What Is AI, and What Types Should Oracle Fusion Users Actually Know About?
A precise taxonomy of AI — from computer vision to autonomous agents — written for every Oracle practitioner, from Finance Director to solutions architect. The vocabulary the entire series builds on.
In this article
00How we got here — a brief history
01What AI actually is
02The AI family tree — three branches
↳Branch A — Perception AI
↳Branch B — Analytical & Predictive AI
↳Branch C — Generative & Agentic AI
03What AI is not
04How Oracle layers these technologies together
05The four types you will encounter most in Fusion
06What this means for your role
07What comes next
Find your reading path
Business & Functional
Finance Directors, HR Leads, Operations Managers
You are evaluating, sponsoring, or adopting AI in your Oracle environment. You need to ask the right questions — and recognise when something is not really AI. The vocabulary in this article is your toolkit for every vendor conversation.
Start here →
IT & Configuration
System Admins, Technical Leads, Fusion Config Specialists
You configure, validate, and govern AI features. The taxonomy maps directly to what each AI type requires — different data prerequisites, different validation, different governance controls. Full read recommended.
Start here →
Consultant & Architect
Solutions Architects, Implementation Consultants, Pre-Sales
This taxonomy is the framework for every AI conversation with a client. Before recommending any feature, know which branch it sits in, what data prerequisites it carries, and what governance it requires. The architecture page maps the broader Oracle AI platform.
Start here →
Not sure? Start from the top — the article works for all audiences.
The Problem With the Word “AI”
There is a problem with the word “AI”.
It is used to describe a spam filter and a system that writes your board commentary. It covers the software that flags a duplicate invoice and the agent that autonomously processes a procurement cycle from requisition to payment. It appears in every software vendor’s pitch deck, in every executive briefing, and increasingly in the release notes of every Oracle Fusion update.
That range is not a marketing accident — it reflects the fact that AI genuinely is a family of very different technologies. But when everything gets labelled AI, the word stops doing useful work. You cannot evaluate what you cannot distinguish. You cannot govern what you cannot name. And you cannot have an honest conversation with a client, a vendor, or your own team about what your Fusion environment is actually doing if the vocabulary is imprecise.
“You cannot evaluate what you cannot distinguish. You cannot govern what you cannot name.”
This article gives you a precise map. It is written for everyone on the Oracle journey — from a Finance Director being asked to sign off on an AI readiness programme, to a solutions architect configuring features in a live Fusion tenant, to a consultant who needs to explain AI outputs to a sceptical end user. The concepts are accessible to any reader; the architecture and implementation implications go deep enough for the most technical.
By the end, you will be able to name, describe, and differentiate every type of AI you will encounter in an Oracle context — and understand how they combine inside the platform you are already running.
00 How We Got Here — A Brief History
AI did not arrive with ChatGPT. It did not arrive with Oracle’s 25D release notes. It has been developing for over seventy years — and understanding that trajectory explains why the technology feels simultaneously overhyped and genuinely transformative right now.
This series is anchored to Oracle Fusion Cloud 25D and 26A — the two most recent active releases as of early 2026. Earlier versions are not covered in depth: the AI landscape shifted enough between 24D and 25D that retrofitting would be more misleading than helpful. If you are on an earlier release, the concepts here apply; the specific features and configuration paths will differ.
01 What AI Actually Is
Before the taxonomy, one foundational distinction.
Traditional software follows rules you write. If an invoice amount exceeds a threshold, route it for approval. If a candidate has fewer than five years of experience, filter them out. The logic is explicit, deterministic, and authored by a human. The software does exactly what it is told, every time, in the same way.
AI software learns rules from data rather than having them written in. You do not tell a machine learning model that invoices from new suppliers with unusual line-item patterns are higher risk. You show it thousands of historical invoices — some that turned out to be fraudulent, some that did not — and it learns to distinguish them. The rules it develops are not written anywhere a human could read them. They emerge from the data.
That shift — from explicit rules to learned patterns — is the fundamental boundary between conventional software and AI. Everything else in this article follows from it.
“AI does not require the internet, does not require a chatbot interface, and does not require you to type a prompt. Most of the AI running in your Oracle Fusion tenant right now operates completely in the background.”
