Want AI outcomes? Yes - but how do customers get there? Inside Oracle's agentic apps news with Steve Miranda

Last week, in conjunction with Oracle AI World London, Oracle announced 22 agentic apps. But is this a major advancement in AI strategy, or an evolutionary step?

For that answer, we'll have to wait for customer field stories. Meantime, I have a different agenda: how do these agents work? 

Last week, Derek reviewed this news in Oracle's agentic Fusion play - from system of record to system of outcomes. I won't repeat that here, but the gist is: 22 new agentic applications across Oracle Fusion Applications (official news release: Oracle Introduces Fusion Agentic Applications). 

During my virtual catchup with Steve Miranda, EVP of Application Development, we pressed into the issues of autonomy, governance, and customer results. I don't think anyone objects to "outcomes," but what does an agentic outcome actually look like? Miranda explains: 

Say you want to optimize your cash position. You want to reduce some risk; you want to put a rush on a given order. You give business outcomes; the AI interrogates the system. It gives you proposals, and then when you accept the proposal, you can actually then use the agents to execute. So it's really kind of a dashboard or cockpit, if you will, but bringing it to life. 

In the old days, we used to talk about an actionable dashboard, but now you've got very smart reasoning agents out there at your disposal for a customer. So we put those throughout the application suite.

AI agents need context: "they inherit your security, your legislative compliance"

I'm pretty sure readers are weary of the SaaS-versus-AI debate. But it's worth a quick revisit via these new Oracle agents. Because as Miranda points out, in an architectural sense, these new agents aren't truly "new." 

It's probably a misnomer to say it's brand new - it very intentionally sits on top of Fusion. 

And why does that matter? Miranda adds:

It inherits your security; it inherits your setup. It inherits approval rules, your legislative compliance, currencies you use, etc. So you don't need to set up anything new there, but you now have a brand new capability, which is a set of agents within a given workspace.

All that company/org data that's governed in SaaS? Yeah, that matters more than ever. Good luck building effective AI agents without it: 

We know where your legal entities are. You've defined that for us. We know what your balance is saying; you've defined that for us. Your historic SEC reporting - that has been in our system... We've already done it. We've already written those things, and now we're adding agentic apps on top of that to leverage. 

Readers know I'm a stickler for agentic AI results. When it comes to AI results, there is no getting around data quality (see our new paper: diginomica independent research - the enterprise data health study).. But in this case, Fusion Apps customers have already made a cloud migration. At least with that data, they've already been through a process of organizational data discipline. Doesn't that come into play now? 

Miranda points to the thorny picture of where (most) customers are at now: legacy, hybrid, and cloud. They all want AI across their messy landscapes now - but will it fly? Miranda says you need to make a distinction between quick AI wins, versus that stubborn issue of technical (and process) debt: 

That's a critically important point. We have many customers with a hybrid environment, both our cloud and then some third-party, and other things. We're not saying there's not quick value add, or quick wins, potentially by adding AI - ours or other third parties - on top of our on-premise products, or on top of a competitor's on-premise product, or your own custom app. There definitely are quick wins. However, you're not really retiring the technical debt from a from a data standpoint, or a business process standpoint. 

If on-premise IT budgets are about keeping the lights on, is AI really going to change that? Miranda isn't so sure: 

If [you look at] most customers IT budgets, you could take your pick. You guys do stories on this all the time. 60 to 80% are keeping the lights on. Well, guess what? If you have an old on-premise platform, can you add AI to it? Absolutely. But whatever you're spending today, that 60 to 80% to keep the lights on, and, you know, keeping that old system up and running and patched, you haven't affected that at all. And so a lot of your spend stays in place. 

When customers do make that cloud applications move, the data/process benefits start to kick in. That translates to AI as well: 

There's other benefits in getting to the cloud, obviously: the pervasiveness of the AI. That journey people have gone through to get there, you're right. They've cleaned up not only data, but business processes and approvals - and everything else that makes a system work. 

Getting results from agentic apps - Miranda's advice for customers

What is Miranda's advice for customers on getting started with these new apps? Any data prerequisites? Or: should you just go where the pain or opportunity is greatest? 

Technically, no, there's no prerequisite. Pragmatically, though, I would start with some other AI for a couple of reasons. First off, most of the large enterprises we deal with have AI security, AI compliance, AI privacy, ethical constraints... What I say is: please go down that path, start with any AI - so you can get your organization ready. 

So, get stuck in organizationally - and follow the best documented (and most impactful) processes.

This is what I'm telling customers all the time: start where your biggest pain point is. If there's a number of pain points then I would say, start where you have the best documented process. As an example, our internal support wouldn't be nearly as successful had we not had a bunch of knowledge articles. [Author's note: Oracle uses its own AI service agent internally]. 

