The History of Time (AI)

“We need to be using AI.”

If you’ve sat in a leadership meeting in the last two years, you’ve probably heard some version of that sentence. It sounds decisive. But ask the obvious follow-up (what does that actually mean?) and the room often goes quiet.

That isn’t a knock on anyone. “AI” has become one of the loosest terms in business. People use it to describe a spam filter, a fraud detection model, a chatbot, and an autonomous agent that books travel, all in the same conversation. When one word covers that much ground, confusion is almost guaranteed, and confusion gets expensive.

Enterprise AI adoption shows a widening divide: according to McKinsey, 40% of large organizations (over $1 billion in revenue) are now scaling AI agents, up from 27% a year earlier, while smaller organizations have stayed flat at 22%. Part of this divide comes from using the wrong kind of AI for the job.

The key idea is simple: AI isn’t one thing. It’s four distinct waves that have built on each other for more than 70 years:

1. Machine learning, which learns patterns from structured data
2. Deep learning, which uses neural networks to learn from larger and messier data
3. Generative AI, which creates new content such as text, images, and code
4. Agentic AI, which plans and carries out tasks on its own

Each wave solves a different kind of problem, depends on a different kind of data, and carries a different cost. Once you know which one you need, you can match the right approach to the right business problem and avoid paying premium prices for work a simpler tool could handle. Here’s how each wave works, and where it fits in B2B commerce.

Wave 1: Machine Learning, the Quiet Workhorse

AI didn’t appear out of nowhere in 2022. Its roots go back to the 1950s, when researchers first began teaching computers to improve through experience. Arthur Samuel, who built a checkers program at IBM that got better the more it played, is widely credited with coining the term “machine learning” in 1959.

Machine learning is a set of algorithms that learn patterns from data instead of following rules a programmer writes line by line. It works best with structured data, the kind that fits neatly into rows and columns, like a spreadsheet of orders with clear fields for customer, product, quantity, and price.

You already use it every day. Spam filters learn what junk mail tends to look like. Recommendation engines notice that buyers of one product often buy another and suggest it to the next shopper.

Where It Fits in B2B

Most B2B companies already hold years of structured data in their ERP, CRM, and eCommerce systems, which makes machine learning a practical starting point. Common uses include:

• Predicting reorders by learning each account’s buying rhythm
• Forecasting demand from order history and seasonality
• Recommending products that buyers typically add to an order
• Spotting fraud and risk in orders and payments

That last use can deliver results at a large scale. In fiscal year 2024, the U.S. Treasury reported that its enhanced fraud detection processes, including machine learning, helped it prevent and recover more than $4 billion in fraud and improper payments, up from about $653 million the year before.

Why the Oldest Wave Still Matters

With so much attention on chatbots and agents, it’s tempting to treat machine learning as old news. For many problems, it’s still the best choice. Many newer AI tools are priced by usage, so running a large model on a simple, repetitive task gets expensive fast. A well-built machine learning model is usually far cheaper to run, produces consistent results, and makes it easier to see which factors drove a decision.

Wave 2: Deep Learning, the AI That Teaches Itself

Machine learning has a limit. Most of the world’s information isn’t tidy. Photos, scanned documents, audio, and free-form text don’t fit neatly into rows and columns.

The second wave took off in 2012, when an image recognition model called AlexNet beat its competition by a wide margin and showed what neural networks could really do. Deep learning is a subset of machine learning that uses these networks, loosely inspired by the human brain, with data passing through many layers that each pick up on something a little more complex.

The biggest difference is who writes the rules. Suppose you want a system to recognize a tail in a photo. In the first wave, a person defines what a tail looks like. In the second, you show the model millions of labeled photos, and it figures out what a tail looks like by itself.

Where It Fits in B2B

That ability to learn from raw examples opens up use cases structured data alone can’t support:

• Visual quality inspection that spots defects on a production line
• Document processing that reads purchase orders and invoices and routes the details into your ERP
• Predicting equipment failures from vibration and temperature data
• Predicting churn by combining orders, support tickets, and website behavior

Deep learning is also a powerful tool for high-stakes decisions such as fraud detection. Visa says it helped block $40 billion in fraudulent activity in a single year, nearly double the year before, and it has expanded real-time risk-scoring services built on deep learning models.

The Catch: It Depends on Good Data

Hand a model millions of clean, labeled photos and it will learn what a tail looks like. Hand it a disorganized pile of PDFs and recordings with no context, and it will struggle. Many companies aren’t ready yet. In a 2024 survey of data management leaders, Gartner found that 63% of organizations either lack the right data management practices for AI or aren’t sure whether they have them.

In our experience, there are two common ways to close that gap: integrate the systems where your data already lives, such as your ERP and CRM, or clean the data a project needs and move it into a dedicated data lake with clear governance rules.

Wave 3: Generative AI, Fluent and Occasionally Wrong

For most people, this is the wave that made AI real. Its foundation was laid in 2017, when Google researchers introduced the transformer architecture, but it was the public release of ChatGPT in late 2022 that turned generative AI into a household term.

The first two waves analyze and categorize data. Generative AI creates it: text, images, code, and answers to questions. Under the hood, it’s a prediction engine. It learns which words tend to follow which, then builds a response one likely word at a time. The result feels remarkably conversational, and you don’t need technical skills to use it.

