Based on an interview with Joyce Li, CEO and Chief AI Strategist at Averanda Partners
Click for Key Takeaways
- AI literacy isn’t about coding – it’s about taste. Genuine AI judgment comes from daily use across domains, not prompt mastery. You develop it by doing, not by reading about it.
- 62% of organizations are experimenting with AI agents in finance. Only 23% have scaled to production. The failure isn’t the technology – it’s the absence of a risk framework, outcome attribution, and leadership alignment.
- Boards want ROI tied to specific business outcomes, not productivity percentages. The next question arrives fast: who stops the agent when it does something unexpected?
- Build vs. buy has a third option: partner. Domain-specialized AI providers are increasingly able to co-build tailored solutions – bridging the gap between generic chatbots and expensive custom builds.
- Be firm on where you’re going. Be loose on how you get there. A five-year competitive vision protects against the arms race of reacting to quarterly model updates.
Joyce Li’s career has a shape that most finance professionals haven’t seen before. She started in computer science, doing analytics work for financial institutions. She then pivoted into investment management — analyst, then portfolio manager across hedge funds, long-only mutual funds, and ETFs. Now she runs Averanda Partners, working with boards and C-suites on AI strategy, governance, and responsible adoption.
That through line, in her own words, is curiosity — and a “can do” attitude that made her say yes to opportunities that didn’t fit neatly into one lane. The result is someone who holds a CFA charter, a computer science degree, and an MBA from Wharton, and who has spent her career asking one version of the same question: what could technology help us do here that we haven’t thought to try?
In a conversation with FP&A Today host Glenn Hopper, Joyce covered AI literacy for finance leaders, the growing gap between AI applications in finance and execution, what boards are actually asking about AI right now, and why the build vs. buy vs. partner decision has become one of the most consequential strategic choices an organization can make.
The Through Line: Curiosity as a Career Strategy
Long before generative AI made it easy for non-engineers to build with technology, Joyce was already asking the question that defines her approach: can we do something differently with what’s available?
As a portfolio manager doing long/short strategy, she describes using early data tools to surface non-standard information sources — government filings, trade show data, alternative datasets — to find what others were missing. The technology was primitive by today’s standards, but the instinct was the same.
“A little bit of technology can do a lot of mileage to uncover interesting insights. When I was doing long/short strategy, we discovered a lot of inflated financial claims or questionable business practices by doing that. The ability to ask, ‘Can we do something differently with the technology available to us?’ benefited my career a lot.”
The same instinct applies now. When she was building financial models, she wasn’t always the one writing the code — but she knew the questions to ask, and she knew where the important levers were.
“I haven’t been building the model for a long time, but I know what to ask. Within 10 minutes of looking at a model, I know what holes I would poke.”
That capacity — to evaluate without executing, to guide without building — is the form of AI literacy she believes finance leaders most need to develop.
AI Literacy: Not Engineering, But Taste
The question of how much technical depth business leaders need to understand AI is one Joyce has a distinctive answer to. The bar is not coding. It is not understanding transformer architectures. It is something harder to define and, she argues, harder to shortcut: taste.
“Gen AI is really easy for a business leader to get onto. They can start prompting, leverage prompt libraries, and be doing amazing things quickly. But the curve stops there. If you want to create truly value-unlocking business strategy thinking, you have to go a step deeper.”
The metaphor she uses is a sphere. The more you use AI — in any domain, not just your professional function — the more the surface of that sphere expands, and the more ideas you encounter that you can carry back to your business context. The insight that unlocks something at work might come from an experiment at home. The taste you develop for what is actually useful versus what looks impressive is only built through accumulated reps.
“For business leaders, we develop that second-level thinking based on our first-level thinking. That taste of what is a good idea, what’s not a good idea in traditional business domains — we’re so used to it. On gen AI, if you can think about it similarly, you’re going to be much more confident on determining what is the right AI initiative for your company to consider, and what is noise that isn’t related to your true competitive edge.”
For board members and executives specifically, Joyce frames the literacy requirement around evaluation — what she calls “eval.” Being able to ask the right questions to assess an AI initiative, understand where it could fail, and hold the organization accountable for outcomes is the competency that matters. You cannot do that without a baseline understanding of what AI can and cannot do.
The Pilot-to-Production Gap
The pattern Glenn Hopper described — AI is strategic, there’s budget, pilots are running, but scaling has stalled — is one Joyce sees consistently. Research confirms the scale of it: 62% of organizations are experimenting with AI agents, but only 23% have managed to scale them into production environments, a gap that cuts across industries and company sizes.
