AI for Finance Enters its Delegation Era

This article is sponsored by Prophix.

Automation has traditionally referred to standardizing and accelerating existing processes. But with the emergence of agentic AI tools, organizations can now delegate workloads to agents, with humans focused on oversight, validation, and key approvals. 

In a company optimized for AI delegation, a human worker might use a suite of agents to interface with core business applications like ERP, CRM, HR, and FP&A, with controls in place to manage oversight in the interest of business goals.

Finance has been slower to adopt this approach than the rest of the business. Anurag Yagnik, CTO of Prophix, an autonomous finance platform built for mid-market finance teams, told MSDW in a recent interview that the gap comes down to shortcomings of probabilistic AI in the context of work that needs deterministic answers.

Finance teams are probably the last departments within most organizations to welcome AI, primarily because the architecture hasn’t worked for them so far, especially generative AI. A generative AI engine that is 99.5 percent accurate might be great for generating content, but that would actually be grounds for dismissal on a finance team. The last thing a board wants to hear is that revenue last year was $100 million or maybe $99.5 million.

Yagnik frames this moment in finance as the Delegation Era, a term introduced at their Prophix Live conferences in Denver and London earlier this year.

We're seeing the finance function move from one that processes and executes to one that delegates work to agents that oversee, interpret, and decide. It's a different job. You're still very involved, you're the human in the loop, but you're doing more of the analytical work you couldn't get to before because you were buried in day-to-day chores.

For finance teams, the barriers to building a new generation of financial capabilities are diminishing rapidly. AI coding tools have made it cheap to create a working prototype in an afternoon. Yagnik agrees but notes that the real value of a financial platform goes further.

The cost of an initial build has never been lower. But was the cost of the initial build really the problem? The real issue is: How do you support version 10, or version 12? How do you manage data ownership changes when there's an acquisition, or when the person who wrote the code leaves the company, and a new security threat shows up in software that's been working fine for years?

The finance professionals with whom Yagnik speaks face a range of potential complications in their businesses including M&A, emerging security threats, and the loss of key personnel with institutional know-how. Any of these can complicate the viability of a financial management solution. With AI-driven development, those problems can become magnified, creating far more pain and cost than issues typically seen in custom software or spreadsheets.

Normally 20 percent of the work is building software and 80 percent is maintenance. That hasn't gone away. With AI tools, the maintenance work is a little harder now, because AI tools are generating so much more code, and there's a temptation to build more than you need.

To avoid these pitfalls, Yagnik reminds companies to define the right technical, data, and governance architectures. They need to work with a software vendor that has a track record of building and maintaining software as a proper discipline that is adaptive and robust. 

It’s not just about building software, it’s about being able to operate the software over longer periods of time. For really serious stuff like finance, it’s about building software that people can reliably use to collaborate, while achieving the right governance and data management architecture.

These risks are not lost on finance leaders, and Yagnik says the Prophix team is hearing from clients and prospects who are curious about a future in which AI plays a bigger role for them. But those same organizations are also saying no to AI touching a number in their books. That sentiment reflects the right instinct aimed at the wrong target, he thinks.

The biggest barrier to autonomous finance isn’t technology per se, it’s the trust gap. The instinct that says 'I won't let AI touch a number' is perfectly rational. It's the correct response to ungoverned AI.

He explained that Prophix has focused on using AI to interpret customer requests, while delegating deterministic workloads to a calculation engine the company has built and tested over years. With that approach, the AI never touches the number directly.

We're not taking all your data, feeding it to an AI, and saying, 'Go figure this out.' The AI interprets your request and converts it into calls to our calculation engine. The engine does the calculation and returns the result, which gets converted into something consumable for the user. Once we walk customers through that, it builds confidence.

One of the company’s guiding principles is to ensure AI solutions are delivered with a “glass box” approach that will show finance teams the steps taken and how an agent arrives at an answer. And the Prophix team is continuing to monitor customer evolution, he said. 

Yagnik maps the path to autonomous finance in four stages:

1. Data. Finance teams need to examine their data and understand if they have a governed data foundation. Making data as clean and structured as possible is a good place to start. 
2. Inquiry. Start small with AI, asking it questions and seeing how it generates response. 
3. Tasks. Delegate specific work like building a report or updating a budget line.
4. Processes. Let AI run an entire re-forecast, a forex update, or a full budget cycle end to end.