Where It Fits in B2B

Generative AI is most useful wherever teams spend time writing, searching, or answering the same questions repeatedly:

• Product content such as descriptions and spec summaries for large catalogs
• Search-friendly content optimized for both search engines and AI answer engines
• Quotes and sales emails drafted from a customer’s history
• Internal knowledge assistants that answer policy and product questions in seconds
• Customer service that handles common questions around the clock

Amtrak’s virtual assistant, Julie, is a well-known service example. According to a case study from Verint, which built the system, Julie answers more than 5 million questions a year and has delivered an 8x return on investment.

The Downside: Hallucinations

Because generative AI predicts the most likely answer rather than looking up a verified fact, it can produce information that sounds confident but is wrong. These mistakes are called hallucinations, and in business the quiet ones cause the most trouble: an invented specification, a made-up policy, a source that doesn’t exist.

The risk grows when companies point AI at everything they’ve ever stored. Research commissioned by Seagate and conducted by IDC found that 68% of the data available to businesses goes unused. Picture a closet packed with years of files, some current, some outdated, some confidential. Giving AI unrestricted access is like inviting in an assistant with a photographic memory and no sense of what should stay private.

The Fix: Start Small and Ground the Answers

Trying to clean up everything first can stall a project for months, and many AI efforts never recover. A 2025 S&P Global Market Intelligence survey, reported by CIO Dive, found that 42% of companies had abandoned most of their AI initiatives, up from 17% the year before.

A focused approach works better. Pick one project, identify only the data it needs, apply your governance rules, load that clean data into a controlled data lake, and instruct the AI to answer solely from it. This is called retrieval-augmented generation, or RAG, and because answers are grounded in data you control, the risk of hallucination drops significantly. When the AI can’t answer something, that usually points to the next piece of data worth adding.

Wave 4: Agentic AI, When AI Stops Answering and Starts Doing

Every wave so far helps people decide or create, but a person still takes the next step. Agentic AI changes that. It can reason about a goal, plan the steps to reach it, and carry them out across different systems.

A canceled flight shows the difference. Generative AI can tell you the flight was canceled and list your options. An agent finds the next available flights, picks the one that matches how you usually travel, holds the seat, updates your calendar, and asks for a simple “yes, book it.”

Adoption is climbing quickly. In August 2025, Gartner predicted that 40% of enterprise applications will include task-specific AI agents by the end of 2026, compared with fewer than 5% in 2025. Gartner also cautioned against “agentwashing,” labeling an ordinary assistant as an agent, which is one more reason to know exactly what you’re buying.

Where It Fits in B2B

Agents are most valuable where a process spans several systems and involves a lot of routine follow-through:

• Resolving order exceptions by finding substitutes or stock in other warehouses
• Replenishing inventory by preparing purchase orders for a buyer to approve
• Following up on quotes before they go cold
• Handling returns and service requests from start to finish

Customer service is moving especially fast. In March 2025, Gartner predicted that by 2029, agentic AI will resolve 80% of common customer service issues without human intervention, cutting operational costs by 30%. Saks Fifth Avenue, for example, launched an AI voice assistant named Sophie in 2025 to handle requests such as order status and returns, reducing call volume for human agents by 20%.

Agents Need Guardrails

Agents connect to other systems through APIs, which act as their hands. That power is exactly why governance matters. Suppose the only seat left on that rebooked flight costs $12,000 in business class, and the original ticket was a $500 economy fare. A person would stop and ask. An agent stops only if it’s been told to.

Good guardrails include spending limits, approval steps for higher-risk actions, access restricted to only the systems each agent needs, and audit trails that record every action.

Think of Agents Like a Bank

When you walk into a bank, a greeter confirms who you are and sends you to the right specialist: a teller for deposits, someone with vault access for your safe deposit box, or a loan officer for a refinance. Each role is narrow, and each person can reach only what their job requires. Well-designed agent systems work the same way, with specialized agents coordinated by one that routes the work. That design is easier to secure, test, and fix.

Bringing the Four Waves Together

The newest wave doesn’t replace the older ones. Each still has its place, and the most cost-effective approach is to match the wave to the work.

Wave	Best for	B2B example
Machine learning	Predictions from structured data	Reorder alerts, demand forecasting
Deep learning	Unstructured data and risk	Invoice processing, fraud detection
Generative AI	Content and communication	Product descriptions, customer chat
Agentic AI	Multistep workflows	Order exceptions, quote follow-up

The waves also work well together. An agent might run a deep learning model to check an order for fraud, then hand the result to generative AI to draft a message to the customer. For content work, an agent can gather research and pass it to generative AI to shape into an article. Knowing the difference is what lets you combine them on purpose.

Your Data Is the Real Advantage

Eventually, nearly every company will use AI, including your competitors. What they won’t have is your data: years of customer interactions, order history, pricing, and service records.

For decades, businesses chased the promise of business intelligence but lacked the tools to analyze everything they collected. AI can finally work through that history, find patterns, and act on them. The companies that organize and connect their data will get the most from every wave.

---

Conclusion: Start Small, Choose Deliberately

“We need to use AI” isn’t a strategy. “We need machine learning to predict reorders” or “we need an agent to resolve order exceptions” is.

The path forward is the same no matter which wave you start with. Pick one well-defined business problem, choose the wave that fits it, get the right data clean and connected, define how you’ll measure success, and prove the value before you expand. AI is moving quickly, and a trusted partner can help you sort through the options and build a roadmap that fits your business.

Ready to transform your B2B eCommerce experience? Let us help you align your technology with your business goals. Reach out to learn more, or check out our blog for insights on digital transformation and eCommerce trends.