Joyce identifies two factors that explain most of the stalling, beyond the usual change management and timing issues.
The first is interface friction. Early AI adoption was largely chatbot-based: a separate window, a copy-paste workflow, a tool that lived beside existing processes rather than inside them. That model had real limits on utilization.
“With AI’s capability continuing to develop, people would love to have AI embedded into their existing workflow instead of having another window or having to copy-paste back and forth. That really affects the utilization rate a lot.”
The second is the build-versus-buy trap. Generic AI tools don’t differentiate the organization. Custom internal builds have higher failure rates and cost more than expected, especially as the underlying techniques keep changing. Many organizations find themselves stuck between two options, neither of which fully serves them.
Her answer is a third option that sits between generic AI budgeting tools and expensive custom builds:
“Partnership becomes a lot more interesting now versus 12 months ago. If you look at some of these AI labs and AI startups, they have the huge advantage of domain knowledge of a specific industry, and they can work with you much more effectively. When we build together, sometimes that’s the best solution — and it can bridge between the chatbot phase and a tool that’s really tailor-made for the business.”
What Boards Are Actually Asking
When Joyce works with boards today, the question that dominates is ROI — and it is getting sharper.
“ROI is still very much a big ask. But the big change I’m seeing is the realization of closer attribution of AI initiatives to business goals. Maybe in the past it would be, ‘Let’s improve productivity by 15%.’ Now it’s more like: if you are a Chief Revenue Officer, what are your growth goals? How can AI help there? And the KPIs of the AI team will be linked to that goal.”
That shift — from general productivity improvement to function-specific outcome attribution — is redefining how FP&A trends are shaping board-level conversations. It means AI accountability is moving closer to the business unit heads who own the outcomes, rather than sitting entirely with the technology team.
The next shift, she predicts, will come around risk. Boards in regulated industries have largely assumed AI is a go-to-market or operational tool, not a core business function. That assumption is changing as agents start appearing in more critical workflows.
“Once you assume machines have agency, the risk alarm bells will start ringing. Who will be able to stop an agent before it does something unexpected? Boards will be very worried about the risk associated with it.”
On the governance side, her position is clear: the risk framework has to come before the implementation.
“From a board angle, it’s almost like you don’t implement anything before you have the risk framework and guardrails in place for big businesses, because the reputational, legal, and compliance risk are highly expensive and very hard to recover from.”
Shadow AI and the Middle Management Dilemma
While enterprise leaders debate frameworks, employees are already using AI — often without permission. Research from IBM found that 38% of employees share sensitive work information with AI tools without their employer’s knowledge — a figure that illustrates just how wide the gap between official policy and daily reality has grown.
Joyce frames this as a middle management dilemma. Middle managers are caught between top-down direction that is often vague or aspirational, and ground-level employees who are already experimenting on their own.
“Depending on whether the head of the division has a clear sense of what AI can do or cannot do, it can be difficult to communicate: how do you go from the goal to the implementation reality? Middle management has a lot of sympathy from me. They’re often caught in between.”
One of the specific translation failures she points to is the automation versus AI distinction. FP&A teams calculating ROI on AI investments often find that process automation looks more attractive by the numbers than genuine AI implementation — because it is easier to model and faster to measure. But organizations that default to automation can miss larger transformation opportunities.
“If you do automation, it’s easier to calculate ROI and sometimes much more attractive. But in order to leverage AI in a way that really changes the business model, you need a lot more communication, alignment, and discussion — because simple ROI comparison may lead you toward automation while you miss out on potentially more interesting things.”
The Agent Question Nobody Is Asking Yet
Joyce and Glenn spend time on a distinction that is becoming increasingly important: the difference between AI tools that are called agents and AI systems that actually are agents.
A true agent, as Joyce notes, has autonomy. It acts without continuous human direction. It interacts with other systems, makes decisions, and produces outcomes without coming back for approval at each step. Most of what is currently marketed as “agentic AI” does not meet this definition.
“At least at the board level, I haven’t met someone who really cares about that distinction. But when agents start becoming core to business workflows, the risk conversation will arrive very quickly.”
She is watching a specific frontier: what happens when your agents interact with your counterparts’ agents. The pricing and business model implications for FP&A professionals are real — agent-based pricing means outcome- or action-based models, not seat licenses.
