.

These agents don’t touch the tax return. They don’t sign anything. They don’t make judgment calls. They handle the structured, repeatable work that sits around the compliance core, and they free up your time for the work that actually requires a CPA license.

You can see the full breakdown of what we build for accounting practices at See Omni for accounting and bookkeeping. The audit itself takes 60 minutes. We look at your current workflow, identify the 10 to 20 hours per week that are candidates for AI assistance, and map out what the agent would do step by step.

The Economics of Partial Automation

If you’re running a $3M practice with four CPAs and two bookkeepers, your fully loaded labor cost is probably $650K to $850K per year. That’s 55% to 70% of revenue, which is typical for firms of this size.

A 25% reduction in time spent on research, drafting, and data prep translates to roughly 10 hours per week per senior staff member. Over a year, that’s 500 hours per person, or 2,000 hours across four CPAs. At a $200 blended billing rate, that’s $400K in capacity that’s currently absorbed by low-value work.

You can’t bill all of that capacity. Some of it goes to reducing weekend work and overtime. Some of it goes to professional development. But if you convert even 30% of it into billable advisory work at a higher rate, you’re looking at $120K to $180K in additional revenue with no headcount increase.

The flip side is that partial automation doesn’t reduce your cost base much. You’re not cutting staff. You’re not shrinking your office. You’re reallocating time from low-margin work to high-margin work. The ROI comes from revenue growth, not cost reduction.

For most firms, that’s the right trade. The constraint isn’t labor cost. It’s partner time and the ability to serve more clients without burning out the team. AI that handles the grunt work and frees up 10 hours per week per person solves that constraint.

The firms that wait for full automation are going to be waiting another two to three years. The firms that deploy partial automation today are building the advisory practices that will carry them through the next decade.

What Happens in the Omni Audit

When you book a 60-min Omni Audit, we don’t show you a deck. We don’t pitch a platform. We walk through your current workflow and build the map for where AI fits.

The first 20 minutes are discovery. We ask about your client mix, your service lines, and the work that takes the most time. We want to know where the bottlenecks are. Is it month-end close? Is it client onboarding? Is it the research and memo writing that happens before you ever touch the return?

The next 20 minutes are design. We pick one workflow and map it step by step. What data comes in? What happens to it? Who touches it? Where does it go? What’s the output? We identify the steps that are structured enough for AI to handle and the steps that require judgment.

The last 20 minutes are scoping. We show you what the agent would do, what the human review step looks like, and what the time savings would be. We give you a build estimate and a deployment timeline. If it makes sense, we move forward. If it doesn’t, you walk away with a clear map of where AI can and can’t help your practice.

The three outputs are a process map, a time-savings estimate, and a build scope. No deck. No sales pitch. Just a clear picture of what AI can do for your firm today, not two years from now.

Most firms that go through the audit end up building one agent in the first 90 days. That agent handles a single workflow, saves 8 to 12 hours per week, and proves the concept. If it works, we build the next one. If it doesn’t, you’ve spent 60 minutes and learned something about your process.

The reliability threshold for tax return preparation isn’t there yet. But the reliability threshold for research, drafting, data prep, and client communication is here today. The firms that draw that line clearly and deploy AI on the right side of it are the ones that will have the capacity to grow advisory revenue while everyone else is still drowning in compliance work.

If you want to see where that line sits in your practice, book my Omni Audit and we’ll map it in 60 minutes. You’ll walk away with a clear picture of the 10 to 20 hours per week that are currently spent on work AI can handle today, and a plan for what to build first.

The automation story for accounting isn’t about replacing CPAs. It’s about giving CPAs their time back so they can do the work that actually requires their license. That story starts with knowing where the reliability threshold is, and building on the right side of it.