Inside Acumatica’s generative AI strategy, and why cloud ERP vendors must become data and innovation partners

Inside Acumatica’s generative AI strategy, and why cloud ERP vendors must become data and innovation partners Summary: Acumatica was careful not to turn the Acumatica Summit keynotes into AI propaganda sessions. But off the keynote stage, candid talks with Acumatica’s product leads brought context. What do customers expect, and what are the adoption barriers? Let’s dig in. In my Acumatica Summit 2024 review, I noted the crossroads Acumatica now faces: Suddenly, it’s a different ERP market. Buyers are asking tough questions about AI and data security. New approaches to ESG bring the possibility of real-time tracking inside ERP, rather than static reporting after the fact. And how do you balance talk of scaling businesses, when plenty of companies have their hands full in a sluggish and volatile market? Given the current risks of building your own generative AI applications, customers are looking to trusted vendors to deliver on AI – and Acumatica is no exception. As I wrote: Whether it’s the AI skills gap, data privacy, or IP concerns, Acumatica customers want to move ahead, without taking on all of that AI risk profile. And yet, customers have employees who are using consumer GPT tools in their work – not the best situation. In that case, any productivity gains come with a big risk tradeoff . Acumatica’s principles of innovation – why now? But where do customers go from here? Getting a handle on a vendor’s AI strategy is not easy. Considerations include: the state of the vendor’s data platform, and how automated their workflows are. After all, “co-pilots” are limited if you can’t use them to invoke automated workflows. Acumatica has done well on these criteria, but generative AI brings new scenarios – and new problems. During a media/analyst discussion with Acumatica’s Doug Johnson, VP Product Management, and CPO Ali Jani, we learned more about what lies ahead. Acumatica’s day one keynote featured Acumatica’s soon-to-be-finalized principles of innovation: “responsible,” with guardrails to protect customer data “practical,” to address real-world needs “valuable,” driven by customer input Acumatica didn’t formally unveil these innovation principles yet, so what else can we expect? Jani said the details behind these principles are important, and Acumatica wants to get them right: For example, when we talk about guardrails in AI, who determines what those guardrails are? It’s a big question. And: how are we transparent on those guardrails? So I wanted to make sure we put some tones into the document, so that they know we always have the best interest of our customers. There’s just some fine tuning we want to do over the next couple weeks, then we’ll put them out. Will Acumatica’s partners adhere to the principles of innovation? I provided context for Acumatica’s AI approach in September’s The principles of customer-centric ERP – how does AI fit in?. But the discussion with Johnson and Ali brought new insights. My next question: will Acumatica’s partners adhere to the principles of innovation? Acumatica certainly can’t dictate the AI terms for all its partners, but wouldn’t some type of validation option be useful? Otherwise, you might have a contradiction between your principles and how your partners approach it. The short answer is yes – Acumatica is revamping its marketplace offering (Jani says this should be formalliy announced by the end of Q1). The marketplace will essentially have two tiers. Each tier will have criteria for marketplace inclusion. But for the second, higher teir, those partners will have to abide by Acumatica’s principles of innovation. Partners will be elibible to pursue that higher standard, and give customers a way to identify validated solutions – in accordance with those principles. But what does Acumatica’s practical AI focus mean in reality? Acumatica has learned from its years of embedding other forms of AI into products, such as QA automation and release management. How much does gen AI shake things up? The three main “buckets of AI” for Acumatica customers Jani says that he puts AI into three main buckets: 1. Things you can automate – “Zero touch automation as much as possible.” 2. Interactive assistance – “You will get help from AI to make decisions; you are assisted as you go along.” (Jani also uses this ‘bucket’ for information search and task assistance as well). 3. Trusted advisor – “It does work behind the scenes, and comes back after the fact and helps you: ‘ Hey, did you know you could do this better? ‘ (from my vantage point, predictive analytics/AI might fit well in this bucket also) With these buckets in mind, the role of gen AI is clearer. As Jani sees it: Generative AI is much more [for] interactive assistance mode. So as I write an email, ‘Can you format my email,’ or: ‘Put tasks that I’m doing in better words, so it looks professional when I send the text.’ Things like CRM, help desk chat, even search, play very well with generative AI. On the other hand Jani says that gen AI doesn’t excel at decision-making: When it comes to decision-making, generative AI is not very good for that at all, and you have to rely much more on traditional AI. We’ve been looking at these transformer models that generative AI uses… There’s a lot of studies that show the way generative AI works is very similar to how our brain operates. But they haven’t been able to cross the bridge for decision-making. That’s going to be maybe 5-10 years away, in our opinion. Jani sees the potential to combine “traditional” AI and generative AI, each playing to their strengths: Traditional AI is working with big data in much more of an analysis form. It’s going to help us short-term on those scenarios. The same thing with the advisor-type AI that goes on behind the scenes. Our GL anomaly detection is a good example of that. We built an engine that can work in traditional ML/AI. Now we’re working on another engine that works with generative AI. Large Language Models and customers’ data challenges Like most ERP vendors, Acumatica has no plans to build their own Large Language Models, or transformer models. Ali: We’re not going to build our own transformers. Big companies are putting a lot of money in this – they will provide these things as a service. We will simply plug it in into the scenarios that we think makes the best sense. Whereas traditional ML/AI, how we diagnose the data, how we find the right anomalies, [that’s where] we have made stronger investments – in terms of the technology. That part is going to change quite a bit over time. It’s always based on the customer. That’s why I said, ‘We’re not in a rush. We don’t want to put out features that are going to be outdated a year later.’ But as fellow analyst Josh Greenbaum argued, most customers have a different set of problems: their data isn’t clean enough to get a good AI result. Jani responded: That’s what’s important to us – [help them] cleanse the data… Right now, we can’t ask them, ‘Trust the AI making a decision for you.’ First they have to trust their data. So we have to help them with pointing out where those anomalies are – and slowly get them to say ‘Yes, how did you fix it’? With a diverse customer base, aggregating data isn’t always easy: The other challenge we also run into is that every one of our customers is unique… So you end up putting a lot of services and effort in them to fine tune for the customer – and that’s just not scalable. Acumatica AI success stories – via Acumatica Summit 2024 Despite these customer adoption obstacles, Jani has success stories to share. He referred us to the day one keynote video with Acumatica customer Titan, and the results they’ve had with automated replenishment. This quote comes directly from the site visit Jani’s team made at Titan Sales and Consulting. In Titan’s video segment, they explained how their buying went from managing replenishment knee deep in spreadsheets to automated processing. They went from three people watching all their suppliers “all day long” to “one person three hours a day” – what Titan called a “dramatic savings.” Now, that buyer is “just eyeballing and managing exceptions.” The most important line of all: And he [the buyer] trusts the system. With the automated system updating the replenishment imbalance, Titan reports that “our inventory dropped probably about 20%.” After showing the video, Jani said: So automated data-driven decision making – that’s a mouthful. Managing by exception, the ability to trust data. Clearly, this had a big impact for Titan and their bottom line. Data-driven analytics and AI are far better equipped to be handled by machines. [The keynote replays are currently available for streaming with free sign up]. My take – on Co-Pilot integration and beyond Jani says he hopes to integrate Microsoft Co-Pilot with Acumatica, assuming the typical licensing sort can get done. Along those lines, there were early Co-Pilot integration demos on site, from Acumatica and its partners. I watched an intriguing Co-Pilot integration demo with Acumatica partner Velixo, which uses Microsoft Co-Pilot’s (still limited) Excel integration to make it easier to create visualized reporting, and drill into spreadsheets with natural language queries. Though I have lingering questions about the accuracy tolerance with this gen AI use case, and how the output could be comprehensively cross-checked by (expert) humans, I can see how these features could extend the use of Excel for things like queries and scenario planning – well beyond the domain of Excel experts inside organizations. Besides Excel, there are obviously other Microsoft Co-Pilot integrations that could improve the workflows of Acumatica customers. As Velixo said to us during the demo, if Microsoft is spending billions on this technology, why not piggy back on it, and bring it into play with Acumatica, where we understand what cloud ERP users really need? Indeed. For now, Acumatica is keeping architectural details like its preferred external LLMs under wraps. Vendors are going to pull and switch multiple external LLMs going forward. What matters is protecting each customer’s data in those scenarios; Acumatica shows every sign of diligence in that regard. We had a spirited back-and-forth with Acumatica about the imperative to move ahead with gen AI, and why. Historically, Acumatica has taken a strong stance on tying next gen tech to customer results. As I argued, the difference with gen AI is that enterprises are under internal pressure to have official answers for gen AI tools, to limit the rogue use of Shadow IT by employees (and, in some cases, to placate impatient CEOs, who are worried about the perception of falling behind). But Jani is firm on this: Acumatica is not going to release premature AI functionality just to have something in the domain (this might seem like an obvious stance, but it’s really not – I’m tracking the PR bogpit of this just about every week in Hits and Misses). As Jani said to us: Public companies are under pressure; they have to show these things they don’t have. We want to make sure whatever we do, we can deliver on. So that’s why the process has to be very practical. A good balance between the two would be to bear down on an “AI readiness” curriculum within Acumatica Open University. My understanding is that some of this type of content is already available to the Acumatica community. If Acumatica can provide customers with substantive preparation they can make, both from a data and internal skills angle, that should help bridge the gap until Acumatica’s fully-vetted solutions are available. In my view, such a curriculum would educate customers as to the risks of rogue ChatGPT use by employees, and provide information on sensible AI policies – and ways to take advantage of external tooling with privacy protections where it does exist, e.g. Co-Pilot integrations via Acumatica and its partners. There is much more to learn, but these discussions advanced the topic – and that’s what makes the tarmac time worth it.