Understanding where agents create value requires recognizing that not all implementations deliver equal returns. Agents that assist with authoring, answer questions through chatbots, and summarize documents do increase productivity. But the real value comes when you use agents to automate things that weren't easily automatable before. The highest tier is when you reimagine the process entirely.

The expense management example shared makes this concrete. Today you might have a chat interface in Oracle that tells you there are twelve expenses waiting. You give it the receipts, it creates expense line items, you approve, and it posts everything. That's 90% of the work done, and it's good. But if you were reimagining the process completely, you would just pay on Amex at twenty different vendor locations during your trip. An agent would pick up each transaction, interrogate whether the hotel bill is real, whether your geolocation is accurate, whether there's fraud, and then post it without you seeing anything. That's reimagining, not just improving.

Know your agent type

The presenters are seeing two types of agents in the market, and the distinction matters for how you allocate resources. Horizontal agents are generic assistants that handle calendars, answer basic questions, and summarize documents. The presenters went to one country where they have many clients, and every single client wanted the same thing. They all wanted a chat interface with an avatar that could automate calendars and answer basic questions. It improves productivity, it's easy to set up, but the impact is marginal.

Vertical agents are different. These are tied to the thirty to sixty core processes that define any business. The presenters spoke with a CTO from one of the biggest telecoms in the world who's building what they call a digital twin. They've been trying to do this for a while but haven't been able to until now. The idea is to understand where your network is weak but demand is high, and where your network is strong but there's nobody using it. If you can balance that using simulations through a digital twin, you can deliver better service to customers. Agents can do that now, along with the infrastructure underneath like optimized LLMs, live inputs, and real time decision making. The same approach translates to the energy sector and other utilities industries. That's what vertical agents mean, something that can really reimagine the business and create significant opportunities.

Start with the process, not the demo

Agents often live and die in prototypes. There are many reasons for this, but the big one is that agents are designed as experiments. They're driven by technologists saying this would be cool if we could do that, but there's no real connection to the business world. People often haven't thought through the process they want AI to assist with. They've seen something AI can do and thought it might be useful somehow, but they haven't really considered automating the entire workflow or making a meaningful contribution to one particular business function. What happens is you get a concept, but then when the business needs to commit investment or widespread usage, it becomes a problem because the agent isn't married to how the business actually works and uses technology.

Treat data access as the product

Bad data or bad access equals a bad agent. You've probably heard that bad data equals bad AI, and the same is true for agents. Many agents don't get direct access to the data they need, or the access breaks because they're not natively integrated, or they get the wrong information. LLMs aren't good at counting. They also have a bias to answer. If you don't have data and you ask them a question, they'll still answer. So bad data equals a bad agent.

Build governance on day one

Governance is an afterthought in most organizations, but it needs to be front and center. There's agent sprawl happening at many clients right now. Agents aren't making it to production, but there are hundreds of them out there, many of them duplicates of each other. Having governance right up front is important. You need an inventory of agents, understanding of how they're performing and whether they're giving the right results. Having the right observability is key.

Use agent teams, not single geniuses

The presenters showed a compensation advisor agent during the demo. Think of the LLM as a brain and the tools as arms and legs. The brain instructs what to do and which tool to use, like which hand should reach for the glass. But the arm actually has to reach out to the glass. The compensation advisor had several tools. One was current salary, so if you're a manager you can find out the salary of individuals on your team. Another was salary history. The most important tool was get user session, because the agent needs to know who's asking for the information and whether they have the right to ask for it. The agent uses the same security model as the Fusion application, so whatever your role is and what's appropriate for you is what you see. You can't see your manager's salary through the agent if you can't see it in the system. The security is natively integrated and inherits all the security posture without you coding it specifically.

Stand on templates to speed value

When they opened Oracle AI Agent Studio during the demo, the screen wasn't blank. It came with a range of tiles, and each one represented an embedded AI agent provided automatically out of the box. They've already gone through key use cases for various industries and business functions, whether back office like HR, finance, and supply chain, or front office like sales, marketing, and customer service. You get agent templates already built to execute key tasks within those business functions. They've been pre engineered and pre tested before you get them. The prompts and logic have been robustly compiled and tested many times. You just overlay your data on top to get answers specific to your organization. You can use the templates as they are, fully customize them, or create a brand new agent. All those opportunities exist.

Design for real time signals

Higher value comes when agents act on live inputs and context, not just static knowledge bases. The telecom digital twin example requires understanding current network conditions, finding capacity mismatches, and simulating load balancing in real time. This needs optimized LLMs, streaming data pipelines, and real time decision making capabilities. The same pattern works for energy grids and other infrastructure intensive industries where conditions change continuously and decisions need to happen fast.

Democratize with guardrails

The tools need to let business users configure agents without being coders. Leaders of lines of business should be able to use click and drag software to create and maintain an agent. During the demo, they showed how agents remain in draft state until you publish them. Publishing means it goes to everybody who's supposed to use that agent within Fusion. The draft capability lets you experiment and make sure you've refined an agent before using it for real. You can create a draft in your production system and it won't do anything until you deliberately mark it as published. You can also test in a stage environment if you have one. The debug view during testing shows what data sources are being accessed, what decision paths are being followed, and what outputs are being produced, all before anything goes live.

Measure business outcomes, not prompts

Define success as cycle time, auto resolution rate, cash collected, forecast accuracy, or cost per case. Instrument the agent and report weekly. If the metric doesn't move, adjust the workflow or retire the agent. Technical health matters, but business impact is what counts. Either the numbers improve or they don't. Sentiment and anecdotes aren't substitutes for measurement.

Plan the operating model

Decide who designs agents, who owns data and prompts, who reviews outputs, and who maintains tools. Create a runbook for failures, rollbacks, retraining, and version updates. Agents operate in dynamic environments where data changes, business rules evolve, and user expectations shift. Without a clear operating model, agents drift from useful to unreliable and nobody notices until something breaks.

Moving forward

The session emphasized that agents succeed when they're designed as business solutions rather than technology experiments. The companies that scale AI agents effectively treat governance, data access, and operating models as first class design concerns from the beginning. Select one vertical workflow tied to a core KPI. Map the data path and permissions end to end. Start with a template and customize it with your tools and policies. Ship a draft in a test environment, observe how it performs, then publish. Report value weekly and scale to the next process only after you've proven the lift is real. Start small, measure relentlessly, and reimagine boldly.