Doing → Directing: The CFO Handbook for the Age of AI Agents

Last July, I had dinner with the CFO of one of Silicon Valley's fastest-growing companies. Back then, he told me: "I don't see how AI agents are going to change anything."

All of 2025, everyone was talking about AI agents. Boards asked about them, investors asked about them, and AI founders promised growth without headcount. But in 2025, agents were either bad or didn't exist. True to the CFO's prediction, nothing really changed.

Then, in December, agents got good. Weirdly good. Claude Code went viral and engineers had a sudden, fundamental shift in how they work. If you know an engineer, look them in the eye. Behind the excitement of shipping without writing code, you will see an existential crisis: if the AI is doing everything, what am I going to do?

Finance leaders should be equally existential. AI agents can already do a significant proportion of finance work. Most finance teams just haven't caught up yet.

Doing → Directing

Engineering is the canary in the coal mine. Before AI agents, engineers measured each other by the amount of good code they could write. Now AI agents can write code faster and better than humans.

The best engineers realized their job changed. Now that agents handle coding, an engineer's time goes to deciding what they should code, guiding architecture-level decisions, and iterating on the outputs. They spend their time solely on the high-leverage work that requires a human's experience and taste.

They went from being "doers" to "directors".

The same shift is happening in finance, where 90% of finance teams' time is spent "doing". Pulling data. Cleaning it. Analysing it. Updating operating models. Writing scenarios. Pulling together KPIs and metrics. Making slides. Creating planning templates. You can name ten more things without thinking.

AI can now do all of that.

When the manual work of answering questions is delegated, a finance team's job becomes deciding which questions to ask and how to turn the answers into operational actions.

In the last year, I've talked to more than 300 CFOs of fast-growing companies. Most of them already know how urgently finance must adapt to the new game. But few have figured out how.

Doing, But Faster

The tools finance teams use today were built with the promise of "doing, but faster". If you've tried to adopt tools, you probably already hit this wall.

ChatGPT analyzes your data, but you're the one pulling it, pasting it in, and checking the output every time. It's not being done for you.

Traditional FP&A platforms like Mosaic or Pigment connect to some of your data, but not the messy stuff that matters most. You still spend your day between them and Excel, building models, reconciling sources they couldn't reach, and doing the work.

Building (vibecoding) your own dashboards with Claude, MCPs and agents starts well, until you find yourself worrying about hosting code, data security, and permissioning. More on that later.

Doing faster isn't directing. Directing is a different way of operating. Nothing on the market today was built for it.

Moving Up Maslow's Hierarchy of Leverage

Maslow said you can't think about self-actualization when you're struggling to eat. You can't focus on strategic direction when you're still trying to figure out why your QuickBooks, Rippling and Salesforce exports don't tie to each other or your operating model.

The mission is simple: spend 90% of your time on the work that determines whether the company wins, and let the agents handle the rest.

The finance leaders who have made the shift are among the highest-leverage people in their companies. They have their agent dig into why margins dropped, not just report that they did; they flag that runway is actually 14 months, not 18, because of overly optimistic sales assumptions; their agent runs a granular unit economics analysis showing that a product deserves more investment.

Agents handle the doing. Your operating model stays current, even if it lives in Excel. Actuals are updated as they come in. Board decks are drafted. Scenario models build on demand. Flux analyses run automatically. Data consolidation is done. Data inconsistencies are raised. KPIs and metrics are always ready. Ad-hoc analyses that took hours take minutes.

You handle the directing. Which market do we double down on? Where do we cut? What's the right hiring plan for next quarter? Should we raise now, extend runway, or squeeze vendors? You spend your time on the questions that determine whether the company wins and reviewing the agents' outputs. You have the data to back up your answers. And you can iterate.

Finance leaders who don't adopt agents stay in the "doing" layer at the bottom of the hierarchy of leverage.

Going from Doing to Directing is Hard

The first step is shifting your mindset. But many CFOs I meet are already aggressive about adopting the new way, and they find AI agents fall short in the messy world of real finance.

There are three obstacles most would-be "directors" hit.

1) Not knowing where to start

This morning, two CFOs said this to me:

• "Things are changing fast. I know I need to be using agents, but I don't know where to start."
• "I've tried pieces, like Claude, Claude Code, spreadsheet plugins, but I can't figure out how to fit it all together."

There's an AI capability overhang.

AI is built by engineers and largely for engineers. Knowing how to get the most out of agents requires staying fluent in a field that moves fast and compounds quickly. CFOs don't have the bandwidth needed to keep up.

2) Real world data is messy

Garbage-in, garbage-out has always been a problem. Most finance teams' experience of agents so far has been: more-garbage-in more-garbage-out faster.

The problem is that finance needs to use all kinds of data:

• Structured: ERPs, CRMs, HRIS, AP/AR, SaaS APIs
• Unstructured: spreadsheets, PDFs
• Custom: custom fields and internal product databases.

It was always "important" to build a clean data layer. Just not necessary while humans were in the loop. To use agents, a clean data layer is urgent and mandatory.

3) Vibecoding drags you into engineering work and creates security risks

AI agents can build dashboards in minutes — if you haven't tried this, you should. But they live on your laptop. When your CEO asks to see it, or your VP of Sales wants the revenue breakdown, you can't just send them a link.

You could go a step further and deploy your own app: host it somewhere (like Vercel), set up authentication, and pipe in live data. But now you're maintaining code, servers, and data pipelines. That code (I promise as a former engineer) will break at the worst possible moment: the night before a board meeting, mid-fundraise, at month-end close. And if you don't secure it correctly, your financial data will be exposed to the internet.

Then there's the internal data access problem. Most AI agents connect via API keys, and the most common setup has the CFO using admin keys. This means anyone with access to those agents can see everything, including the CEO's salary and company financials. You can build permissioning logic, create user accounts, and maintain access controls. But this isn't what you do for a living.

At best, it's a distraction. At worst, you put your company's data at risk.

Your Toolkit for Directing

Getting to the top of the hierarchy of leverage shouldn't require you to become a data engineer, software engineer and frontend designer on top of your job. That's why we built Toolkit.

1) Start with agents built by CFOs like you

We've partnered with CFOs of fast-growing companies to build a library of agents. So far, we've built agents across reporting, planning, modelling, analysis, and business partnering.

Talk through your processes with our team and AI, and we'll identify which agents you need. After that, they are yours to direct with no engineering required. Toolkit ensures you're always running the most capable agents available.

2) Data you and your agents can rely on

Toolkit's data layer combines traditional consolidation with AI fuzzy matching to give agents the clean, reliable context they need to work.

We integrate with your entire data stack:

• Structured: ERPs, CRMs, HRIS, AP/AR, SaaS APIs
• Unstructured: spreadsheets, PDFs
• Custom: custom fields and internal product databases.

3) An AI arsenal your whole team can access, with the permissions to match

Your CEO can open it. Your VP of Sales can open it. Online, in Excel and Google Sheets, or in Slack. No deployment, no servers, no code to maintain.

And you control exactly who sees what. A permission layer sits between your data sources and the AI agents, so agents can only access as much as the user. The FP&A team sees everything. Your VP of Sales sees revenue and pipeline data, but not payroll. Your board sees the board deck, but not the raw operating model.

The Shift is Happening Now

We're in pilot with some of the fastest-growing VC-backed companies in the U.S. We've seen the finance leaders who moved first compound their advantage.

Doing is getting commoditized. Your leverage is in directing.

Thanks to all the CFOs who proof-read this.

Note: numbers in this article, e.g. 10x, are illustrative. The actual gains from AI coding are increasing week to week and depend on the skill of the engineer.