AI-native Head of Finance says to avoid forcing automation where AI’s ceiling is genuinely lower than a human’s

Nicolas Frendo is the Head of Finance at Acephalt, an AI-native fintech company building due diligence tools.

We sat down with him to learn how AI is changing financial reporting. Here's what he told us.

From finance to tech

I'm Nicolas Frendo, head of finance at Acephalt, a Toronto-based fintech building AI-driven due diligence tools for private equity, venture capital, and family office investors.

My path here wasn't a straight line from finance into tech. I studied business law at Carleton, then did a Master's in Financial Management at the University of Ottawa, picking up an honors certificate in financial markets from Yale along the way. Early on, I worked as a data analyst and financial analyst across a few organizations, then worked as a research assistant and fellow inside Ottawa's innovation ecosystem, at places like Invest Ottawa and Treefrog Accelerator. That's where finance and technology started to overlap for me.

The pattern I kept running into was the same everywhere: Investment decisions still ride on manual, judgment-heavy work. Quality of earnings reports, deal diligence, financial modeling — all of it takes weeks and still comes down to a handful of overworked analysts doing their best with incomplete information. That gap between how much rigor a deal deserves and how much time anyone has to give it is what pulled me out of traditional finance roles and into building a company.

At Acephalt, we're closing that gap with a multi-agent AI platform that compresses weeks of diligence into hours, without losing the explainability investors need to trust the output. I run finance and go-to-market, so my day splits evenly between the numbers and the people, turning what deal teams need into what we build next.

The moment we're in right now — AI moving from a novelty to a working part of the investment process — is exactly the shift I got into this industry to drive.

How AI is changing financial reporting

Before, our team spent roughly 15 hours a week manually pulling transaction data from our accounting system, reconciling accounts, and assembling the close package in spreadsheets. We also paid for an outside bookkeeping service to cover part of that work.

Now, Claude and Daloopa handle the pull-and-structure work: pulling the data, categorizing it, and flagging anything that looks off. A custom Claude Skill packages it into our close report format, and our team reviews the flagged items and signs off instead of building the report from scratch.

The close cycle that used to take about ten business days now takes about three, freeing up roughly 15 hours a week for higher-value analysis.

Which finance tasks must stay with a human

We rely on AI for financial modeling and forecasting outputs, reconciliation and variance analysis (bank vs. QBO matching, vendor spend checks), and first-pass risk flagging on deals.

But capital allocation and fundraising decisions, final deal and investment calls, pricing and client negotiations, and risk sign-off all stay human.

So, AI handles the parts with a checkable right answer — numbers reconcile, a formula is correct, a variance has a cause. But humans handle decisions that carry judgment, relationship context, and accountability that isn't in the data.

The good and bad results of AI in finance

Here are the good results that we're seeing with AI:

• Financial modeling accuracy on client work increased from ~37% to 80-90% after we added the symbolic validation layer.
• QoE-style due diligence that used to take weeks now turns around in hours to days. Note: Claude isn't capable of doing this with accuracy, so I'm referring to our own product here.
• Reconciliation and vendor-spend checks that used to be manual now take minutes and catch things a manual scan misses. For example, it flagged 13 duplicate charges from one vendor.

As far as bad results:

• Early versions relied on raw LLM output for modeling and produced plausible-but-wrong numbers; the 37% accuracy figure was a real failure mode, not a hypothetical one.
• Clients pushed back on trust: Valuation discrepancies (DCF vs. market price) surfaced without enough explainability, so we had to build more transparency into outputs before people would rely on them.
• We had some platform/usability issues that still required human troubleshooting rather than the AI catching its own failure. For example, specific automated workflows hanging or behaving inconsistently.

Where AI tooling falls short

We expected AI to meaningfully help early-stage deal sourcing, but internally, we've kept it a lower priority since it hasn't been worth the build effort yet; relationship-driven sourcing still outperforms it.

Our buyers — private equity professionals, often older and relationship-driven — don't adopt through digital or tech-forward channels the way a typical SaaS buyer would. AI, standard lead-generation tools, LinkedIn outreach — all of it resulted in low engagement and difficulty reaching decision-makers. Because the market does not operate that way. We shifted to manually building lists of target firms and hosting exclusive dinners to build the personal relationships and warm intros these buyers respond to.

