AI-native finance expert shares his workflows — and how he makes them auditable

Sergei Mochtchenkov has spent more than 20 years in finance, including serving in CFO roles. He is now a fractional CFO who has advised over 300 clients in a wide range of industries.

We sat down with Sergei to learn about the workflows and tools he's using in finance organizations. Here's what he said.

An AI-native CFO

I'm Sergei Mochtchenkov, CFA. I'm a fractional CFO based in Kyiv, working with companies in North America, Europe, and beyond. I've spent more than 20 years in finance: financial modeling, valuation, investment analysis, and data visualization (Excel, Power BI, SQL). Along the way, I've advised 300+ clients, supported over $1 billion of debt and equity transactions and several tech-sector IPOs, and served as CFO or head analyst for businesses ranging from an AI-powered influencer marketing platform to a California real estate fund.

The structure is lean by design. I sit at the top as CFO and lead analyst; a small team handles bookkeeping, reconciliations, and reporting; and AI agents work in between — parsing bank statements, categorizing transactions, auditing models, and drafting reports. The complexity comes less from headcount than from variety: multi-entity structures, intercompany accounts, investor waterfalls, SAFEs and cap tables, and a mix of accounting stacks (QuickBooks, Xero, Odoo, Shopify, and Amazon Seller Central).

My path to AI was practical, not theoretical. Fractional work means serving many clients at once, and the bottleneck was always the same: hours spent cleaning data, rebuilding models, and auditing spreadsheets before any real thinking could start. Once AI became reliable enough to read a general ledger, parse bank statements, or trace a broken formula chain, I rebuilt my practice around it.

Today, I position myself as an AI-native fractional CFO: AI does the heavy lifting on data and first drafts, and I spend my time on judgment — what the numbers mean and what the owner should do next.

How AI streamlines data prep in finance operations

Here's the biggest change I've made with AI: I stopped treating raw data preparation as human work.

Every new engagement started with days of grunt work — downloading PDF bank statements from several banks, converting them semi-manually, then categorizing thousands of transactions line by line before any analysis could begin. Errors crept in, and the most expensive person on the engagement was spending time on the least valuable task.

So I built an AI-assisted pipeline. Claude writes and runs Python (with pdfplumber) that parses statements from different banks into clean Excel, and every output is validated with a full balance walk — opening balance plus transactions must equal the closing balance, period by period — so nothing is accepted on trust. Transactions then land in a Supabase database where AI categorizes them against mapping rules I define and approve, and anything ambiguous is routed to me. In one recent engagement, this covered four entities across four very different banks — US and Saudi, neobank and traditional.

As a result, the turnaround on the data phase dropped sharply, accuracy went up because validation is systematic rather than spot-checked, and my time moved from keying to judgment — reviewing exceptions, intercompany flows and related-party items, which is where the real findings are.

How AI powers model audits

I often perform financial model audits and rebuilds. Here's how I do it:

• Step 1 — Set the rules: Create a Claude project with written instructions: naming conventions, English-only deliverables, formulas as plain text, and in-scope and out-of-scope files.
• Step 2 — Inventory and dependency map: Claude, using Python (openpyxl), catalogs every sheet, named range, external link, hard-coded number, and formula pattern, then builds a map of what feeds what. We do not touch any cells until we understand the model's structure.
• Step 3 — Error hunt: AI traces calculation chains against the model's intended logic — waterfalls, IRR, debt schedules, cash flow — and produces an error list, detailing cell locations, severity, and dollar impact. Before we classify anything as an error worth fixing, we check if anything downstream reads that cell.
• Step 4 — Two-chat execution: I run a main chat for scoping and review, and a separate execution chat. The execution chat receives one written task brief at a time, carries it out, and independently checks the main chat's reasoning. When two AI passes disagree, it signals me to investigate further.
• Step 5 — Verification: I perform full recalculations, zero-error scans, tie-outs to source documents, and stress tests on key scenarios. Then, I review the changes and logic myself.

Why AI handles data and humans handle judgment

As far as where to use AI and where to use humans, my rule is simple: AI prepares, tests, and drafts; a human decides and signs.

