What 2.5 Weeks at Harvard Taught Me About Agentic AI in Fractional CFO Work

By Andrew Wood, Founder, Wood Consulting Group

Agentic AI in fractional CFO work is the use of autonomous AI agents to execute repeatable financial operations — bank feed coding, reconciliations, reporting packets, forecast updates — while a human CFO retains judgment over the decisions that actually move a business forward. That distinction sounds simple. Living it for two and a half weeks at the Harvard Data Science Initiative's Agentic AI Intensive showed me how much rigor it takes to draw that line well.

I just completed the program — Agentic AI: Contextualized and Applied — and walked away with something more useful than a certificate. I walked away with a redesigned month-end close workflow, a framework for deciding where agents belong and where humans must stay in the loop, and a much sharper view of what the next chapter of fractional CFO services looks like.

This post is my attempt to share what shifted for me, and what it means for the clients we serve at Wood Consulting Group.

What I Thought I Knew Going In

I started the program with a working hypothesis: AI was a productivity layer. Better autocomplete, faster drafts, smarter search. I was already using it that way every day.

What I underestimated was the difference between AI as a tool and AI as an operating system. A tool waits for you to pick it up. An operating system runs in the background, takes initiative, makes handoffs, and only interrupts you when it needs a decision. That's the shift agentic AI represents — and it changes the question I ask about every workflow we touch.

That reframe sounds small. In practice, it changes everything about how a finance function is designed.

The A.G.E.N.T. Playbook — and Why It Pairs With DECIDE™

The Harvard program is built around a methodology called the A.G.E.N.T. Playbook (developed by DAIN Studios): Audit, Gauge, Engineer, Navigate, Track. It's a five-phase approach to redesigning a workflow for agent-first execution rather than just bolting AI onto a human-shaped process.

What struck me is how cleanly it pairs with our own DECIDE™ Framework. DECIDE™ operates at the strategic altitude — Direction, Economics, Capability, Decision Discipline, Execution. It answers the question, what should this business be optimizing in the first place? A.G.E.N.T. operates one level down: given that direction, how do we actually rebuild this workflow so agents and humans each do what they do best?

I went into the program thinking I'd be learning a new framework. I came out realizing I'd been handed a missing layer for the one I already had.

The Workflow I Rebuilt: Month-End Close

Every fractional CFO knows the rhythm. The first days of the month hit. You pull bank feeds, review memorized transactions, reconcile accounts, review financials for anomalies, refresh the reporting workbook, format it, PDF it, update the FP&A forecast, assemble the client packet. Mostly mechanical. Punctuated by a few moments of real judgment.

That close cycle takes hours per client. Multiply it across a growing roster and you hit a ceiling — not on revenue, but on the strategic time you have left to actually advise.

Through the A.G.E.N.T. process, I redesigned the close from a sequence of human steps into outcome-based stages where agents do most of the execution and humans focus only on exceptions. The redesign uses four agent archetypes that I'm now baking into our broader architecture:

• An Orchestrator that triggers the close, loads client context, and routes work
• Tasker agents that handle bank feed coding, recurring entries, workbook refresh, and packet assembly — running in parallel rather than in sequence
• An Analyst agent that proposes adjustments and forecast updates
• A Guardian agent whose entire job is to double-check for anomalies, compare against prior periods, and escalate anything that doesn't smell right

The result is a materially faster close with meaningfully fewer errors — not because anyone is rushing, but because the work is finally happening in the order and configuration it should have always been in.

The Question I Get Asked Most: Does This Replace the Accountant?

This is the question that comes up — sometimes openly, often quietly — in nearly every conversation I have with leadership teams about agentic AI in finance. If a CFO can run a close with agents, what happens to the bookkeeper, the accountant, the controller, the analyst?

It deserves a direct answer.

The honest truth is that agentic AI changes what accounting and finance work looks like. Pretending otherwise is the real disservice to the people doing that work today. So here's how I'd encourage every leader to think about it.

