99% GL Coding Accuracy & 83% Straight Through Invoice Processing: Inside Hyperbots' Self-Learning Finance AI

Kelly O'Neill has been a go-to-market CFO partner with Hyperbots for two years, giving her a front-row seat to the platform's evolution, especially as AI capabilities have rapidly advanced. In this interview, she shares how Hyperbots' self-learning co-pilots are reshaping day-to-day finance workflows, from GL coding to vendor management, and why the ROI of adaptive AI far outpaces static automation.

Introducing Kelly O'Neill: Two Years Watching Hyperbots' AI Roadmap Evolve

Brad: Thank you, Kelly, for joining us. Really, what we're focused on today, of course, is Kelly O'Neill. I'm Brad Boehmke with Hyperbots. The conversation that we're having today is primarily around Hyperbot's self-learning capabilities. So I have a number of questions I wanted to run by you, Kelly. And again, thank you for joining us. We'll first start with a little bit of background about your experience with Hyperbots and kind of go from there, if that sounds good with you.

Kelly: Absolutely. I've been a go-to-market partner with Hyperbots for, I think, this is my second year. So I've been watching the roadmap evolve and additional features that come out, and especially with the development in AI, we've seen a lot of progression in the capabilities.

Brad: That's fantastic. Awesome. Well, thank you so much. Your experience here is valuable, specific to Hyperbots as well, and excited to hear some of your thoughts as we kind of continue through this today. But really just to kind of kick us off, you know, your team rolled out Hyperbots six months ago.

What Self-Learning Looks Like in Day-to-Day Finance Operations

Brad: What does self-learning look like from the day-to-day?

Kelly: The co-pilots are quietly watching every override that is made in the GL: recodes, tolerance tweaks, approval escalations, and they adjust their behavior as they go. Within two closes, the invoice processing co-pilot stopped asking about penny rounding variances. It had learned that we already accepted those, as an example.

Brad: Awesome. As you're explaining that, adaptability comes to mind and how the co-pilots are assuming and understanding your workflows, and that gets sharper over time.

How Co-Pilots Adapt to Company-Specific Processes Without Writing Code

Brad: How do the co-pilots adapt to company-specific processes without code?

Kelly: If our accountants are approving due date mismatches, so maybe up to three days as an example, the pattern registers. After about ten of those approvals consistently, the co-pilot raises its mismatch tolerance to three so that future invoices flow straight through.

Brad: Awesome. So obviously, with that understanding, you take a little bit more of any manual steps, you're really getting rid of those and allowing the AI to make those decisions.

GL Recommender Agent: From 94% to 99% Accuracy Through Continuous Self-Learning

Brad: Let's talk about the GL recommender agent. How has it improved?

Kelly: It has improved quite a bit from day one accuracy. It has gone up to 94% on line-level coding, and after analyzing about two months of overrides from the users, it is now at 99%. It has even learned that Vendor A's prototype boards always hit our R&D prototypes account, something generic systems still miss.

Incremental Learning vs. Full Nightly Retraining: How Hyperbots Avoids Catastrophic Drift

Brad: Does the AI retrain the entire model every night?

Kelly: No, it doesn't need to do that. Hyperbots uses incremental learning, so only the deltas with high-confidence human confirmation are injected, and that avoids catastrophic drift and keeps audit teams happy.

Brad: Yeah, absolutely.

How Confidence Is Measured Before a Behavior Change Goes Live

Brad: How is confidence actually measured before a behavior change goes live?

Kelly: Great question. The platform needs a quorum, so generally five identical human actions plus consistency across two periods. Until then, it logs the suggestion, but it stays in an ask mode.

Self-Learning in Vendor Management: Auto-Rejecting Low-Quality Document Uploads

Brad: Can you give an example outside of invoices, say in vendor management?

Kelly: We have tightened up document quality thresholds for W9 images, and after we repeatedly rejected low resolution uploads, the vendor co-pilot learned the pattern and now it auto-rejects any file under 200 DPI before we even see it.