One important clarification: AI does not require the internet, does not require a chatbot interface, and does not require you to type a prompt. Most of the AI running in your Oracle Fusion tenant right now operates completely in the background — it reads data, applies a learned model, and produces a score, a flag, a forecast, or a draft. You may never see it work. You only see the output.
02 The AI Family Tree
AI is not a single technology. It is a family of technologies, grouped here into three branches based on what each one does and how it works. Understanding the branches matters because they have different data requirements, different failure modes, different governance implications, and different implementation considerations.
Perception AI processes raw, unstructured inputs — images, documents, audio, natural language — and converts them into structured data that other systems can use. It is the part of AI that reads.
Computer Vision teaches machines to interpret images. In an Oracle context, this means reading a scanned invoice, identifying a supplier logo, or extracting line items from a photographed receipt. The model has learned from millions of labelled images what text looks like, where amounts typically appear on an invoice, and how to distinguish a purchase order number from a vendor reference.
Document Intelligence and OCR extends vision specifically to structured documents — contracts, invoices, forms, expense receipts. Where basic OCR simply converts pixels to text, modern document AI understands document structure: it knows that the number in the top right of an invoice is probably the invoice number, and that the table in the middle contains line items. Oracle’s AP Invoice Automation uses document intelligence to extract fields from supplier invoices before matching them to purchase orders.
Natural Language Processing (NLP) enables machines to understand human language — written or spoken. This underpins contract clause extraction (understanding what a sentence in a contract means legally, not just what words it contains), sentiment analysis in CRM, and the voice input features appearing across Oracle’s newer interfaces in 26A.
Speech Recognition converts spoken language to text, and increasingly to intent. Oracle is integrating voice input into several Fusion workflows in 26A — the ability to dictate into a submission form, or navigate by voice — built on speech AI that has learned the acoustic patterns of language.
Key characteristic: Perception AI takes something a human can see or hear and makes it machine-readable. It is the input layer — the foundation everything else builds on. AI that cannot read an invoice, understand a contract, or interpret a query is AI that cannot act on the world.
Once data is structured — whether extracted by Perception AI or already structured in your Fusion data model — Analytical and Predictive AI finds patterns in it and uses those patterns to generate scores, forecasts, anomaly flags, and recommendations.
Supervised Machine Learning is the most widely deployed form of AI in enterprise software today, and the type most prevalent across Oracle Fusion 25D and 26A. It works by learning from labelled historical examples: past invoices marked as matched or unmatched, past employees marked as having left or stayed, past journal entries marked as anomalous or clean. The output is typically a probability or a risk score. Oracle uses supervised ML across Financials (Smart Close, AP matching), HCM (attrition prediction, candidate scoring), Procurement (supplier risk), and EPM (variance flagging). It is the workhorse of Fusion AI.
Unsupervised Machine Learning finds patterns without labelled training data. Rather than learning from examples of good or bad, it groups similar items together and flags things that do not fit any group. In a Fusion context, this is most visible in expense audit AI — identifying claims that are statistically unusual relative to the population rather than matching to a known-bad list.
Predictive Analytics uses statistical and ML models to forecast future values from historical patterns. Oracle’s cash flow forecasting, demand sensing, and EPM predictive planning all fall here. The model learns the historical relationship between inputs and outcomes, then extrapolates forward.
Prescriptive AI goes one step further: rather than telling you what will happen, it tells you what to do. Collections AI that recommends which accounts to prioritise, resource optimisation that recommends workforce assignments, and next-best-action in CRM are all prescriptive. This is the AI most directly embedded in workflows, because it produces an action rather than just an insight.
Key characteristic: Analytical and Predictive AI operates on structured data, requires historical data to learn from, and its quality degrades when data quality is poor or data history is thin. This is why data maturity is a recurring theme across this series — without clean, complete, historical data, these models cannot learn effectively.
The newest and fastest-moving branch. Where Analytical AI observes and predicts, Generative AI creates. Where Predictive AI produces a score, Agentic AI takes a series of actions.
Large Language Models (LLMs) and Generative AI learn from vast amounts of text and can produce new text — coherent, contextually appropriate, human-quality — in response to a prompt or a structured input. In Oracle Fusion 25D and 26A, generative AI appears in variance commentary, sales proposal drafting, and candidate communications. The model does not look up a template — it generates the text from scratch based on what it has learned about language and what it knows about the context.
“Generative AI can produce confident-sounding text that is wrong. Human review of generative AI outputs is not optional — it is a governance requirement.”