I rely on AI for:

• Data ingestion and reconciliation: statement parsing, transaction categorization, AR and AP analysis.
• Model auditing: tracing formula chains, finding broken links, hard-coded overrides, and latent errors in client models. AI reads thousands of formulas without getting tired.
• Forecasting: For e-commerce clients, I've built demand-forecasting systems that clean the data, normalize for seasonality and ad spend, and produce sales ranges with 95% confidence bands that feed a purchase plan.
• Variance analysis
• First-draft commentary
• Scenario runs
• Research such as competitive benchmarks and market data

What stays explicitly human:

• Assumptions: A model is only as good as the growth rate, cap rate, or churn assumption behind it, and those require understanding the business and the people running it.
• Capital allocation, valuation conclusions, and anything that goes to an investor, lender, or board under my name: As a CFA charterholder, I am accountable for the conclusion, and accountability cannot be delegated to a model.
• Decisions about people and relationships: Cost cuts that affect roles, negotiations, and the conversation with a founder whose plan doesn't work.

The two main benefits of AI in finance

The scale of verification has been a big benefit of AI. It exhaustively checks work previously sampled. Recent examples include: a 134-item metrics verification workbook for a film production's investor pitch, where we traced every figure in the deck back to source documents; a client quoting tool with 126 tasks and about 1,200 formulas, delivering zero formula errors; and a four-entity bank statement conversion where every period passed a full balance walk.

The cost is another win. AI lets organizations automate repetitive tasks and processes. It also makes it easy to cut unused software seats and marketing spend in underperforming channels.

Lessons learned from the places where AI struggles

AI errors are dangerous precisely because they look professional. I've seen polished output with a wrong range reference or a misread line item that would pass a casual review. One small but telling example: formulas quoted inside a report included a leading equals sign, so the spreadsheet tried to evaluate them, producing broken references and spurious circular-reference warnings. Now, a standing rule dictates formulas in deliverables are written as plain text.

AI will also happily "fix" things that don't matter. Before deciding whether to recover or correct a value, I first check if anything downstream actually reads it. This simple discipline saves much wasted effort.

And here are a couple of specific places where AI doesn't deliver:

• Large legacy Excel models: I expected AI to 'just understand' and fix a 30-tab model. It doesn't. Big workbooks with external links, hidden sheets, old VBA or ActiveX controls, and years of patches overwhelm it — it loses the thread across tabs and proposes fixes that break something elsewhere. Slower, more disciplined approaches work: first, build an inventory and dependency map of the model; then, let AI work one bounded task at a time, with a separate pass checking the result. AI is a strong analyst, but not yet an architect.
• Forecasting under regime change: Statistical and ML forecasts excel when the future resembles the past — seasonal e-commerce demand, for example. Living and working in Ukraine, I often see how badly history-based models handle war, sanctions, currency shocks, and sudden cost inflation. In those environments, AI produces confident numbers with no real basis; scenario thinking and human judgment matter far more.

And a smaller point: transaction categorization plateaus. AI handles the bulk of it well, but the ambiguous remainder — related-party transfers, owner expenses, intercompany flows — is exactly where accounting risk sits, and that still needs a human rulebook.

How AI introduces hidden cost risks in finance

CFOs need to keep an eye on compute cost as a hidden variable cost of goods sold.

Many leaders still think of AI as software — a fixed subscription. For AI-native products, it isn't. Every user action can trigger model calls, and inference cost scales with usage, prompt length, and model choice. That turns what looks like a high-margin SaaS business into something with real, volatile variable costs. I regularly see AI startups whose gross margins quietly erode as they grow because their model's design assumed software-style margins, and nobody tracked cost per customer, per feature, or per transaction.

The fix is basic finance discipline applied to a new cost line: Measure unit economics at the level of the AI call, put compute in cost of goods sold where it belongs, set budgets and alerts, match model size to the task, and price the product with compute in mind. Compute cost control is now a core CFO responsibility in any AI-heavy business — not an engineering detail.

How to manage AI

AI needs management, not just prompts. At first, I treated it like a very fast tool, but it behaves much more like a brilliant new hire with no memory of yesterday and a strong urge to be helpful. Here are three things I'd tell my earlier self if I could:

• Write the rules down once. Every firm has unwritten conventions — how percentages are shown, which accounts appear in which report, how formulas are documented, what language deliverables are in. If they live only in your head, AI will "reasonably" do something different every time.
• Separate questions from changes. Ask AI why a number looks odd, and it will often "fix" it. My standing rule now: when I ask a question, AI answers first and waits for approval before changing anything. Early on, that would have prevented several silent edits I had to unwind.
• Keep tasks small, and the original untouched. My first attempts gave AI too much at once — "review this whole model and fix it." The results looked impressive and were hard to verify. Bounded task briefs, one change at a time, a clean copy of the original, and a change log make every step checkable.

