ClickUp CFO: You’re Asking the Wrong Questions About AI
Dan Zhang explains why finance leaders should redesign workflows for AI, automate repetitive work, and keep human judgment at the center of every decision.
Dan Zhang
CFO at ClickUp
AI Approach: AI handles middle work; humans focus on strategic decision-making and final judgment.
Model Co-Building: Co-building finance models with AI improves institutional knowledge and scenario speed.
Codifying Skills: Codifying knowledge into AI skills automates tasks and retains expertise after contributors leave.
System Control: Control is now inherent in the system, not reliant on human oversight.
AI Limitations: Full autonomous finance isn't here; human oversight remains crucial in AI-driven processes.
Dan Zhang is CFO at ClickUp, a pre-IPO AI productivity company where the agent-to-human ratio is greater than one to one.
We sat down with Dan to learn where finance leaders must focus their attention. Here's what she told us.
The most telling numbers
I'm Dan Zhang, CFO at ClickUp, a pre-IPO AI productivity company. I oversee Finance, Accounting, People/HR, and Legal, helping the company reach scale through capital allocation, strategic decision support, and public readiness.
Continue Reading for Free
Create a free account to finish this article, plus get ongoing access to timely insights and practical resources.
Have an account? Log In
I lead the full finance stack: FP&A, Corp Dev, accounting, and Deal Desk. Accounting alone spans corporate accounting, technical accounting, revenue accounting, payroll, tax, and procurement. We carry the complexity of a public-company-ready function, but we run it deliberately lean.
We crossed $300 million in ARR last year, and we're growing fast. We serve over 200,000 customers and 30 million users globally, with about 1,000 employees across six legal entities.
But here's the number I find most telling about where we are: Our agent-to-human ratio is now over one to one. Every person on my team works with multiple AI agents, and that ratio keeps climbing. Our team consists of the humans who own judgment and the agents who carry the workload underneath them.
Ask If the Work is Ready for AI — Not the Other Way Around
Most companies ask if AI is ready. We flipped the question: Is the work ready for AI?
When you break down what a finance org does, most of it sits in what I call "the middle." Initiation happens on one end: framing the question, setting the goal. Final judgment happens at the other: the decision, the sign-off. Everything in between — gathering, reconciling, drafting, and formatting — is middle work. That's where AI belongs. Humans own the front and the back. AI runs the middle.
We made real decisions based on this. We ran process mining across our workflows to find where the middle work lives, rather than guessing. We automated the close process, running it faster with 40% fewer people. I built my own agent stack, including a context agent that carries my institutional knowledge and a decision agent that pressure-tests my thinking before I bring anything to the board.
In the end, it's not about what you built, it's about the decisions you've made differently because of it. The role you didn't fill. The close you shortened. The number you trusted faster. The control vulnerability you wouldn't have caught without AI.
That's the transformation.
Why Finance Models Should Be Co-Built with Claude
We changed how we build models. We now co-build every new model with Claude, like training a new analyst.
It's painful up front. I've heard the complaints: "It'd be faster if I just built it myself." Yeah. For you. Today. Not for the team, and not next year.
Co-building closes the two gaps every FP&A team has.
- Institutional knowledge: The knowledge trapped in someone's head or in a fragile spreadsheet that nobody wants to touch is no longer an issue. When you co-build with AI, documentation isn't an afterthought; it's the build itself. We capture every assumption.
- Scenario speed: Answering the 1AM executive Slack message, "What happens to topline and cash if I push X into marketing and hold Y headcount," used to take days; now it takes seconds. Planning tools promised this for years, but never delivered, because the problem was never model mechanics. It was the context — how everything connects. Once a central intelligence holds that context, execs run infinite scenarios on the fly.
Why Finance Teams Must Codify Their Knowledge into Claude Skills
Our accounting manager built 27 Claude skills. 27 codified, expert-written playbooks that aggressively automate the preparation layer: reconciliations, close steps, entity-level prep. Each one documented, repeatable, and able to survive its author leaving.
Here's why that matters. We're adding entities and new business motions. The old playbook says to staff up the frontline. We're not. We build super agents that orchestrate those specialized skills, sub-agents that relentlessly execute the tasks, and a workflow with a human final reviewer accountable for the quality of the AI’s work. The accountability never transfers to the machine.
If a CFO wants to copy one thing, copy this: Have your best expert codify what they know into skills, put an orchestration layer on top, and put a named human at the end. Incentivize your best experts so they are rewarded for the automations they build.
Why Control Is Now a Property of the System
More human touch does not equal more control.
I came up believing control meant humans reviewing everything. AI forced me to admit that this was theater on top of a fragile system. Undocumented spreadsheets, assumptions living in people's heads, key-person risk everywhere. I still get calls from my old company about a spreadsheet I built, a year after I left. That's not a control environment. That's a hostage situation we've all normalized.
Control has now become a property of the system, not of the headcount touching it. Inputs are locked. The system logs every agent action and ties it to a workflow ID. Assumptions are explicit, and scenarios are reproducible. My auditors can ask for the trail. They could never audit what was in somebody's head.
Humans didn't leave the control system. They moved to where they always belonged: the exception layer and the final review, named and accountable. Fewer touchpoints, each carrying real weight, in a function where 99% accuracy is missing the mark completely.
