Most Accounting Firms Are Stuck at Level 1 AI — Here's What Level 2 Looks Like

It's Tuesday morning. A 12-person CPA firm. One of the senior accountants has been using AI to sort client documents for six months — it saves her maybe 45 minutes a day. But this Tuesday she spent three hours manually pulling transactions from a client's bank portal, pasting them into a spreadsheet, categorizing them one by one, and building a trial balance by hand. The AI on her laptop could have done that. It just didn't. Because the firm's tools don't talk to each other, and no one has set up the connection that would make it happen.

That's Level 1. Most small accounting firms live there.

There's nothing wrong with Level 1. It's a real productivity gain. But it's not where the 3-4x efficiency improvements begin. Those start at Level 2. And the path from Level 1 to Level 2 has almost nothing to do with picking a better AI model — and almost everything to do with your data foundation.

Wolters Kluwer, one of the largest professional services technology research organizations in the world, recently published a practitioner framework defining four distinct levels of agentic AI for accounting firms. Their finding: 87% of firms with highly integrated technology stacks (75%+ of tools connected) experienced revenue growth, compared to firms running fragmented systems. The gap isn't about which AI tools a firm uses. It's about whether the firm's infrastructure is capable of running them at full capacity.

This post walks through all four levels — with concrete examples for a 10-25 person CPA firm — and explains exactly what's blocking most firms from making the jump to Level 2.

---

The Four Levels of Agentic AI for Accounting Firms

Level 1 — Taskers (What You're Already Doing)

Level 1 AI automates individual, repetitive tasks. It handles one discrete thing at a time, then stops.

What this looks like at a 12-person CPA firm:

• AI scans and categorizes incoming client documents
• An AI assistant drafts routine client emails
• Copilot summarizes meeting notes
• ChatGPT generates a first-pass memo on a tax question

Level 1 is the stage almost every small accounting firm has reached in 2026. Wolters Kluwer data shows 70% of US accounting firms use AI weekly — nearly all of that is Level 1 usage.

The efficiency gains are real but limited. You're saving 20-60 minutes per person per day. That adds up. But you're still doing the work linearly, one step at a time, with a human touching every handoff.

The ceiling: Level 1 tools don't complete processes. They help with tasks within processes. Your accountant still has to assemble the outputs, check the handoffs, and move things from one system to the next. The AI is a better keyboard, not a different way of working.

Level 2 — Automators (Where the Gains Actually Start)

Level 2 AI runs a complete workflow end-to-end without human intervention at each step.

This is the pivot. This is where accounting firm efficiency metrics start showing 3-4x improvements. It's also where most small firms hit a wall they don't fully understand.

What this looks like at a 12-person CPA firm:

The transactions-to-trial-balance workflow is the most common Level 2 implementation for small CPA firms. Here's the automated version:

1. Bank feed connects directly to your accounting platform via API
2. AI ingests transactions as they arrive
3. AI categorizes transactions using prior-period patterns and client-specific rules
4. AI flags anomalies (duplicate entries, classification mismatches, unusual amounts)
5. AI generates a review-ready trial balance
6. Accountant reviews exceptions in the morning — not each step

Instead of a three-hour Tuesday afternoon, this is a 20-minute exception review on Wednesday morning. The work still gets done. The accountant's time goes toward judgment, not process.

Other Level 2 workflows for accounting firms:

• Tax prep queue management — AI ingests client documents, checks for missing items, and populates prep checklists. Staff reviews status rather than chasing documents.
• Monthly close sequence — AI runs the standard close checklist steps (reconciliations, journal entries) and delivers a reviewed-for-exceptions trial balance.
• Engagement letter generation — AI builds a draft from the prior year's engagement, the client's current services profile, and a fee schedule. Staff reviews rather than drafts.

The common thread: Level 2 AI delivers a finished output for human review, not a partially completed task requiring human assembly.

Level 3 — Collaborators (The Review Layer)

Level 3 AI acts as an intelligent partner during complex, judgment-intensive work. A Level 3 system doesn't just run the trial balance — it flags the entry that looks statistically unusual given this client's prior 18 months. It surfaces the tax position that conflicts with a recent IRS guidance change. It identifies the expense pattern that may indicate a classification risk before the return is filed.

