Predictive AI Doesn't Compete with GenAI; It Fixes It!

Summary

Most generative AI systems never leave the pilot stage because they make too many serious mistakes, and Eric Siegel argues that predictive AI is the fix.

In this DataCamp webinar, Siegel, CEO of Gooder AI, founder of Machine Learning Week and author of The AI Playbook, joins host Richie Cotton to make the case for predictive AI vs generative AI as partners rather than rivals. Siegel describes a reliability layer: a predictive model that scores every LLM interaction for risk and sends only the riskiest cases to a human reviewer. Twilio, NextGen Healthcare and Instacart already run versions of this design.

He then turns to an older problem. Predictive AI projects also fail to deploy, and Siegel blames the way teams measure them. Data scientists report precision, recall and AUC, while the managers who must approve deployment want to know about profit and savings. Siegel calls the alternative ML valuation and demonstrates it live with profit curves in Gooder AI, using churn and fraud detection examples built on Jev, a foundation model for zero-shot prediction that launched in September 2026.

The session closes with a parallel point about generative AI: public benchmarks such as MMLU say little about how a model will perform on your task. Watch the recording to see the live demo and the audience Q&A on data security, uncertainty and who the tool is for.

Key Takeaways

• Predictive AI and generative AI solve different problems, and Siegel argues most organizations should invest at least as much in predictive AI as in generative AI.
• Generative AI systems that act autonomously with external users often stall at around 95% reliability, because the remaining errors are too costly to accept.
• A predictive reliability layer can score each LLM interaction for risk and route only the riskiest cases, such as the top 20%, to a human reviewer.
• In Siegel's example, that design keeps about 80% of interactions fully automated while cutting the error rate from 5% to roughly 1%.
• Technical metrics such as precision, recall and AUC tell you a model beats guessing, but they do not tell you how much money it will make or save.
• A profit curve plots business value against the share of cases you act on, which lets teams choose a decision boundary in terms executives understand.
• Business inputs such as the cost of a false positive are uncertain, so value estimates should be tested across a range of settings.
• Foundation models like Jev make zero-shot predictive AI possible, letting teams skip building a custom training set for some projects.
• For generative AI, standardized benchmarks are poor proxies for value; test the model on examples from your own use case.

Deep Dives

Predictive AI vs Generative AI: Partners, Not Rivals

Siegel opened by placing the two technologies side by side. Both are built with machine learning, but they do different jobs. Generative AI writes text, code and images. Predictive AI, which used to go by the names machine learning or predictive analytics, estimates the probability of an outcome for each case so an organization can decide how to treat it: who to contact, which transaction to block, which claim to review. Its classic uses include targeting marketing, detecting fraud, scoring credit and predicting equipment failure.

For four years, generative AI has taken most of the attention. Siegel thinks that has left predictive AI undernourished. As he put it, "Predictive AI is older but not old school." Much of its potential value is still untapped, partly because so many projects never deploy. He then made what he called a subversive claim: "most organizations should be investing at least as much in predictive AI as in generative AI." He added that the right choice depends on the problem, and that predictive projects often offer lower-hanging fruit with less complexity.

Siegel used an analogy to explain why the two should not compete. A camera and a telephone serve different purposes, yet combining them gave us the smartphone. Predictive and generative AI work the same way. Each covers gaps in the other, and the most valuable systems will use both.

The rest of the talk followed that logic in both directions. First, predictive models can make LLM systems safe enough to launch. Second, generative AI can make predictive tools easier to use. Gooder AI includes a chatbot that answers questions about a project, which models to compare and where to set thresholds. Siegel summed up the second direction: "Use Gen AI. Use a chatbot to help" business users make sense of the numbers.

Richie Cotton closed the session on the same note, observing that a rivalry has grown between traditional machine learning practitioners and newer generative AI teams. "There shouldn't be. I love that they're very complementary technologies." For data teams deciding where to spend budget and training time in 2026, the message is that a strong predictive modeling practice is not a legacy skill. It is a requirement for getting generative AI into production.

Why LLM Agents Fail to Launch: Mental Scope Creep

Siegel drew a line between two ways people use generative AI. When you use ChatGPT or a similar tool to draft code, documents or messages, you check the output yourself. "You are the human in the loop," he said, and that use case is valuable and well proven. The overhyped case is different. It is the system that acts on its own with someone outside your team, such as a customer service agent talking to the public.

He described a pattern he calls mental scope creep. A chatbot that answers questions from 20 pages of FAQs may be reliable enough to launch without supervision. The trouble starts when a team stretches the plan: 20,000 documents instead of 20 pages, external customers instead of internal staff, financial transactions instead of general information, sensitive health or account data instead of public facts. Each step raises the cost of a mistake. Eventually the system cannot launch because its errors are too grave for the business to accept.

Siegel was careful about the numbers. Suppose an autonomous agent handles customers correctly 95% of the time. "Now 95% correct is incredible," he said, noting that he worked in natural language processing research in the 1990s and never expected to see this level of performance. But for many enterprise uses, a 5% rate of serious failures means the whole project gets shelved. A shelved system delivers nothing.

He also took aim at the word agentic, which he thinks overpromises to non-specialists because it implies granting real agency to a machine. He cited LinkedIn executives who wrote that "AI is not coming for titles. It's only coming for tasks." Siegel said he agrees, though he acknowledged that some in the audience might think generative AI is underhyped rather than overhyped.