Why AI Is the Draft, Not the Decision
The pattern across all these activities is that AI earns autonomy in stages. Observe, recommend, and then narrow automation with full audit trails…The faster we automate, the more judgment matters. Speed without judgment is just faster mistakes.
As far as what AI can handle, I use three rules:
- AI aggressively handles preparation, synthesis, and monitoring.
- A human always takes the last step if it changes a financial statement or moves cash.
- If it requires judgment under ambiguity or someone accountable for the outcome, that's human, because AI's output is a draft, not a decision. I own the outcome 100%.
So, for example:
- Forecasting: AI-powered, human-owned. We co-build every FP&A model with AI. Assumptions documented, logic explicit, nothing living in one person's head or a fragile spreadsheet.
- Variance analysis: My most AI-reliant activity. An analyst agent sits on live data, flags anomalies, and drafts the narrative. Humans move from the transaction layer to the exception layer. I don't want my team reviewing every line. I want them reviewing every anomaly the system surfaces.
- Cash flow management: AI forecasts it, reconciles it, and monitors it. But one rule has no exceptions: An agent can never initiate a cash outflow. Agents are read-only for financial statements. Aggressive on preparation, prudent on review.
- Budgeting: AI prepares the consolidation and the BvA. Humans handle the trade-offs.
The pattern across all these activities is that AI earns autonomy in stages. Observe, recommend, and then narrow automation with full audit trails. Humans concentrate where they outperform the machine: judgment, accountability, and new context from the real world. The faster we automate, the more judgment matters. Speed without judgment is just faster mistakes.
Why Finance Leaders Must Inject Tension into Resourcing Decisions
Speaking of budgeting, I've injected tension into every resourcing decision.
People love hiring, but most of the time, they're hiring into broken processes and calling it scale. So, put in a hiring gate. Leaders prove the AI workflow before adding headcount.
It's a sexier version of zero-based budgeting. If the workflow genuinely still needs a human after a rigorous test, great, hire. But if you never inject the tension, you'll never know how much you can tighten.
Why Finance Leaders Must be Wary of "Autonomous" Finance Tools
The biggest gap between hype and reality is end-to-end automation. It's not here. Hands-off, fully autonomous finance does not exist yet.
And I'll go further: Even Anthropic's own product team tells their finance team the same thing. Full automation is not here, even inside the company building the frontier models. If the lab that makes the model won't run its own books hands-off, neither should you.
Anyone selling you autonomous finance is selling you a demo.
How AI Impacts Capacity and Speed
The biggest positive impacts I'm seeing are in:
- Efficiency: Before AI, my FP&A team spent 80% of its time on mechanical tasks. Cleaning data, formatting reports, pasting numbers into decks. Only 20% went to synthesis and storytelling. That ratio is upside down, and it's the industry default.
- Headcount: Leaders who built the workflow proactively removed 50% of future hires. We reduced 15% of the current headcount, mostly in the lowest-tier, process-heavy shared services work.
- Capacity multiplication: A Corp Dev leader went from working 1-2 deals a month to 6-8 simultaneously. One person, no new hires. The agent widened the funnel; the human worker kept the last mile.
- Scenario speed: This is the clearest before-and-after. As I said, the 1AM executive question, "What happens to topline and cash if I push X into marketing and hold Y headcount?" went from days to seconds.
- Vendor consolidation: Internal software built on Claude has already killed multiple tooling line items, including closing automation, procurement, and ticketing tooling. So far, we've saved six figures of vendor spend.
The negative impact was the AI Slop Tax. But the failure mode wasn't AI errors; it was humans shipping unreviewed AI output. Finance slop erodes trust quietly, and it burns out your best people fastest. Nothing triggers an A-player like reading a document with zero human input.
The fix was cultural. I have to lead and hold my team accountable. They must know that I read everything they put in front of me.
Why Finance Leaders Must Focus on Workflow Foundations First
There’s a difference between adopting AI and designing a system…Make the work legible to AI before you automate anything. Codify the knowledge into skills, lock the data sources, and capture every action. A mediocre workflow on a solid foundation compounds. A brilliant demo on no foundation evaporates.
There's a difference between adopting AI and designing a system.
My first year was tinkering: building toys, bolting AI onto existing workflows, collecting impressive demos. Almost all of it was disposable. The demos died with their owners, the prompts lived in people's heads, and nothing compounded.
What I know now is that the foundation comes first. Make the work legible to AI before you automate anything. Codify the knowledge into skills, lock the data sources, and capture every action. A mediocre workflow on a solid foundation compounds. A brilliant demo on no foundation evaporates.
If I'd figured that out earlier, I could have avoided the wasted cycle: months of one-off experiments that produced zero durable assets. I also could have gotten my accounting manager writing skills from day one.
Advice for Finance Leaders
So, my advice? Build the foundation AI-first. Hold quality like the books depend on it, because they do. And spend your humans where they beat the machine: judgment, accountability, and contact with the real world.
The future org isn't a smaller, cheaper version of today's. It's a leaner team of amplified A-players operating at a level that wasn't possible before.
Follow Along
You can follow Dan Zhang's work on LinkedIn.
More expert interviews to come on The CFO Club!