The mechanical floor of finance work — coding transactions, matching receipts, reconciling accounts, formatting reports — is exactly what agents are good at. That's not where most accountants find their professional satisfaction anyway. Talk to any experienced bookkeeper or staff accountant and they'll tell you the part of the job that drains them is the repetitive cycle. The part that energizes them is when they catch something nobody else noticed, when they can explain why a number moved, when they help a manager understand the story behind the financials.

That second part — the judgment work, the noticing, the explaining — is exactly what agents can't do well. And it's the part that has historically been undervalued because there was never enough time left at the end of the close cycle to do it properly.

So the change I'd argue for, with empathy and with intention, is this: agentic AI gives finance teams the chance to elevate their best people into roles they've quietly been ready for all along. The bookkeeper becomes an operations analyst. The staff accountant becomes a controller-in-training. The controller becomes a strategic finance partner. The mechanical work that used to fill their week becomes the agent's responsibility, and the human work — the work that requires context, relationship, and judgment — becomes the center of the role.

But that elevation doesn't happen automatically. It happens when leadership commits to it. That means investing in upskilling, redefining role descriptions honestly, paying for the new shape of the work, and being transparent with staff about the transition. The leaders who do this well will keep their best people and get more strategic value than they ever did before. The ones who use AI as a quiet substitution play will lose talent, trust, and ultimately the institutional knowledge that makes their finance function actually work.

Wood Consulting Group's role, when we step into a fractional CFO engagement with a company that has internal finance staff, is not to engineer those people out. It's to help leadership redesign the function so the people they already trust can do work that's worth their experience.

The Part of the Playbook That Mattered Most: Navigate

If I had to name the single most important phase of the playbook for a fractional CFO, it wouldn't be Audit, Engineer, or Track. It would be Navigate — the phase that asks where humans and agents should interact, where they shouldn't, which decisions require oversight, and how trust and accountability are maintained.

This is where finance gets specific. A reconciliation an agent runs autonomously is fine. A forecast adjustment that recommends a client delay a hiring decision is not. The judgment call about which client conversations the financials need to support is mine, and it always will be.

So I built explicit interaction rules into the redesigned workflow:

• Agent-autonomous for trigger setup, parallel data work, reporting refreshes, and packet assembly
• Collaborative for reconciliation and final delivery — the agent does the work, but a human weighs in when something is off
• Human-led for financial review and approval — the agent presents, flags, and recommends; a human approves before anything reaches a client

This is what I mean when I say agentic AI doesn't replace finance professionals. It removes the mechanical floor of the work so the strategic ceiling can rise.

Safety Isn't a Disclaimer — It's an Architecture

The most common second objection I hear, after the staff question, is some version of: "That makes me nervous." It should. Finance isn't a place where you tolerate hallucinations, fabricated numbers, or autonomous decisions about money.

The Harvard program reinforced something I'd already begun designing into our multi-agent architecture: safety in agentic AI isn't a paragraph in a policy document. It's a structural property of the system. It looks like:

• Hard rules. Agents do not provide legal, tax, or investment advice. They flag and recommend; licensed professionals decide.
• No fabricated data. When information is missing, the agent states the assumption explicitly rather than guessing.
• Least-privilege access. Read-only tokens. No password requests. Client consent for every data source.
• Human approval gates. Anything client-facing — proposals, financial packets, recommendations — passes through a human before it leaves the system.
• Guardian oversight. A dedicated agent whose only job is to ask, does this look right compared to history, and if not, who needs to know?

What This Means for WCG Clients

Practically, here's what this changes:

• Faster, more reliable monthly closes — not because anyone is rushing, but because the work is sequenced correctly.
• More strategic time per engagement. Hours that went into mechanical execution now go into conversations about pricing, capital allocation, and growth.
• More transparent reporting. Every recommendation our agents surface is paired with the assumptions behind it — no black boxes.
• A pathway for your existing finance team to grow. When you bring WCG in, we work with the people you already have, not around them.