We had documentation on frequently asked questions, how to answer technical questions. So once you have that, putting support against it works great. In other cases, to automate payables or, if you have a very good, old-fashioned, what used to be called the desk manual, or operating procedures, it's great, because then you deploy the agents, and you can upload the manual. They know your processes; they know your workflow, and they know your approval rules. They know the regulations. It just makes it much faster, and they're much more efficient. 

As for pricing, Derek covered that off in detail, but the short version is:

On pricing, agentic AI is included within existing Fusion subscriptions, with a consumption model - measured in action units - for usage beyond the base allotment. 

My take - enterprise leaders are in the AI pressure cooker, but field lessons can help

Enterprise leaders are in the AI pressure cooker. I devote an absurd amount of time chasing down AI rabbit holes for definitive information, but that's my day job.

I told Miranda: I can't imagine a heads-down enterprise leader trying to parse the daily distinctions between AI coding advancements on the one hand, and Amazon taking down their own systems with AI code on the other. Or: making sense of supposed "AI layoffs" in your industry, which are often much more complicated stories (see: Block's sensationalized layoffs). Add to the list: competing vendor claims on their "agentic" capabilities, as in: agent-washing run amok. Miranda acknowledges the point: 

Unfortunately - I'll say it that way - in our business, we come out with things and we label them agentic applications. You know, that's the big press release. 

What to do? Miranda advises customers to look beyond the tech, to the issues the tech is aiming to solve: 

[Agentic apps] are fundamentally new features, but we try not to get lost in the tech of it. So my advice to customers is: what are you trying to do in your business? Oh, you're trying to optimize your cash position. Great. Got a supply chain problem? Why don't we look at your supply chain and get agentic apps, and articulate what that problem is, and fix it?  

Some of this can only be resolved by documenting results with customers - as it should be. But Oracle checks other boxes that are important to me, such as model agnosticism, and giving customers control over how autonomous they want their agents to be (notice that Oracle hasn't been overdoing it with the "autonomous enterprise" hyberbole). 

Oracle asserts that its full stack AI efficiencies are key to applying one of the more sensible enterprise AI pricing models, with many AI capabilities included in core products (to be fair, AI pricing will be a moving target for everyone, vendors and customers alike, until we get further down the road towards inference cost stability and outcome tracking, e.g. if you have to ask the AI three times to get the thing you wanted, do you pay for three inference requests?). 

As I noted in my what works/what doesn't piece on AI agents, you can have the same exact agent/process and get an entirely different result. It's all in the AI use case design - and evaluation framework. I've seen mostly poor examples of agentic service bots, but I've also seen good ones, which include intuitive human escalation, and availability of so-called "intelligent" bots when human service would have been unavailable (evenings, weekends). 

Miranda cited one key to getting agentic service right: iterating the process, and tackling the edge cases that emerge. As I see it, you start with a goal. The goal is everything. If the goal is "reduce our service headcount with agentic AI," I don't like your chances. But if your goal is "improve our customer experiences, leveraging opt-in customer data - and AI - to get a better service result," now the power of that iterative process comes into play. Miranda shares Oracle's internal agentic service results: 

We used to measure time to resolution in days, because it would take time, either for us to give the answer, and then time for customers to respond. For the first time in my career  here, we're having time to resolution in hours - and learning about outliers, etc.

Miranda says Oracle customers have similar light bulbs go off during their own adoption cycles. That leads to the real winning use cases: 

As they've deployed some agents, the feedback from end users is tremendous. The end users come back with, 'Well, here's what AI can do. Can you write me an agent that does this.' We've seen that happen largely outside of the applications, and that has taken off. So I think that experience, that change management, is extremely valuable. 

I can't resist a parting shot at the fanciful notion that enterprise user screens are going away: not in compliance-related industries! That's why the context (and source documents) provided by vendors like Oracle is so important. I might not always click on those source documents, but at times, the ability to review the original data/transaction is a very big deal. Miranda says he's calling this user behavior "exploration": 

Do you want to explore and look at things to make sure, or do you want to transact?

This is leading Oracle to a rethink what the purpose of Fusion Apps screens will be: 

 By definition, we have all the screens to do the transactions. The question is: are you going to want to go into the same screens?

No arguments here - optimizing enterprise screens for exploration/validation rather than step-by-step transaction execution makes total sense. My core UX principle: brute force simplicity is almost always a fail. Hiding the complexity behind intuitive navigation is the gold standard so often lacking in enterprise apps. If Oracle changes its UI by letting users vote with their clicks, that will be a smart move indeed. I guess we should buckle up now...