These three things would have saved me hours of rework, a few embarrassing near-misses caught in review, and the false belief, for a while, that the main value of AI is speed. The main value is that you can afford to verify everything — but only if you design the process for it.

How AI can be used in working capital

Most finance teams still manage receivables, payables, and inventory with monthly reports and gut feel, even though this is where cash is made or trapped. But AI is unusually strong here because inputs are high-volume, repetitive, and contain patterns humans lack time to see.

Two areas I've redesigned:

• Inventory and purchase planning for e-commerce: For brands in sportswear, swimwear, toys, and protein bars, we replaced "reorder what sold last season" with a forecasting system that I already touched on briefly. We clean the sales history, normalize for seasonality and ad spend, forecast demand with 95% bands, then build a purchase plan that explicitly trades off storage cost, gross margin, return on equity, volume discounts, and supplier lead time. The result is a purchase decision based on a range of outcomes, with the cost of being wrong in either direction visible up front — rather than a single guess.
• Receivables, payables, and communications: I build Claude-powered analysis over AR, AP, and the business's own communications, so finance sees not just that an invoice is late, but why — a dispute, a delivery problem, a customer under stress — and can act earlier. The same lens catches duplicate or unusual payables and supplier terms that have quietly drifted.

Why Claude Code and Cowork are game changers in finance

Claude is a game changer — specifically in its agentic form (Claude Code and Cowork), where it doesn't just chat but works with files, writes and runs Python, reads and builds Excel models, and connects to databases and business apps.

With it, I can parse bank statements, audit a 30-tab model, build a forecasting system, draft an investor memo, and build a Power BI-ready dataset — and, importantly, it writes code I can keep, re-run, and audit, rather than just giving me answers. Project-level instructions let me encode each client's rules once, so the work stays consistent across sessions.

But Excel remains invaluable. It's still the language investors, lenders, and boards read, and the best AI workflows I run end in a clean, transparent spreadsheet that a human can check line by line.

How AI is changing teams and training paths

The pyramid has become a diamond. Traditionally, a finance practice needs many junior people doing data entry, reconciliations, and first-draft models, a few seniors reviewing them, and one partner on top. Today, AI does most of that base layer.

The talent mix I look for has shifted accordingly. Less value in data processing speed; much more in three things: understanding how a business makes money, the ability to write a precise brief and test the result, and the confidence to tell a founder that a number or a plan is wrong.

The open question for the profession is the training path. Juniors used to learn by doing the grunt work. If AI does it, we have to deliberately teach judgment — by having people review, challenge, and explain AI output rather than produce it from scratch.

Five pieces of advice for CFOs

Here's my advice to CFOs:

• Use AI yourself, this week, on real work. Don't delegate AI to an innovation committee or wait for the perfect enterprise rollout. Take one painful, repetitive process — a reconciliation, a model audit, a board-pack first draft — and do it with AI personally. You can't lead a transformation you haven't felt.
• Design the controls before you scale. Every AI workflow needs a built-in check that doesn't depend on trusting the AI: a balance walk, a tie-out to the general ledger, a second independent pass. AI makes errors faster and more convincingly than a junior analyst, so your control design must be better than before.
• Keep judgment and accountability human. Let AI prepare, test, and draft; keep assumptions, capital allocation, and anything signed in your name with people. Write that boundary down so your team knows where it is.
• Rebuild the team, not just the tools. The future finance team is smaller, more senior and more analytical. Invest in people who can frame questions, challenge numbers, and talk to the business — those skills compound with AI; pure data processing does not.
• Measure value, not activity. Faster reporting is nice, but the real question is whether you are finding more value: margin you didn't see, cash tied up in the wrong place, a deal priced wrong. If AI frees up the finance team's time and that time doesn't go into decisions, you've only automated the old job.

Follow along

You can follow Sergei Mochtchenkov's work on LinkedIn. Or check out aistartupcfo.com and financialmodeling.net.

More expert interviews to come on The CFO Club!