It is important to understand that generative AI does not “know” things in the way a database does. It has learned statistical relationships between words and concepts, and generates outputs that are statistically consistent with those relationships. This means it can produce confident-sounding text that is wrong. Human review is not optional — it is a governance requirement. Oracle’s implementation reflects this: generated commentary is presented as a draft for human review, not a final output.
Multimodal AI combines multiple input types — text, images, audio — within a single model. An AP automation workflow that reads a scanned invoice image (vision), extracts the vendor name (NLP), cross-references against a supplier database (retrieval), and drafts a query email if there is a discrepancy (generative) is a multimodal workflow. Oracle’s document processing increasingly blends these capabilities across Procurement and Financials in 26A.
AI Agents represent the most significant architectural shift in enterprise AI. An agent takes a sequence of actions autonomously to achieve a goal — calling tools, querying data, making decisions, producing outputs — without a human in the loop for each step. Oracle introduced agentic capabilities in 25D and extended them in 26A: the Agentic Planning Assistant in EPM can receive a natural language instruction and autonomously create, version, and populate planning scenarios. Service resolution agents in CX can handle a complaint from receipt to resolution recommendation without human management at each step.
Agents introduce governance requirements that do not exist for predictive AI. A predictive model produces a score; a human decides what to do with it. An agent takes actions. The audit trail, override protocols, and escalation paths for agentic AI are fundamentally different — the series returns to this in the Governance layer (Articles 19 and 20).
Key characteristic: Generative and Agentic AI produces outputs that did not exist before — text, decisions, actions. This makes it the most visible and the most consequential branch, and the one that requires the most careful governance framework.
03 What AI Is Not
Three categories that frequently get labelled as AI in vendor conversations, and are not.
Rules-Based Automation
Executes workflows you define explicitly. An approval routing rule that escalates invoices over £50,000 to the CFO is not AI — it is a conditional statement. Workflow automation, approval hierarchies, and threshold-based alerts in Oracle are not AI features, even when described as “intelligent” in marketing materials.
Test: could a human write it as an if-then statement? → Automation.
Traditional BI & Reporting
OTBI reports, FRS financial statements, and Analytics dashboards are not AI — they are visualisation and aggregation tools. They make data visible; they do not learn from it or make predictions. Oracle Analytics is the platform that can host AI-generated insights, but the visualisation layer itself is not AI.
Test: does it learn from data over time? → If no, it is reporting.
Robotic Process Automation (RPA)
RPA mimics human interaction with interfaces — clicking buttons, copying data between systems, filling in forms. It is scripted automation. It breaks when the interface changes. It does not learn. Distinguishing RPA from genuine ML-based automation matters particularly in Procurement and Finance, where vendors frequently conflate the two.
Test: does it improve as more data accumulates? → If no, it is RPA.
The practical test for all three: does the system learn from data and improve its outputs as more data accumulates? If yes, it is AI. If it executes a static logic tree, it is automation.
04 How Oracle Layers These Technologies Together
The three branches do not operate independently inside Oracle Fusion. They are layered — each level building on the one below it, and multiple levels combining within a single feature. The diagram below shows Oracle’s AI architecture as a stack, from the data foundation upward through Perception, Analytical, Generative, and Agentic layers.
Oracle Fusion AI Architecture Stack — Releases 25D / 26A. Each layer builds on the one below it; most Fusion AI features combine two or more layers simultaneously. Full platform architecture at simplitconsulting.com/architecture →
Reading the stack in practice through three real Fusion features:
AP Invoice Automation 25D / 26AOracle docs ↗
L1Perception AI reads the invoice image and extracts structured fields — vendor, amount, line items, PO reference.
L2Supervised ML matches extracted data to purchase orders and scores match confidence.
L2High confidence → auto-posts. Low confidence → routes to human review.
L4In 26A, an emerging agentic layer begins handling defined exception types autonomously.
Smart Close 25D / 26AOracle docs ↗
L2Supervised ML scores every journal entry against historical patterns, flagging statistical anomalies before close.
L3Generative AI drafts period-end commentary explaining the flagged variances in plain language for the finance team.
Agentic Planning Assistant 26A PreviewOracle docs ↗
L1NLP interprets a natural language instruction from an FP&A planner.
L2Predictive ML draws on historical actuals and identified planning drivers.
