Where CFOs Should Start with AI — and What They Should Avoid
Mitch Petracca explains why CFOs should start AI with repeatable finance workflows, automate confirmation work, and keep judgment and accountability human.
Mitch Petracca
Founder of Finsider.ai
Start Small: Successful AI adoption starts with one measurable workflow, proves value, then expands across finance operations.
Human Judgment: AI should handle confirmation and pattern detection, while finance professionals retain conclusions, recommendations, and liability.
Data Quality: Poorly maintained financial data limits AI reliability, making cleanup essential before investing in sophisticated tools.
Real-Time Finance: Continuous accounting replaces stale month-end snapshots with current reconciliations, faster anomaly detection, and more timely leadership decisions.
Proven Results: Forward Firm cut engagement time 40 percent while expanding transaction review from sampling to complete ledger coverage.
After serving in various finance leadership roles, Mitch Petracca founded Forward Firm, a remote CPA and advisory service. Then, when AI removed limitations within his workflows, he built Finsider.ai to support his operating model.
We spoke with Mitch about how AI changes the work of finance professionals. He told us where to start and where caution is required.
When the AI Transformation Kicked In
My name is Mitch Petracca. I'm a CPA, Managing Partner at Forward Firm, and recently, the founder of Finsider.ai.
Create a Free Account to Read More
You'll also get access to a growing community of modern CFOs and finance executives accessing proven frameworks, tools, and insights to navigate AI-driven finance.
Have an account? Log In
The short version of my journey is that I started my career at PwC and quickly noticed how much time even elite teams spent wrestling data into shape before they could advise. Cleaning, tying out, rebuilding the same analyses from scratch, every single engagement. Tech couldn't fix this then; it was simply the cost of doing business.
When I left and built Forward Firm, I set out to run a different kind of firm. Remote-first, lean, and built entirely on leverage. We do transaction advisory, tax, bookkeeping, and about 50 quality of earnings engagements a year. We achieve this without the headcount a traditional firm would need by aggressively systematizing everything.
But about three years in, I kept running into the same wall: The market's tools weren't designed for practitioners like us. They were for internal FP&A teams. So, I built the tool I wished I had. That became Finsider, an AI-powered financial intelligence platform designed to transform messy client data into an advisory-ready data pack in a fraction of the time.
That's where the AI transformation kicked in for me. It wasn't abstract. It came from watching hours disappear on work that a well-designed system could complete in minutes.
How to Get Started with AI
AI is a force multiplier, not a transformation strategy by itself. It makes whatever you already have faster, more thorough, and more consistent. If your processes are well-designed, AI amplifies the output. If your processes are broken, AI scales the broken output faster and at higher volume.
So, the first question a CFO asks should not be which AI tool to use. It should be, "What exactly am I trying to do faster and better, and is that thing worth doing faster and better?" A CFO who automates a reporting process that nobody reads is not transforming finance. They are automating noise.
This matters because most AI initiatives in finance fail not because the technology is wrong but because they defined the problem incorrectly. The technology worked exactly as designed. It just optimized the wrong thing. Before you start, define what "better" looks like in a number: cycle time, transaction coverage, headcount per engagement, days to close. If you cannot name the metric you are moving, you do not have an initiative yet. You have a tool looking for a problem.
Once you define that, start with one workflow, prove it completely, then expand. When you see the potential, the temptation is to build everything from day one. This costs you production readiness because everything takes longer when you build across too much surface area at once. Choose something that has a clear right answer, a measurable time cost, and an obvious proof point. Ship with it, live with it, then move to the next workflow.
I didn't do that, and the cost was trust.
Why Financial Leaders Should Automate 80% and Keep 20% Human
Regarding what humans should do and what AI should do, I view it as a stack. Confirmation work forms the bottom, analysis forms the middle, and judgment forms the top. AI owns the bottom and supports the middle, but it does not touch the top.
So, AI handles transaction classification, cash reconciliation, variance analysis, anomaly flagging across the full ledger, KPI reporting, AR and AP aging, and the first-pass data pack. These activities prioritize speed and coverage over interpretation. A machine can scan 100 percent of transactions against 300 flags faster and more consistently than any associate.
We also heavily use AI in the firm’s operational layer: meeting notes and client summaries from every call, daily bookkeeping classification with a 48-hour maximum for uncategorized items, and a health score on every QBO file before we touch a QoE engagement. These are workflow automations, not finance decisions, but they create conditions for better decisions.
Explicitly human tasks include every QoE conclusion, every tax sign-off, every add-back judgment, and every client recommendation. The CPA license attaches to these outputs. AI can surface the data, flag the anomalies, and run the first pass. I decide what it means and what to do about it.
