…but you are missing out on so much if you let it end there!
Are you part of the 95% of businesses that haven’t seen any meaningful returns on the investments made in AI? Reference, MIT: The State of AI in Business 2025.
I don’t think anyone is arguing that AI is worthless or hasn’t impacted how people live their lives today, not to mention how businesses are conducting their day to day operations.
But how do you get into that elusive 5% where a real difference is being made?
Let’s listen to BNY, who started their AI journey in 2023.
Part of their success has been an early investment in the technology, but also a company-wide investment, 97% of their workforce has received instruction on Agentic AI. “AI is in the hands of everyone, for everything, everywhere.”
I’m guessing the bank doesn’t employ 100% software engineers. They surely have folks that are involved with accounting, accounts payable, quality control, legal, financing, sales…you know, stuff that doesn’t generally involve coding or application development.
Which brings me to my first big statement.
AI (and our MCP Server for Oracle Database) isn’t just for developers or development use cases.
I’ve had the honor and privilege to talk to thousands of people about our MCP Server and how it can transform how you work with your Oracle Database, or more importantly, how you can unlock the value from the data inside it.
The #1 question I get is simply –
What are the use cases for this technology?
I’m still workshopping my answer to this question, but it basically boils down to these points:
- all of the use cases apply!
- where does it hurt? what is causing you or your organization the most pain right now?
- which persistent and difficult problems would you tackle today if you suddenly had double the workforce or resources available?
- imagine if you could easily bring your AI to YOUR data?
Let’s consider a business task vs a coding project.
We gathered input (via survey) from more than 2000 developers last year, across 95 countries.
One of the questions we asked dealt with the type of tasks they though AI could provide assistance for. While Kris and I have both spent our careers in R&D, we both believe the true value lies in the core business application of AI.
The business wants to expand, to make more money, to return value to the shareholders, to …
Let’s look at a couple of prompts.
Disclaimer: I am not a banker, I have never worked for a bank, and in fact I’m currently a member of a Credit Union, but I like to pretend to know how things work.
Compare this:
🤖 How many personal banking accounts have been opened in our branches for zipcode 12345 in the last 6 months?
To this:
🤖 We’re looking to open new retail bank branches in zip code 12345. Using the Census income data available for that same area, correlate our existing branches and their revenue generation (deposits on hand, and average debit card transaction fees), to find the ideal possible locations from available commercial leases in local real estate listings. This data can be found in our XYZ database.
For each proposed branch, forecast the number of accounts and revenues captured for 3, 6, and 18 month timelines. You can use average rates from growth from any of our existing branch locations within 15 square miles of the targeted zip code. The goal is to maximize profits, so we should open fewer branches vs more if the return per location is lower than the average.
Ok, so in the first prompt, that’s clearly a single (or just a few) SQL query to be generated and then executed.
But for the 2nd prompt, this is clearly an Agentic experience, pregnant with opportunity.
Our Agent will need to generate a pretty involved plan, break that down into individual steps, and will need to rely on a bunch of data that can only be found in our database.
I imagine this fake scenario isn’t completely fake, and that there may be entire teams of real people out there working on figuring out these types of projects. Imagine that team having access to this technology AND not needing to know their way around a database or even SQL. Their productivity and perhaps even accuracy could both see very dramatic improvements!
But they only get there if they think BIGGER than a single, simple query.
We have data quality issues
This isn’t an uncommon problem, and in general it’s not a core technology issue. So how can technology or AI help?
Let’s take a fake but mostly real world scenario: “our biggest problem is data quality,” from a very well known shipping and logistics company.
Addresses are fun data models to crack at design time, and even more challenging to ensure your data is ‘correct.’ I’m assuming most of you have had problems shipping or receiving packages. Shipping companies DO NOT LIKE this! Happy customers equal returning, customers!
Deliveries include GPS coordinates, where the truck or driver was when doing the scan/signature for dropping off a package. And in many cases, a picture is also taken.
With a decent amount of context (next level prompt) engineering, one could have their Agent find all the different addresses for a single ‘location,’ and find a strong correlation on which one was MOST correct. This information could be sussed from your data via SQL – SQL that your agent could generate and execute (for you via our MCP Server), safely and securely in your Oracle Database.
But wait, remember those pictures? Vectorize those! Use similarity searches to further ‘weight’ the value or correctness of a physical location. Or perhaps for some reason the delivery only has the picture data and is missing the GPS coordinates? It’s easy to figure out where that picture’s from based on our AI Vector Search feature in Oracle Database 23ai.
And I don’t need to learn or re-learn how to run these spatial queries or how to interact with vectors, I can simply ‘instruct my Agent to do that on my behalf.’
Finding bad data or anti-patterns could also be something an LLM excels at. Our attention span as humans is only getting worse while LLMs are only getting better at ‘reasoning.’ Find those soul-crushing tasks that need to be done, but not by you. Hand them off to your ‘digital assistant!’ They can at least give you a really good running start.
So yes, maybe we are using NL2SQL to get the data, but value is delivering that data to the agent so it can summarize and/or identify the anti-patterns, which leads to more investigation and analysis.
Does that sound like a simple matter of taking your request and turning into a SQL statement?
One doesn’t simply ‘walk into the database.’ Our MCP Server enables a trusted and secure pathway.
So yes, we are doing quite a bit of NL2SQL here. But delivering the contents of your database to an Agent/LLM must be done with care!
- access to sensitive data must be limited, controlled, and audited
- server resources can’t be overwhelmed by ‘greedy’ agents
- credentials and database geographies should remain private
Our MCP Server provides controlled, credential-managed access, ensuring AI interactions are secure and auditable, with read-only options to prevent unauthorized changes. How is this accomplished? Simply by providing database connections for users having:
Minimal Privileges – user XYZ can only query these 3 tables, and in those tables, XYZ has no access to the compensation or PII attributes.
Limited Resources – user XYZ belongs to a resource consumer group that only gets a max of 5% server CPU. This prevents an army of ‘bots’ taking down a database by submitting hundreds or more queries, that haven’t been optimized.
Visibility and accountability – all actions being served by an MCP Server are marked as such in the database (v$session, v$sql).
The biggest question I get: Jeff, how do we get started???
You know what might be worse than not getting the biggest return on your AI investment? Maybe not doing anything at all. I’ve been talking to a lot of customers who are just not getting started in their AI journey. I STRONGLY recommend the time to take action is NOW, NOT TOMORROW.
If you wait too long, you might wait yourself out of a career, business, or organization.
Kris puts it simply here –
That’s for the single user. What about for the Enterprise? I will leave that discussion for an upcoming post!