“If you sell AI-type products, agents means the pricing strategy will be very different. Outcome-based or action-based, usage-based — definitely not seat-based. For FP&A professionals who can think through a framework for modeling that, that’s a great way to stand out and differentiate yourself.”
Governance: The Black Box Is Burstable
The fear that AI in finance is inevitably a black box, opaque, unreproducible, unauditable, is one Joyce pushes back on directly. It is also a fear that most organizations haven’t formally addressed: only 27% of boards have incorporated AI governance into their committee charters, leaving the majority of companies without any structural accountability for the AI decisions being made on their behalf.
“There are sweet spots where the generative part is useful, but there are also certain parts you can control precisely. You can run deterministic code that always produces the same result from the same formula. It’s not like doing anything with gen AI means the whole machine is a black box. That misconception is something people should know so that it’s not treated as unsolvable.”
On vendor strategy, she flags an emerging risk: lock-in. As organizations build around specific models or platforms, the ability to swap components without rebuilding from scratch becomes a competitive and operational concern.
“People now realize that one of the biggest risks for AI implementation is locking in with one vendor. The vendor has to be exchangeable. If you have to exchange models or data sources, that should be very modular.”
Advice for Finance Leaders Who Want to Start
For CFOs and FP&A leaders who want to move forward but aren’t sure where to begin, Joyce’s advice cuts against the short-term grain.
“Focus on where you think your business’s competitive edge will be five years down the road — and walk that back. For finance leaders so used to thinking about NPVs and discounting back what needs to happen now, that framing is actually easier to get people aligned around than what’s happening in the next six months.”
The long-term horizon serves a second purpose. It protects the organization from the arms race of reacting to quarterly model updates.
“That also allows you to be very firm on where you are going, but very loose on how to get there.”
The distinction matters in practice. Tactical AI decisions — which model, which vendor, which tool — will be wrong within 12 months. Strategic AI decisions — what we are trying to enable, what advantage we are building, what capability we want to own — are durable. Finance leaders who anchor on the latter can survive the turbulence of the former.
Where Datarails Fits In
The work Joyce describes — developing AI taste, attributing AI investments to business outcomes, building a governance framework that enables rather than blocks — requires a foundation of reliable, governed financial data. Without it, the AI conversation stalls before it begins.
Datarails is the AI-powered FP&A platform built for Excel users. It replaces fragmented spreadsheets with financial consolidation tools that connect ERPs, accounting systems, and existing workflows into a single, governed source of truth, without requiring teams to abandon the workflows they already trust. From that foundation, it enables real-time variance analysis, dynamic scenario planning, and AI-generated narratives — the forward-looking work that defines what business partnering looks like in practice.
When finance teams have clean, consolidated data, they can evaluate AI initiatives the way Joyce describes: tied to specific business outcomes, measurable against a baseline, and built on a foundation that an auditor can follow.
To learn more about how Datarails supports FP&A teams navigating AI transformation, visit datarails.com.
About Joyce Li
CEO and Chief AI Strategist at Averanda Partners, where she advises boards and C-suites on AI strategy, governance, and responsible adoption. Joyce holds a computer science degree, a CFA charter, and an MBA from Wharton. Her career has spanned financial institution analytics, portfolio management across hedge funds and long-only strategies, and AI governance advisory work. She serves on the advisory board of OpenBB and co-authored the Athena Alliance AI Governance Playbook. She is also, by her own admission, an introvert — and would love to hear from you one on one.
Connect with Joyce on LinkedIn.
FAQs
Not engineering-level depth, but more than prompting proficiency. The goal is developing “taste” the accumulated judgment that separates genuine AI for financial analysis from initiatives that are just noise. That judgment only develops through consistent daily usage across contexts, not just within your professional domain.
Interface friction and the build-versus-buy trap are the two most underappreciated causes, and both recur across AI applications in finance. Chatbot-based AI creates a parallel workflow that limits utilization. Internal builds are expensive and have high failure rates as the underlying technology evolves. Partnership with domain-specialized AI providers has emerged as a viable third path.
The risk framework needs to come before the implementation — not as a blocker but as the foundation that makes scaling possible. Key principles include: separating deterministic and probabilistic AI tasks by use case, ensuring the vendor stack is modular rather than locked in, and maintaining audit trails that allow outcomes to be traced and reproduced.
Five years. Finance leaders are skilled at discounting back from a future state — it is how they think about NPV. Applying that same framing to AI strategy produces better alignment than reacting to short-term tool updates. Be firm on where you are going; be flexible about how you get there.