The point he wanted everyone to accept, whatever their view on the hype, is that ambition will always run slightly ahead of capability. Teams will keep scoping LLM projects a little too broadly. That makes reliability the core obstacle to deploying AI agents, and it is an obstacle that model improvements alone will not remove. Something has to manage the residual risk, and that is where predictive AI comes in.

The Reliability Layer: Predictive AI as a Guardrail for LLMs

Siegel's answer to the reliability problem is to bring the human back, but only where it matters. Human reviewers cost money, so they need to be used selectively, according to risk. Deciding which cases are risky is a prediction problem. In his words, "We can think of a large language model as this incredibly powerful beast that must be tamed with a reliability layer so that we can make use of it."

This is familiar ground for predictive AI. Risk management has always been its home: fraud detection, hospital readmissions, predictive maintenance, health code violations, and even which manhole covers might explode or which bridges might fail. In each case, inspectors cannot check everything, so a model ranks cases by risk and directs human attention to the top of the list. An LLM in production is one more large-scale operation that benefits from the same approach.

Siegel walked through the architecture. A predictive model monitors every interaction between the LLM system and the external user. If it flags a case as high risk, the case goes to an internal reviewer who can intervene. Suppose the model routes the riskiest 20% of cases to a person. If that person catches most of the problems, the net error rate could fall from 5% to about 1%. The system is now perhaps 80% autonomous and 99% reliable, which may be enough to launch.

That trade, Siegel argued, is the realistic version of the autonomy promise. You do not get the full story the hype describes, but you get most of it, which beats the zero value of a project that never ships. "This is inevitably the next killer app" of predictive AI, he said.

Companies already use this design. Twilio applies it to a customer service agent. NextGen Healthcare fills in insurance claims automatically, then has a human review the ones a model predicts the payer is likely to deny. Instacart picks a replacement product when an item is out of stock, but only swaps it in automatically when a model predicts the customer will accept it, since a rejected substitution is costly. Siegel noted that the same reliability modeling can also be built with traditional custom models or with newer foundation models, depending on the data.

How to Measure the Business Value of Machine Learning Models

Predictive AI has its own deployment problem. Siegel cited a Rexer Analytics survey in which only 22% of data scientists said their new predictive initiatives usually deploy. He described the pattern as reaching the first kind of excitement, seeing that the technology works, without reaching the second, seeing it deliver value in operations. "It's like we're more excited about the rocket science than the launch of the rocket."

A core cause, he argued, is that teams answer the question of how good a model is with the wrong numbers. Data scientists are trained to report precision, recall, AUC and accuracy. These metrics show the model predicts better than random guessing, which matters. Siegel calls this the prediction effect: "A little prediction goes a long way." But they say nothing about how much profit or savings the model will produce. "If you're not measuring value, you're not pursuing it," he said, and stakeholders presented with technical jargon often feel lost. "They should be expecting the data scientists to talk business."

His alternative is ML valuation alongside ML evaluation. The main tool is a profit curve, also called a value curve. Cases are sorted from most to least likely to respond, churn or commit fraud, according to the model's score. The x-axis is the share of cases you act on. The y-axis is a business metric such as profit or savings. In a marketing example, value climbs steeply at first as you reach the most likely buyers, then flattens and falls as you contact people unlikely to respond.

That shape shows where to set the decision boundary. In a direct mail example, contacting the top 24% maximized profit, while going to 73% would break even, which might suit a team trying to build awareness of a new product. For fraud, blocking too few transactions lets fraud through and blocking too many annoys legitimate cardholders. As Siegel put it, "You don't wanna treat too many. You don't wanna treat too few."

Siegel noted that this kind of chart has been possible since the 1960s, when banks first used linear models for credit and fraud. Yet it is still rare in practice. Every predictive AI project, he said, should be able to show its value in a two-dimensional chart where both axes mean something to the business.

Hands-On with Jev and Gooder AI: Zero-Shot Prediction and Uncertainty

For the live demo, Siegel used Jev, a foundation model built on large language models but adapted for predictive tasks. It outputs a probability rather than text. Jev launched on September 15, 2026, drew copycats from OpenAI, Amazon, Nvidia and Databricks within weeks, and raised $870 million at a $7.5 billion valuation. "Predictive AI's back, baby," Siegel said, though he stressed that he was not promoting Jev over its many competitors, some of them free and open source.

What these models offer is zero-shot predictive AI. Instead of preparing a labeled training set and building a custom model, a team can ask the foundation model a question about each case. In the demo, Jev read customer support tickets and answered whether each customer was expressing an intent to cancel, request a refund or discontinue service. That is a churn model built without training data. A second variant, which Siegel called aggressive Jev, was told to treat missed churners as five times as costly as false alarms.

Gooder AI then plotted both models on a profit curve. Two sliders set the business inputs: the value gained when a retention discount reaches a customer who was about to leave, and the cost when the discount goes to someone who was staying anyway. Moving the sliders reshaped the curves and moved the best decision boundary. In one setting, sending the offer to the top 4.41% of tickets produced the highest estimated return.

Siegel explained why the interface is interactive. "There are always business factors that inform the estimated business value," and those factors are uncertain. Testing a range of values shows whether the choice of model and threshold holds up. In a fraud example comparing a random forest and two XGBoost models, changing the cost of a wrongly blocked card moved the dollar totals a lot but did not change which model or threshold won.

Cotton asked about data security. Siegel said Gooder AI runs locally in the browser, with an offline version for banks that want more control. It needs only scored test data: labels, model scores and perhaps one or two fields such as transaction size. "It doesn't care how the model works. It only cares how well the model works." Siegel said the main users are data scientists who already have test data, though managers comfortable reading a chart can use it too.