The product could be accurate and fast, but without that relationship-building capability, none of it would have reached the people who needed it. AI tooling didn't close that gap; distribution built around how the buyer trusts and decides did.

Why AI must show its work in finance workflows

AI struggles with auditability, trust, and compliance.

For example, DCF-generated valuations didn't match market pricing, and initially the model's output lacked sufficient explainability for clients to understand why the numbers diverged. A confident number with no visible reasoning path doesn't earn trust just because it's accurate; clients pushed back until we built more transparency into how it derived outputs.

AI accuracy improvements do not automatically solve auditability and compliance. A model can be mathematically correct and still fail on trust if it cannot show its work. And the platform around it can fall short if it lacks the compliance credentials that the client's risk process requires.

Nicolas Shares

Planning and treasury also benefit from AI, but reporting is where the redesign has real teeth, because it’s the one place where “close enough” was never acceptable, and AI without validation just made a slower version of the same trust problem.

Why reporting needs an AI redesign

I'm biased, but reporting is the part of the finance lifecycle that needs an AI redesign — specifically, the audit-adjacent reporting that underpins deal diligence (quality of earnings, financial statement analysis).

It's the area we've redesigned where the old model breaks down first: reporting has traditionally been priced and structured around analyst hours, not output quality or speed, which is why a QoE report costs $25,000-plus and takes weeks regardless of how repetitive the underlying work is. That structure doesn't survive once a validated AI layer produces the same CPA/CBV-aligned output in hours.

It's also the part of the lifecycle most exposed on trust: reporting sits right at the boundary with audit, so accuracy and explainability aren't optional, as they might be for internal planning. Every output must withstand the same scrutiny a human-prepared report would, which is why we built the deterministic validation layer before touching the pricing model, not after.

Planning and treasury also benefit from AI, but reporting is where the redesign has real teeth, because it's the one place where "close enough" was never acceptable, and AI without validation just made a slower version of the same trust problem.

Why Claude Code is a game changer for data hygiene

Claude Code is a game changer because it's the one tool that underpins everything else in this workflow — not a single-purpose solution. I use it to clean and normalize raw data before it touches a model (PitchBook exports, bank downloads, messy CSVs), to reconcile accounts and catch discrepancies that would otherwise take hours of manual line-matching, and to build the QBO-ready outputs that close the loop on bookkeeping.

If I lost every other tool but kept Claude, I could still get from raw, messy source data to a usable financial output. Lose Claude and keep everything else, and I'm back to doing that cleanup and reconciliation by hand.

It's also the layer I trust to tell me when something doesn't add up rather than producing a confident-looking answer, which matters more to me than any single feature. Everything else in the stack assumes clean, reconciled incoming data. Claude brings the data to that state in the first place.

Why rigor is no longer about time and headcount

There's an assumption that rigor scales with time and headcount, that a more reliable number requires more analyst hours or another layer of manual review. That link isn't real.

A deterministic validation layer catches errors a stack of tired analysts would miss, and it does it in the time a single review pass used to take. The old proxy for trustworthiness — weeks of work, multiple sign-offs — stopped meaning anything once the actual check was: "Does the math hold?" It wasn't "Who touched it and how many times?"

The result is that our team remained small and generalist longer than a traditional finance organization at this stage would. Judgment and client-facing roles absorbed what used to require a larger analyst bench.

Three pieces of advice

My advice? CFOs navigating this shift should treat AI as a tool for anything with a checkable right answer, while keeping capital allocation, deal decisions, and pricing human, since those carry accountability and context no model has. Never trust raw LLM output on financial calculations without a deterministic check behind it; fluency isn't accuracy, and unchecked models will produce plausible but wrong numbers.

Expect resistance where legacy systems are entrenched: AI has to integrate with the tools your team already trusts, not replace them, or it loses credibility fast. Build explainability into outputs from day one; a number nobody can trace back to a reason won't get used no matter how accurate it is.

And know where AI's ceiling is genuinely lower than a person's; don't force automation into places like relationship-driven deal sourcing just because the technology is available.

Follow along

You can follow Nicolas Frendo's work on LinkedIn and his personal website. And check out Acephalt.

More expert interviews to come on The CFO Club!