The same logic applies to pricing and deal structuring. AI can give me a comp set, a margin benchmark, a variance analysis. But determining if a client's EBITDA add-back is defensible to a sophisticated buyer's QoE team requires context, experience, and accountability. That never moves to AI.
My operating principle is: Automate the 80 percent involving confirmation and pattern recognition. Protect the 20 percent involving judgment and liability. Inverting that is either inefficient or dangerous.
The Benefits of AI Integration
With AI, we reduced engagement time by 40 percent at Forward Firm. Our lean team runs about 50 QoE engagements a year, and we compressed delivery from 3-4 weeks to 2-3 weeks. Our short-term target is 1-2 weeks. That is not a projection; that is our trajectory with current tooling.
We increased transaction coverage from the industry standard of 5-10 percent to 100 percent. We scan the full ledger against about 300 keyword flags on every engagement. This changes the quality of the work product, not just the speed.
COA mapping, standardizing a client's chart of accounts against our model, achieves 95 percent accuracy automatically. This was a full day of manual work at the start of every new engagement. It is now an afternoon spent reviewing the 5 percent that the system flags for a human.
Where AI Needs Human Support
But it isn't all good. Here are a few places where the impact of AI has been less than expected.
- Data quality: This is the key variable. AI is only as good as what you feed it. We have seen client books with GLs so inconsistently coded that the first-pass output requires significant human correction before becoming usable. You cannot shortcut the garbage-in problem with a better algorithm.
- Trust calibration: Getting experienced finance professionals to rely on AI output without running manual checks is a change management problem, not a software problem. It took longer than the build. If I had known this, I would have designed the rollout to prove trust in small, verifiable steps, rather than deploying the full capability and expecting behavior to follow. Behavior follows evidence.
- Coverage gaps: The AI flags certain categories correctly — complex intercompany transactions, non-standard revenue recognition, and certain tax adjustments — but cannot tell you what to do with them. Senior judgment is still necessary. The system surfaces the question; the CPA answers it.
- Forward-looking analysis: AI has been transformative for historical analysis. Reconciliation, variance, anomaly detection, first-pass data packs. All of these work. It has not delivered as I expected in forecasting. Building a defensible 13-week cash forecast, a revenue projection that separates price from volume, a scenario model that a CFO or lender will rely on still requires significant human construction. Tools can surface historical drivers. They do not yet produce usable forward-looking models.
- Written narratives: A QoE report is not just a data pack. It includes a conclusions section, a findings narrative, and an executive summary that a buyer, their counsel, and their lender will read closely. AI can draft an outline. It can structure a section. But the quality of written narratives from AI tools is not yet at a level I can hand to a counterparty without significant editing. I expected more progress on this by now.
A Real-World QoE Workflow
The throughline: AI owns data ingestion, standardization, confirmation, and pattern detection. The CPA owns every conclusion.
For decades, financial controls and transaction testing relied on the assumption that a well-designed sample is practically as good as full coverage. The industry standard for a QoE is 5 to 10 percent of transactions. This assumption was not arbitrary. A real constraint drove it: achieving full coverage was not feasible in a reasonable time. So the profession developed methodology around this constraint and called it best practice. AI removes the constraint.
I will walk you through a QoE engagement from the moment a deal lands to the moment we deliver the report. This is the workflow as it runs today.
Before we touch the engagement. When a new client comes in, we first run their QBO file through CheckTheLevel, a tool that scores the health of a QuickBooks file on a 0-to-100 scale. It flags duplicate vendors, miscoded accounts, missing bank feeds, and classification inconsistencies. The score tells us how much manual cleanup to budget and where to focus our first conversation with the client. We avoid surprises two weeks in.
Data ingestion. We use our tool to connect to the GL and the bank. It ingests through Plaid and Railz, and the sync runs in about five minutes. The platform maps the chart of accounts to our standardized model automatically, at 95 percent accuracy. It flags the 5 percent it cannot map with confidence for human review.
First-pass analysis. Once the data is in, three processes run automatically. First, the cash proof: GL tied to bank, every account, every period. Second, the full GL scan: 300 keyword flags across 100 percent of transactions. Third, flux analysis: variance period over period, with percentage and dollar movement automatically surfaced.
Human review layer. The CPA work starts here. I review the flagged transactions, the flux anomalies, the COA exceptions. I form the add-back thesis: identifying which items in this P&L are non-recurring, owner-personal, or non-operational, and determining if I can defend each one to a sophisticated buyer’s QoE team. The platform does not provide that judgment. The platform hands me a structured, prioritized set of questions. I answer them.
