Most of the hesitation we hear about AI in ERP environments isn’t really about whether it’s worth pursuing. Almost everyone agrees it is. The real hesitation is whether it will actually work, because a lot of organizations that have tried to work through it haven’t been happy with where they landed.
We’ve seen this play out enough times to recognize the pattern, and it usually isn’t about the technology itself. It’s more about the way organizations approach it. That tends to show up in two forms.
Two ways organizations stall
The first is building it yourself. Sometimes it’s a desire for more control over a fast-moving technology. Sometimes it’s a belief that the organization’s own team understands its systems and people better than any outside partner could. Whatever the reason, the response tends to look similar: license a model directly, assign a team, and start building something custom. It’s an understandable instinct, and also one of the harder bets in enterprise technology today, because AI doesn’t move at the pace of a normal IT roadmap. It moves every few weeks, with new models, new capabilities, and new approaches, and an organization whose core competency isn’t AI is trying to hit a target that moves before they’ve finished aiming. Most don’t have the bandwidth to keep up, so rather than a clear failure, what tends to happen is a slow stall. Months of analysis lead to a decision that’s already out of date by the time it’s reached, and the organization drifts gradually toward starting over.
The second pattern comes from a different, equally understandable place. Rather than build something new, an organization leans on whatever AI capability is already built into the software they’ve purchased. It’s the path of least resistance, and because the capability is already sitting there waiting to be turned on, it can feel less like a decision and more like the next software update. The challenge is that AI built into a single application, as delivered, can only see what’s inside that application. It can answer questions about Salesforce using Salesforce data, or about your HCM system using HCM data. The moment a question needs information from more than one place, that AI reaches its limit.
Neither pattern reflects a bad decision. Both are reasonable choices, made by people doing their best to keep pace in a market that’s moving faster than almost anyone can evaluate in real time. But both tend to lead to a similar place: a lot of time and investment, and results that fall short of what the organization hoped for.
Here’s what that cost two organizations we’ve encountered, in different ways.
Two stories that show the cost
One large public university system spent nearly three years conducting intense AI evaluations, trying to land on the perfect AI approach before committing to one. Each time they got close to a decision, they found the landscape had shifted underneath them. They’re still working through it today. Had they committed three years ago to a platform built to update and expand alongside the technology, we estimate they would have already realized $5 to $10 million in operational savings. Instead, the clock has reset more than once, and they’re effectively starting over.
A different kind of organization ran into a related challenge from another direction.
A regional hospital system chose to build a custom AI solution themselves, working with an outside consulting firm to build it on top of one specific automation framework within their existing platform. Since then, investment in that particular framework has narrowed industry-wide, and the tool their whole solution depends on now has an uncertain future. That’s the risk of owning a solution outright. The organization also takes on the responsibility of tracking every shift in the market underneath it, and adapting to it alone. They paid for the solution once, and now, as the framework underneath it changes, they’re the ones left to manage what comes next.
Both organizations are still feeling the effects of these decisions. Here’s what a different path looked like.
What it looks like when it works
Seneca Polytechnic is the clearest example we have of this working well. Today, Seneca supports roughly 20,000 automated conversations a month through their AI deployment, providing around-the-clock support for an international student body spread across dozens of countries. Using standard help desk benchmarks, that volume represents an estimated $4.3 million in annual value.
What got them there wasn’t a bigger AI budget. Seneca identified a specific problem: their people needed support when no one was in the office to give it. From there, they did the work to prepare their own systems and processes internally, and brought in a partner for the expertise they didn’t have in-house. They didn’t hand over a finished spec and ask for it to be built exactly to plan. They leaned into the partnership, let that expertise shape the approach, and let the solution take form from there.
That’s the difference between this story and the two before it: a clear sense of the problem, and a willingness to let an experienced partner help solve it. They didn’t have more money or more ambition. They just chose a path more likely to succeed.
What this comes down to
Across the organizations we’ve seen succeed with AI, two things tend to be true. They start with a specific problem they want to solve, and they bring in the right expertise to solve it. That combination, their own expertise in their processes, their people, and their priorities, paired with a partner who brings deep platform and implementation experience, is what makes the difference.
None of this is out of reach. It’s also not easy to do, because conventional wisdom right now urges moving fast over moving thoughtfully. The push to do something quickly is strong. So is the temptation to simply turn on what’s already available. Both are understandable, but neither one is quite the same as having a clear, well-informed plan to guide the way.
If there’s one thing we’d want every organization weighing an AI initiative to take from this, it’s that the question was never whether AI is worth it. It’s whether the approach you’re taking can get you where you’re trying to go. We’ve watched this play out enough times to recognize what works, and we’d rather see you get there the first time.
This is the third post in our five-part series, Steady in the Storm: What We’ve Learned About AI in ERP. The next post, What Good AI Implementation Looks Like, publishes Wednesday, September 16.
