Table of Contents
Key takeaways
- When Pooja joined Docker, she was looking for a company that was still building its finance infrastructure and had a genuine appetite for AI. At Docker, her team negotiates contracts, models investment scenarios, and builds execution plans alongside the business.
- The first FP&A process Pooja began automating at Docker was variance commentary – a task that was data-rich, structurally consistent, and cognitively punishing during the most compressed time of the month.
- AI gave her team back 30% to 40% of the time they had been spending on variance analysis. But establishing their AI workflow took months of iteration across different tool combinations and prompting methods.
- Pooja's approach to finding AI opportunities is to notice the moments when you wish you could hand something off as you go about your day. If you could teach a junior analyst to do it, you can probably teach AI to do it. Her advice for anyone starting out is to stay on it, refine it as you go, and let the value compound over time.
- In the human-AI partnership, people supply business context, own the output, and decide what it means. They’re the ones sitting in the leadership meetings, and that judgment is entirely theirs.
- Pooja believes FP&A and strategic finance functions are converging into a single, unified role and that Docker's current structure offers a preview of where most finance teams are headed.
There’s a version of starting a new FP&A job where you quietly settle into routines like tackling monthly close, applying budget templates, and drafting variance reports at the end of every quarter.
Pooja Agrawal has never been particularly interested in that approach. She began her career at a health tech company where her role covered everything at once: budgeting, forecasting, ROI quantification, acquisition due diligence, post-integration work, and even implementing an entire finance system. It was a lot, but Pooja also views it as the foundation she needed. The breadth of her role in finance gave her a clear sense of what she was looking for and what she enjoyed.
When Pooja left San Francisco, the moment had a particular texture to it. ChatGPT had just launched, and AI was the only thing anyone was talking about. Pooja knew that she wanted to work at a company that was still actively building out its finance infrastructure and had a genuine appetite for AI. She wanted to play an active role in shaping how finances were handled in the business. That aligned perfectly with what Docker was looking for.
Building the financial engine
When Pooja joined Docker, she knew the company operated with a startup mentality that would keep finance on its toes. She’s built a financial function that works for Docker by taking what she sees across cost, revenue, headcount, and cash flow and using it to help leadership make better decisions.
Pooja is interested in the space between numbers and perspective. Numbers may be her starting point, but she's really drawn to teaching people what they mean and what to do about them.
The most painful task for Docker’s finance team
In Pooja’s experience, commentary analysis is one of the most painful parts of the finance workflow.
It’s a repetitive process that usually starts the same way. Every month, Pooja’s team goes through all of the P&L, headcount, and cash flow, writing explanations for why performance landed where it did. Then, they take those explanations and repackage them for different audiences and occasions. A single analysis might yield an executive summary, a functional leader view, a board deck, and any other materials needed to tell the financial story.
The timing of variance analysis is part of what makes it such a challenge. All of this typically happens during close, when deadlines are tight and everyone on the team is already stretched thin.
Beyond the strain, what made variance analysis a good candidate for AI was its structure. The Docker team had built their variance analysis tables in Pigment. The numbers were already clean and organized, and they had standardized how they wanted commentaries written – including tone, format, and level of detail for each audience. All of that structure made it easier to give AI exactly what it needed to produce a useful first draft.
The workflow Pooja built connects Pigment directly to Claude. Pigment holds the data and keeps it organized, with the right contextual notes in the right places. Because Claude has been trained on examples of Pooja’s commentaries at every level, she can do everything from within its conversational interface. Today, Claude handles the aggregation, the consolidation, and the first draft.
Instead of formatting and repackaging, Pooja’s job is now focused on refinement. That includes tasks like verifying accuracy, checking that the right drivers are being called out, and making sure the story being told is the one that matters.
Months of iteration, followed by a breakthrough
Since adding AI to the variance analysis workflow, Pooja has seen a roughly 30% to 40% reduction in dedicated commentary time, with close moving from a day and a half to closer to a day. But the impact she emphasizes most is not the hours recovered. It’s the cognitive load that went with them.
There was also a secondary benefit she had not anticipated. The process of building the AI workflow forced the team to standardize everything more rigorously. That better standardization meant less rework, fewer rounds of back-and-forth, and consistently higher-quality outputs.
But getting there didn’t happen overnight.
For months, Pooja tried different AI apps and workflow combinations. Nothing happened the way she hoped at first. The data structure wasn’t quite right. The prompts weren’t refined enough. And the outputs weren’t always consistent.
Her advice for anyone starting out with AI is to stop trying to think about it in the abstract. Instead, pay attention to what’s happening in your average workday. If you can imagine teaching a junior analyst how to do something, then you can probably teach AI to do it.
Once you’ve identified the right problem, she says, AI itself can help you figure out what to do next.
The goal is to start small, prove value, refine the prompts, and let it compound.
Clean data, clean output
Pooja returns to a principle that governs how her team approaches AI: the quality of the output is determined by the quality of the input.
When Pooja built the commentary workflow, she was deliberate about keeping the data clean and focused, scoping each task to its own clearly defined input. P&L commentary runs on P&L data. Feeding Claude cash flow and balance sheet data simultaneously introduces ambiguity about which numbers belong to which analysis – particularly when naming conventions overlap. Clarity of input is what produces clarity of output.
The same principle governs how she is building out Pigment more broadly. Before opening the analyst agent to wider use, Pooja is doing the foundational work of ensuring the data environment is clean enough to produce reliable answers. She has watched AI improve significantly over the past two years, as competition among the major models has pushed each toward greater accuracy and consistency. With well-structured inputs and refined prompts, she trusts the output considerably more today than she did when she was first getting the workflow off the ground.
For all the time Pooja has invested in making AI work inside her function, she is clear-eyed about where it stops. "AI can probably say, oh, revenue is trending a certain way, and here is the information. It tells you objectively what is happening. But the team says, this is why revenue is trending the way it is."
AI is faster at processing data, identifying patterns, and surfacing structure, but the decision about what that output means and what to do about it belongs to humans. That conviction shapes who Pooja hires. She wants people who are already thinking about how they would embed AI into their workflows.
What finance looks like from here
One of Pooja’s favorite ways to spend time outside of work is putting together logic puzzles. It’s a hobby that she sees as parallel to finance. You start with a set of scattered, incomplete clues, and you try to solve for something. You look for patterns and move as fast as you can. The puzzle doesn’t change over time; what changes is how quickly and clearly you can see it.
As AI removes the mechanical layer from her day-to-day work – the aggregation, the summarization, the reformatting – Pooja sees her team's actual value coming more into focus.
FP&A and strategic finance, she believes, will increasingly operate as one. Docker has already made sure of this. Still, Pooja thinks most companies will find that day arriving faster than they anticipate. Beyond expediting processes, it’s an outcome that makes finance a job Pooja is more interested in and excited about.