Today, many Prophix customers are still in the first stages of autonomous finance adoption, asking questions of AI and getting comfortable with responses. From there, some finance teams ask Prophix questions from directly inside Microsoft Office or Teams. A team member working in PowerPoint to build a board deck can access Prophix directly from the application, using Copilot capabilities and the latest financial data controlled via their ERP. And leading-edge customers are already taking things further, Yagnik explained.

They've connected our MCP server to Microsoft Copilot, ChatGPT, or Claude, and they're accessing Prophix data directly inside the larger workflows they've built in those AI ecosystems. That's a more advanced use case, where they're combining data across multiple systems to build their own workflows.

Prophix customers that have moved their AI initiatives ahead are leveraging Prophix One Intelligence, a suite of AI agents and capabilities that allow users to delegate tasks, get immediate insights and outputs and streamline processes with the appropriate privacy and security guardrails in place. Some of these agents include: 

• Prophix One Reporting Agent - provides quick data insights, does analysis and calculations, and creates board reports.
• Prophix One Budgeting Agent - creates and edits budgets and forecasts.
• Prophix One Autonomous Planning Agent - autonomously executes recurring FP&A processes including budget kickoffs, reforecasts, monthly maintenance cycles on a set schedule autonomously with human in the loop checkpoints built in.

Measuring the return on any of this comes down to two main questions, Yagnik says.

One is simple: how many hours a week is your finance team saving? If you're only doing inquiries, maybe a couple hours a week. Delegating more complex tasks? Maybe 4-8 hours. Into autonomous processes, saving 10 to 20 hours a week, freeing people up for higher-value analytical work.

The other question matters more to him: What were you avoiding doing more than once? 

Reconciliation is usually a once-a-month, everybody-hates-it process. The real question is why only once a month? Can you run it continuously instead? The same goes for budgeting. If that grunt work is being handled by agents, can you run your budget every quarter instead of once a year? A lot of our customers can, and that's the real value.

Companies preparing for autonomous finance should also prioritize data validation and the effort to deliver a single source of truth, Yagnik advised. And he encouraged companies to use real-time data rather than a copy of the data, which is inherently stale and can raise additional complications with data accuracy, and data authorization.

Security and accuracy are essential. Once your AI is running on real-time data, let it do the orchestrating and communicating with you. But ensure a robust, deterministic calculation engine built for finance does the calculations.

They should also ensure they are supported by a vendor that is up to the task, Yagnik explained. The Prophix team believes that the demands on finance teams will be changing more quickly than many expect, and that vision is guiding their product planning. He advises that organizations should evaluate vendors based on what has actually shipped. 

I'd focus on track record. Look at the product roadmap. Ask your vendor: what have you released in the last 18 months? There's no bulletproof answer here. You have to rely on your own research, word of mouth, and asking hard questions in a live demo, not in a deck.

We are all-in on autonomous finance. We think a large portion of the repetitive, multi-step work finance teams do today will be handled by agents, and not in ten years, in the next one to two years. Humans will be involved, of course, but their jobs will become different. They're assisting, they are reviewing, they are validating, they're approving, and auditing, instead of doing the grunt work. We're fully committed to that future.

The Prophix team is developing AI-based capabilities alongside other advances in their financial performance platform, assisted by new tooling that bridges deterministic financial capabilities and generative capabilities. Yagnik noted that today’s high standards will apply to software vendors and end users of next generation solutions. Those end users may also be exploring new rapid development options enabled by AI-based tools to fill gaps with custom solutions. 

Accelerating development is a great thing for everybody, but I think it’s important to keep a question in mind: was the initial build really the problem? The real issue is being able to support the software in version 10, 12, or 15, as well as managing data ownership changes.

When evaluating AI delegation, finance teams should look at many of the same factors they would consider in other IT scenarios. How is data being governed and can it be audited? How will proper authentication and authorization work? How will software scale and adjust as conditions evolve – or change suddenly? 

…

On September 15, 2026, see how Prophix is delivering on its vision for Autonomous Finance. Sign up today. 

FREE Membership Required to View Full Content:

Joining MSDynamicsWorld.com gives you free, unlimited access to news, analysis, white papers, case studies, product brochures, and more. You can also receive periodic email newsletters with the latest relevant articles and content updates.
Learn more about us here

Get full access

or login