Call cadence. Fathom captures every client call, auto-joins, transcribes, and produces a summary. Within 24 hours of any client call, a written summary is in the file. No notes fall through the cracks. Action items are explicit and assigned.
Delivery. We provide a QoE report with audit-friendly workpapers, lineage from source to conclusion, and an executive summary written by the engagement team, not generated by AI.
The throughline: AI owns data ingestion, standardization, confirmation, and pattern detection. The CPA owns every conclusion. The workflow ensures that by the time I make a judgment call, I have better information, faster, with higher coverage than any prior process I have run.
Why CFOs Should Redesign Finance for Continuous Accounting
I would most strongly advocate for redesigning the close cycle itself. Not just a part of it, but its entire architecture.
Most finance organizations still run accounting as a batch process. Transactions accumulate. At month-end, a team performs reconciliation, classification review, variance analysis, and reporting. The close takes days, sometimes weeks, and leadership reads a snapshot that is already stale. Human-speed constraints designed that architecture. The constraint is gone, but the architecture remains.
Continuous accounting is the redesign. Instead of accumulating and processing transactions in bulk, you classify and reconcile them in real time, flag anomalies as they occur, and maintain a live, rather than periodic, financial position. Month-end close still occurs, but it becomes a review and sign-off, not a construction project.
We run this model on the bookkeeping side. We classify every client's GL daily. Uncategorized items have a 48-hour maximum. Anomalies surface the same week they occur, not six weeks later when someone pulls the monthly report. Before we start a QoE engagement, we run a health score on the QBO file to identify data quality gaps. By the time we perform transaction advisory work, the books support it.
This has a significant downstream effect. A CFO with a continuous, accurate financial picture can advise in real time. Currently, most CFOs advise using data that is weeks old and models that are built on a closed period, which may have had errors corrected after the fact. That is not a CFO problem; it is a system design problem.
The redesign does not require impressive tooling. It requires three decisions: define a classification latency standard and hold the team accountable, deploy a reconciliation check that runs after every sync rather than at period-end, and move anomaly review from a scheduled to a triggered activity. Any finance organization can make these process and tooling decisions today.
Why CFO Authority Shifts with Real-Time AI Data
The question is: "What happens to the CFO's authority when the data is always current and always available?"
Right now, a CFO's judgment carries weight partly because they have context that others in the room lack, built from living inside the numbers over time. When AI makes that context accessible to everyone in real time, the CFO's edge shifts entirely to interpretation, communication, and accountability, and leaders who have not made that transition will find the role eroding underneath them without understanding why.
The CFOs who will matter most in five years are not the ones who know the numbers best; they are the ones who can tell you what the numbers mean and stand behind them.
Five Pieces of Advice for CFOs
I have five pieces of advice for finance leaders:
- Start with confirmation work, not judgment work. The highest ROI, lowest-risk place to deploy AI is work that currently consumes skilled hours just to verify truth. Cash reconciliation, variance analysis, transaction coverage, COA standardization. That work has a right answer. AI gets there faster and more completely than a person. Start there, prove the value, and build internal trust. Do not start with forecasting or strategic narrative and then wonder why the output is not good enough.
- Fix your data before you buy the tool. Nobody wants to hear this advice because it is slow and unsexy, but it is the ceiling. I have watched firms deploy expensive platforms on top of inconsistently coded books, then blame AI when the output is unreliable. The tool is not the problem. If your GL is a mess, clean it first. Leverage comes later.
- Separate what carries your license from what does not, and be explicit with your team about it. The permanent dividing line in finance is not AI versus human. It is accountability. Whatever output carries your name, your sign-off, and your recommendation to a board or a lender is yours. AI can build the analysis. Judgment and liability belong to you. That line does not move, and if you let it blur, you will eventually own a mistake the tool made.
- Do not just buy tools. Build a system. A tool is a component. The operating layer around it creates leverage: the routing logic, the SOPs, the defined owners, and the handoffs. I have seen finance teams with great tools and no system produce no better output than they did before. The system is the thing.
- AI does not risk the CFO role. The CFO who ignores it risks falling behind the one who does not. This is not a threat; it is a market structure observation. If your peer delivers a QoE in 10 days and you deliver it in four weeks, the market will show you which one is the future. The window to build this competency while it remains a differentiator is shorter than most people think.
I have five pieces of advice for finance leaders: Start with confirmation work, not judgment work. Fix your data before you buy the tool. Separate what carries your license from what does not, and be explicit with your team about it. Do not just buy tools. Build a system. AI does not risk the CFO role. The CFO who ignores it risks falling behind the one who does not.
Follow Along
You can follow Mitch Petracca's work on LinkedIn. And check out Finsider.ai.
More expert interviews to come on The CFO Club!