Level 3 AI assists the expert rather than replacing the expert's judgment. It shortens the path from "raw output" to "confident sign-off." Your accountants do more work per hour that requires their expertise — and less that doesn't.

78% of accounting firms plan increased AI investment in 2026 (Wolters Kluwer). Most of that investment is aimed at Level 3 capabilities. The catch: Level 3 systems only function well on top of Level 2 infrastructure. Review assistance requires clean data to review.

Level 4 — Orchestrators (The Integrated End-to-End)

Level 4 AI coordinates multiple agents across multiple systems without manual handoffs. From client intake to filing-ready return: intake form triggers document request → AI ingests, categorizes, and identifies missing items → prep agent populates the return → review agent flags exceptions → client communication agent generates the summary → accountant reviews and signs.

Every step that currently requires a staff member to move something between systems is automated.

Level 4 is the full realization of the agentic accounting firm. Few small CPA firms are there yet. Firms that have built the Level 2 foundation are 12-18 months from Level 4 reality. Firms that haven't built the foundation are still measuring in years.

---

Why Most Small Accounting Firms Stall Between Levels 1 and 2

The jump from Level 1 to Level 2 requires agents to run complete processes. Agents can only run complete processes if they can access the data they need, without manual intervention, at every step.

Most small firms are missing at least one of three foundations.

The Three Foundation Problems

1. API-First Platforms: What This Actually Means for a 10-Person Firm

"API-first" sounds technical. In practice, it means one thing: your tools can share data with each other automatically, without you exporting a file and importing it somewhere else.

QuickBooks Online, Xero, Gusto, Bill.com, and most modern practice management tools have APIs. FreshBooks, some older desktop software, and many specialty tools do not — or have partial APIs that require expensive middleware.

Before any agentic workflow can run, the AI needs a direct line to your data. If the line requires a human to export a CSV, the process is not automated — it's semi-automated with a human bottleneck in the middle.

Audit question: For each of your three most-used platforms, can they send and receive data to other tools automatically? Or do you regularly export data from one to import it into another? If you're regularly doing manual data transfers, you don't have an API-first stack. That is your Level 2 blocker — not your AI tool.

2. Role-Based Access Controls: The Security Layer Agentic AI Requires

AI agents access systems on behalf of your firm. They need to know what they're allowed to see.

Without role-based access controls (RBAC), an AI agent connecting to your accounting platform either has access to everything (a security and liability risk) or access to nothing useful (a functional failure). The Level 2 prerequisite is not complex RBAC — it's any RBAC. You need to define, at minimum: which AI tools can access which client data, under what conditions.

Most small CPA firms have never explicitly defined this. They have password-sharing setups, general admin accounts, or informal permission structures that were fine for a human team. They are not fine for AI agents running automated workflows.

Audit question: If an AI tool accessed your accounting platform right now, what client data could it see? Is there a defined answer? If the answer is "everything" or "I don't know," you need to define access controls before deploying any Level 2 automation.

3. Data Governance: The Unglamorous Pre-Condition

An agentic AI that encounters inconsistent data stops working or produces garbage. It is not creative about inferring what you meant. If client names are spelled three different ways across your systems, the agent sees three different clients. If transaction categories have been applied inconsistently across quarters, the trial balance agent will flag hundreds of exceptions it can't resolve.

Data governance means: consistent naming conventions, standardized input formats, and documented rules for how your team handles exceptions. This is not glamorous work. It is also the work that makes everything else possible.

The firms that have built clean data foundations — often because a managing partner was simply obsessive about it — find that agentic AI deploys in weeks. The firms that haven't spend months wondering why their AI pilot keeps breaking.

---

What Level 2 Looks Like in Practice: The Transactions-to-Trial-Balance Workflow

Let's walk through the Level 2 workflow a 12-person CPA firm can run with current tools (QuickBooks Online + an agentic AI connector):

Prerequisites: QBO API access enabled, staff roles defined in QBO, transactions categorized consistently for at least one full quarter.

The workflow:

1. Bank feed connected — QBO ingests transactions from your client's connected bank account nightly.
2. Categorization rules applied — Based on prior-period patterns you've established in QBO (or via a tool like Botkeeper, Docyt, or Vic.ai), the AI categorizes each transaction.
3. Exception queue generated — Any transaction the AI can't confidently categorize (new vendor, unusual amount, ambiguous category) goes to an exceptions list.
4. Trial balance generated — QBO generates the trial balance from confirmed categorizations.
5. Morning review — Your accountant opens the exceptions queue (typically 5-15 items for a well-managed client) and makes judgment calls. They do not touch the 85-90% of transactions that went through cleanly.

