Interview by Jon Reed on Diginomica
Talk of "autonomous agents" is pervasive - but are customers on board? Next up in my event highlights: a discussion about autonomous agents for finance that took me by surprise. Workday partner Auditoria.ai has a different way of approaching autonomy - and the projects to back it up.
The biggest theme of agentic AI this spring was, without question, autonomy - but we didn't find much AI clarity.
To kick off the spring event season, vendors launched into hyperbolic "autonomous agent" talk. But by the end, there was more emphasis on humans-in-loops - a likely concession to customer uncertainty.
Confusion: how much autonomy is viable? How much is desirable for your project or use case - and how do you achieve it?
Workday's latest AI partner news - and how Auditoria.ai fits in
The most coherent discussion of agentic autonomy happened at Workday's Innovation Summit, specifically around the session with Workday partner Auditoria.ai, which provides "purpose-built agentic AI solutions for the office of the CFO."
During the Workday Innovation Summit partner session, Auditoria.ai Co-Founder and CEO Rohit Gupta shared an instructive view of how Auditoria.ai handles degrees of autonomy on their projects, putting the customer in control of the pace.
In my view, those answers raised more questions. So after the analyst session, Gupta told me how they've achieved high accuracy levels, earning the autonomous trust of finance customers - often in collaboration with Workday.
But first - a note on Workday's changing AI partner program. I've written about the momentum of Workday Extend, via strong customer stories. As per Workday, there are now 2000+ Extend apps in production, from 1000+ customers and 30 certified partners. But there is more to the story now, including Workday's Extend Pro, and a new partner AI certification (Extend Pro includes Workday's developer Copilot, and AI Gateway access to AI/ML services). Since the Innovation Summit, this story advanced at Workday's DevCon 2025 event in June. As my colleague Phil Wainewright wrote:
Workday DevCon opened this week with the launch of a new AI Developer Toolset, new capabilities in Developer Copilot, its conversational AI companion, availability of AI agents from both Workday and partners in the Workday Marketplace, and a new Agent Gateway.
And, to the point of our today's topic:
The event also marked the launch of the Workday Agent Partner Network, an ecosystem of accredited third parties whose agents integrate with the Workday platform. Many of these partners already offer apps and other components built on Workday's Extend developer platform, but now there's a new network dedicated to those with agentic add-ons.
Auditoria's 2019 epiphany: build natural language models for finance
Auditoria.ai is one of these launch partners, alongside the likes of AWS, Adobe, Google Cloud, Accenture, etc. And that's where we pick up the story. Gupta is no stranger to finance automation - he's worked with RPA technologies for 15 years. So why start Auditoria.ai in 2019? As Gupta told us at the Innovation Summit, it was a light bulb moment: Natural Language Processing in finance has arrived.
I had this epiphany in the summer of 2019 that one could use natural language to drive automation for the finance function... Obviously with what we've seen with ChatGPT and things of that sort, we're really seen an inflection point here in the last couple of years.
To get this right, Auditoria ended up building their own language models for finance. Gupta:
We've got our own domain-specialized language models for finance, accounting and procurement that is purpose-built for this function. This allows finance teams to use Auditoria's AI agents to take on tasks in AP and AR and such that hopefully can improve productivity, improve efficiencies, reduce errors and really help in getting the work done faster and quicker.
And why Workday? Short answer: finance buyers want to innovate around proven solutions. Gupta explains:
I think it's reasonable to say that the finance buying center, by definition, is risk averse and conservative, right? They're skeptical. They love the tools that work, and that they've tried and tested. So it's typically ERP; it's email; it's Excel - the three Es, as I call them. What we were looking for were some of the early innovators.
Right off the bat, we were looking for folks who essentially had already moved into the cloud. They were on a foundational data model that essentially was integrated - and they fundamentally believed in open APIs, and believed in the ability to have partners that could essentially integrate with them... It was a no brainer to partner with Workday.
Sounds like a go-to-market plan - but then agentic AI takes hold. But while LLM agents have strengths, they also have accuracy limitations, due to their probabilistic nature. How do you make that work for finance leaders? Gupta:
I'll say this for finance teams: you don't have the luxury of accurate or being right 70% of the time, or 80% of the time. Megan from the PGA Tour brought up a really interesting use case. Suppliers send in requests to the AP team, asking them about the payment status of invoices. Now let's take it to another dimension: let's say a supplier sends you a statement to reconcile, which has 30, 40, or 50 invoices.
How do you think it would look if you were right on 80% of them, or 90% of them? I don't think that's going to fly.
What to do? Auditoria.ai built their own agentic architecture for finance:
That is where the opportunity to re-imagine these sorts of workflows with agentic AI comes in. And so, from an Auditoria perspective, what we built are foundational building blocks. We've got our own domain-specialized language model, so we're able to understand things like detecting intent. We've got the taxonomy of things like vendors, bills, payments, journal entries, the chart of accounts. We know how to essentially query Workday - we know what to look for. We know how to assemble information that's being requested. And then we've got the grounding to ensure that accuracy.
Do finance leaders trust autonomous agents?
But even if you have a good AI finance model, CFOs might not be ready to activate autonomous agents. That's where Auditoria's autonomous modes come in:
The other unique innovation is we've essentially allowed users to run our software in a couple of different modes. They can run it in what's called collaborative, human-in-the-loop mode... But over time, as they build trust with the agent, you can turn on what's called 'fully autonomous,' and just let the agent complete the transaction on its own.