Brad: Wow, that's awesome. So obviously understanding that this is something we're rejecting, it's low resolution, getting it out of there.

In-Tenant Real-Time Learning: No Cross-Company Data Bleed and No PS Fees

Brad: How does Hyperbots compare when it comes to the learning approach?

Kelly: Rivals retrain generic models offline every quarter, so that's expensive and slow. Hyperbots learns in-tenant in real-time, bounded by our own data, so there's no cross-company bleed. And no PS fees.

Brad: Gotcha.

Preventing Bad Habits: How Co-Pilots Are Kept on the Right Track

Brad: How do you prevent co-pilots from learning bad habits?

Kelly: Only validated actions count. So if an override is later reversed, the data point is tagged low-confidence and it is excluded. So outliers trigger an alert rather than becoming a policy automatically.

Data Privacy and Security: Self-Learning Inside an Isolated Tenant

Brad: Is any of our data sent to the shared cloud for training?

Kelly: No. Self-learning happens inside an isolated tenant with our encryption keys, and Hyperbots' global models stay frozen. Only our private adapter layers evolve in that manner.

Pausing Self-Learning During Audit Windows for Full Compliance Control

Brad: As far as self-learning is concerned, can self-learning be paused to lock processes during audit windows?

Kelly: Absolutely, yes. There's a toggle. We freeze learning two weeks before a year-end close and then we re-enable it after audit sign-off.

How Self-Learning Shrinks Month-End Close Time in the Accruals Co-Pilot

Brad: How does self-learning manifest in the accruals co-pilot, for example?

Kelly: It noticed that we often move small GRNI items under around $5,000 to period accruals, and once confidence was built, it auto-booked that rule. So our month-end was shrunk by about 30%.

Payment Timing Intelligence: How the Model Eliminates Needless Treasury Approvals

Brad: Thinking about payment timing recommendations, does the model adjust to those specific recommendations?

Kelly: Yes. The payment co-pilot learned that Vendor B never honors early-pay discounts, so it stopped flagging those invoices, saving treasury needless approvals.

Measurable Gains from Self-Learning: 99% GL Accuracy and 80%+ STP Within One Quarter

Brad: How fast would you say you see measurable gains from self-learning?

Kelly: GL coding hit 99% in two closes. Invoice STP jumped from 74% to 83% in one quarter, and manual touch points continue to fall each and every month.

Brad: That's amazing. Whenever you're seeing 99%, close to 100%, that is amazing. And straight-through processing going from the mid-70s to already jumping, within the first quarter, over 80% is awesome, and you'd expect that to continue.

Compliance Guardrails for Regulated Industries Like Healthcare

Brad: Any guardrails for regulator-heavy sectors like healthcare, for example?

Kelly: Yes, there are. We can whitelist or blacklist fields from the learning. And as an example, for HIPAA-sensitive CPT codes, learning is locked unless compliance approves a rule update.

Brad: Awesome.

ROI of Hyperbots Self-Learning vs. Static Automation: A Living System That Pays Continuous Dividends

Brad: If you could summarize the ROI of Hyperbots self-learning versus static automation?

Kelly: I would say it's phenomenal. Static bots plateau, whereas Hyperbots improves on its own and we've reduced manual GL fixes by 95% and cut exception queues in half. And we did that without extra licenses or consultants. So it's a living system that pays continuous dividends.

Brad: That's amazing. Well, awesome. I really appreciate the discussion today and it was great to hear some of the efficiency gains, but also just how, as you continue to use the product and as Hyperbots goes and the agents continue to self-learn, it makes a better and better experience and one that's a lot more touchless, and less where a human needs to spend time on some of those redundancies. So great conversation, Kelly. Really appreciate the time per usual and I'm looking forward to continued discussions.

Kelly: Absolutely. Thank you.