L3Generative AI drafts scenario narratives and commentary.
L4The agent layer orchestrates all three autonomously across multiple planning model dimensions.
This stacking pattern — perception feeds analytical, analytical feeds generative, generative feeds agentic — is the architectural logic of Oracle’s AI roadmap. Features that appear to be “just” document reading today are being built as Layer 1 foundations for agentic workflows in the near future.
Where OCI and EBS sit in this picture: Oracle Cloud Infrastructure (OCI) AI Services is the platform layer below Fusion — it provides the GPU compute, base model training, and AI services (Vision, Language, Document Understanding, Speech) that Fusion’s embedded AI is built on. On-premise EBS customers can access some of these capabilities via OCI integration, though the embedded AI features covered in this series are specific to Oracle Fusion Cloud. Future articles will address OCI AI Services and the AI strategy for hybrid EBS/Fusion estates.
05 The Four Types You Will Encounter Most in Fusion
Zooming from the full taxonomy back to what matters most for day-to-day Fusion work in 25D and 26A:
Type
What it does in Fusion
Data requirement
Key risk
Document & Language AI
Reads invoices, contracts, expense receipts, free-text fields. Enables AP automation, contract clause extraction, expense audit.
Document volume and format variety
Extraction errors on non-standard document formats
Predictive & Scoring AI
Forecasts values, ranks items, flags anomalies. Powers attrition, cash flow, supplier risk, demand sensing, GL anomaly detection.
Clean historical data at sufficient volume
Poor predictions when data history is thin or quality is low
Generative AI
Drafts text, explanations, structured content. Powers variance commentary, proposal drafting, candidate communications, board-pack narrative.
Structured context data to ground generation
Plausible but incorrect outputs — human review is mandatory
Agentic AI
Takes sequences of actions autonomously. Powers planning assistant, service resolution, emerging autonomous processing workflows.
All of the above, plus reliable tool integrations
Actions with real consequences — governance and audit trail are non-negotiable
These four types map directly onto Oracle’s module-by-module AI feature set, which Article 2 covers in full.
06 What This Means for Your Role
Business & Functional Users
Finance Directors, HR leads, operations managers. The most important thing this article gives you is the ability to ask the right questions. When a vendor tells you a feature is “AI-powered”, ask which type. A predictive model that scores attrition risk is only as good as the HR data you have been maintaining for the past three years. A generative model that drafts variance commentary still needs a human to review it before it goes to the board. Understanding the type tells you what the feature requires from your organisation — and what it cannot do on its own.
IT & Configuration Teams
System administrators, technical leads. The taxonomy maps directly to what you need to configure and govern. Perception AI features require document format testing and extraction validation. Predictive AI features require data quality assessment before enablement. Generative AI features require output review workflows and user training. Agentic AI features require audit trail configuration, role-based access controls, and escalation paths. Each type has a distinct implementation checklist — the module articles in this series provide them.
Consultants & Architects
Solutions architects, implementation consultants. This taxonomy is the framework for every AI conversation with a client. Before recommending any feature, know which branch it sits in, what data prerequisites it carries, and what governance it requires. Clients who enable predictive AI on thin data get poor predictions and lose trust in the feature within weeks. The right question is never “should we turn this on?” — it is “are the conditions in place for this type of AI to work?”
Continue your reading path
07 What Comes Next
The taxonomy in this article — Perception, Analytical and Predictive, Generative and Agentic — is the vocabulary the entire Oracle Fusion AI Series uses. Every subsequent article names which AI type it covers, what it requires, and what it delivers.
Article 2 takes this taxonomy and maps it directly onto your Oracle Fusion tenant as it stands in releases 25D and 26A. It answers the question every implementation team eventually asks: of everything Oracle calls AI across Financials, HCM, Procurement, EPM, CRM, and Projects — what is actually live today, what is genuinely useful, and what still needs the right conditions to deliver?
Coming later in this series: Oracle’s AI capabilities on OCI AI Services, the AI roadmap for organisations on Oracle EBS, and cross-platform AI governance considerations that apply regardless of which Oracle product you are running. Those articles build on the foundation established here.
About the Oracle Fusion AI Series — Published by Simpl’IT Consulting. 20 articles covering Foundation concepts, Module deep-dives, Industry verticals, and Governance across Oracle Fusion Cloud 25D / 26A. Series hub and all articles →