This workflow requires no new AI tools beyond what most QBO users already have access to. It requires clean bank feed connections, consistent prior-period categorizations, and 30 minutes spent configuring the exception thresholds.

The upgrade is not AI sophistication. It's connecting the pieces you already have.

---

The Self-Assessment: Which Level Are You? (3-Question Test)

Answer these three questions honestly:

1. Do any of your core processes require someone to manually export data from one tool and import it into another?

• Yes → You are at Level 1, and an API integration gap is your primary blocker.
• No → You may be at Level 2 or higher on this dimension.

2. If an AI tool connected to your accounting software today, could you immediately define what client data it should and shouldn't access?

• No clear answer → You have an access control gap blocking Level 2.
• Yes, clearly defined → You have the RBAC foundation for Level 2.

3. Are your transaction categories applied consistently across clients and quarters, with documented rules your staff follows?

• No (each accountant categorizes differently) → You have a data governance gap.
• Yes (documented standards exist) → Your data foundation supports agentic workflows.

Most small CPA firms will answer "yes, yes, yes" to these gaps. That's not a failure — it's a roadmap. Each gap you close is a direct upgrade to your firm's capacity to run agentic workflows.

---

Your One Action: Audit Your Foundation Before Adding More AI

Before you evaluate another AI tool, subscribe to another AI platform, or build another demo workflow: audit your three foundations.

This week: Identify your three most central platforms (accounting software, document management, communication tool). For each one, answer: Does it have an API? If yes, is it connected to your other platforms? If not, what would connection require?

This is a 30-minute audit. Open each platform's settings. Look for "Integrations," "API," or "Connected Apps." Note what's connected, what's disconnected, and what doesn't support API integration at all.

What you're looking for: If any of your three core platforms has no API integration capability, that platform is your Level 2 blocker. Before you spend another dollar on AI tooling, understand what it would cost to replace or integrate that platform.

The Level 2 upgrade path is 80% infrastructure, 20% AI tool selection. The 87% of highly integrated firms that grew revenue did not get there by picking better AI models. They got there by building the foundation that makes any AI model effective.

---

Frequently Asked Questions

What are the four levels of agentic AI for accounting firms?

According to Wolters Kluwer's framework: Level 1 (Taskers) automates repetitive work like document sorting and data entry. Level 2 (Automators) runs complete processes like transactions-to-trial-balance end-to-end. Level 3 (Collaborators) provides intelligent guidance during complex workflows. Level 4 (Orchestrators) coordinates multiple agents from client intake through a filing-ready return. Most small accounting firms in 2026 operate at Level 1.

Why do most accounting firms stall between Level 1 and Level 2?

Wolters Kluwer identifies three missing foundations: API-first platforms, role-based access controls, and clean data governance. Most small firms are missing at least one. Agents can't run complete processes if they can't access data automatically at every step — fragmented stacks make Level 2 impossible regardless of which AI model you select.

What does a Level 2 workflow look like for a small CPA firm?

The most common implementation is transactions-to-trial-balance: bank feed ingestion → AI categorization → exception flagging → trial balance generation. With QuickBooks Online or Xero, this runs overnight. Your accountant reviews exceptions in the morning rather than performing each step — from doing the process to reviewing the output.

Is the 87% revenue growth statistic for integrated firms real?

Yes. Wolters Kluwer's 2026 research found that firms with 75%+ of their technology stack fully integrated experienced revenue growth at significantly higher rates than fragmented-stack peers. When tools share data cleanly, time spent on manual reconciliation drops and capacity shifts to advisory work.

What should a 10-person accounting firm do first to move toward Level 2?

Audit your three core platforms for API connectivity before evaluating any new AI tools. The upgrade path is 80% infrastructure, 20% AI selection. If your accounting software, document management system, and communication tools are not connected via APIs, that gap is your actual blocker — not your AI model.

---

The Crossing Report covers AI adoption for professional services firms every week. Subscribe free for the top insights, or go premium for implementation guides and tool comparisons.