That's at the heart of what we do. We've got about a half a dozen agentic use cases available today on the Workday Marketplace; through Connect; we're looking at Built on Workday as well. All these different options essentially are in production implementations with Workday's Fins customers.
Customer control over autonomy is one trust factor, but there is more. Gupta says distilling field lessons into meaningful KPIs is crucial. That includes industry extensions:
From the 60-plus Workday Finance implementations that we've done, we now have best practices that we recommend to clients, including specifics on industry extensions. So if you're a healthcare client, here are the attributes; here are the KPIs that you should expect.
Setting granular expectations week-to-week pays off. So when do customers go into autonomous mode? Gupta:
We tell customers: here's what you're likely going to see at the end of week three, week four and week seven. What we're typically finding within the Workday Finance install base is that customers like to run agentic software in human-in-the-loop collaborative mode for about 10 to 12 weeks. That's approximately the time where the trust starts emerging, and they're like, 'Okay, you know what, I'm willing to turn on autonomous for that transaction.'
But it's not an all-or-nothing switch. As Gupta explains, some clients (or scenarios) might be different:
What we're seeing is there's going to be a certain set of transactions, or maybe a certain set of clients that are different. Let's say the agent is servicing a client where the customer says, 'You know what, I want, white glove service for that client forever. The agent may do most of the work, but I still want to do a spot check before the transaction gets completed.' And then there's going to be the long tail where the 90% of the work is run in full autonomous mode, where the agent completes it.
How should we measure agentic AI results?
Implementing agentic AI just to say that you did it doesn't make sense. Finance leaders expect a result. How does Auditoria.ai measure success? Gupta:
We have about four to five core KPIs that essentially allow customers to truly demonstrate the ROI of their investment in agentic tech. It starts off at the highest level with the ability to deliver time savings and efficiencies, cost reductions.
This is something that we measure; we instrument our software.The agent itself, by the way, reports on the level of efficiency and productivity that it has delivered to the accounts payable team, or the accounts receivable team. There's another dimension, which is around the ability to accurately assess spend - and spend anomalies - and essentially track that over time. It's the opportunity, if you will, particularly for companies that essentially have rather complicated close processes. They want to be able to approve for their expenses in an accurate fashion, and assess where there are opportunities for optimization. So that's a second dimension that we provide.
We talk about risk management of AI. But as Gupta points out, agentic AI can also help with finance risk:
The third dimension - which is probably the other aspect that drives a hard ROI - is risk mitigation. So, the ability to look at indicators like duplicate invoices being submitted, invoice fraud... The ability to essentially assess that, validate that, ensure that it doesn't get processed... All of that is handled autonomously by the agent in a true, touchless fashion - that makes for real value.
As readers know by now, I am a stickler for accuracy - and matching accuracy levels with use cases. So after Gupta's talk, I pressed him on the limitations of probabilistic LLM technology for finance. He asserted that Auditoria's combination of finance-specific LLMs, along with determining user intent via the LLM, is delivering the level of accuracy finance users expect. Gupta also made a good point about finance queries: when you ask about ten open invoices, there is a specific, quantifiable aspect to that query that leads to a more deterministic output.
Gupta says that Auditoria's finance output accuracy levels are in the 90 percent and higher range, and in the high 90s on document processing, e.g. payment recognition - much better than OCR-level performance. He also credits this to Auditoria.ai's 300+ languages and currencies, which increases the global comprehension of the agents.
If anything, Gupta says there is a user education aspect, where users need to get across one clear intent in an agentic query for an optimal result - rather than mix three or four different instructions or questions into one query.
My take
I press on accuracy issues so that customers can peel back the vendor hype - and make the right decision for their use case, industry, or regulatory environment. Accuracy is also dependent on the amount of agent-to-agent handoffs within a process; another reason to design agentic workflows with care.
Auditoria.ai has one of the most coherent approaches to autonomy I've seen - including audit trails for output evaluation. To give customers a choice of modes, and allow them to choose which clients (or transactions) to put in autonomous mode, is clearly the sensible way forward. Auditoria's finance-specific models are also appealing. One downside, as I see it: if a model is trained on relevant finance data, it could occasionally backfire, in a sense, by reverting to what it believes is the correct finance output, versus the customer-specific data we want the model to utilize in that query (perhaps via RAG context).
But Gupta says that's another advantage of their approach: since their models are not trained on the entire Internet like OpenAI, retraining/updating the model is faster and easier as well. In two years of AR/AP automations via these methods, Gupta says their customers are not flagging the types of issues I am raising. That sets up future customer use case discussions, perhaps at the next Workday user event.
Meanwhile, it's full speed ahead for Auditoria.ai. At the April event, Gupta told us Auditoria is launching their take on the deep research approach, which they call smart research - a "strategic finance analyst," in the form of an AI agent. This research agent pulls from Workday's transactional data, combining it with "unstructured data" feeds from services like Bloomberg and Capital IQ. Via LinkedIn, Guptal told me that "This agent is being piloted with several enterprises currently, and I am planning to have this ready for launch at Rising later this year."
I'll be at Workday Rising US; with any luck, I'll get to test this agent in action.