About the Author
Predictive models and generative AI are both AI. Knowing which fits your question is what turns data into decisions.
Recently, I’ve had many clients ask me some version of the same question: “Why bother implementing predictive machine learning models in Fabric when I can just ask AI?”
It’s a fair question, and a good sign. A couple of years ago, most of the leaders I worked with were still asking whether AI was relevant to their business at all. Now they’ve used ChatGPT, Claude or Copilot, they’ve seen it draft an email or summarize a long document in seconds, and they reasonably wonder why anyone would spend weeks building a forecasting model when a chatbot will happily give them a forecast in a sentence.
The short answer is that you’re already talking about AI either way. Predictive modeling is AI. The real question is which kind of AI fits the problem in front of you — and getting that right is the difference between a number you can plan a business around and a number that just sounds convincing.
Here’s how I explain it to clients, and how I’d suggest you decide where to start.
Two Kinds Of AI, Built For Two Kinds of Questions
When most people say “AI” today, they mean generative AI: large language models such as ChatGPT, Claude and Copilot. These models are trained on enormous amounts of public or private text and data and are remarkably good at working with language. They read, summarize, extract, translate, code, draft and explain. Ask one what a contract says, what customers complained about on last month’s sales calls, or how to word a difficult message to a supplier, and you’ll usually get a useful answer quickly.
Predictive machine learning is a different branch of the same family. A predictive model is trained on your historical records — years of orders, leads, projects, service visits, weather, staffing, whatever is relevant — and learns the patterns that connect what you knew at a point in time to what actually happened next. Its output isn’t prose. It’s a number: how many, how likely, when, how much. And critically, it comes with a measured margin of error, because the model has been tested against outcomes it never saw during training.
A simple way to remember the distinction: if the answer is language, think generative AI. If the answer is a number you have to trust, think predictive model.
Why “Just Ask AI” falls Short For Forecasting
Nothing stops you from pasting last year’s sales into a chatbot and asking for next quarter’s forecast. You’ll get an answer. The problem is what you can’t get alongside it.
Consistency
Ask the same question twice, or have two people on your team ask it slightly differently, and a language model may give you two different numbers. That isn’t a flaw. Generative models are designed to produce plausible, varied language and visuals. But a forecast that changes depending on who asked and when isn’t a forecast that you can hand to finance, build a budget on, or hold anyone accountable to. One of my client BDMs put it perfectly: he didn’t want a different answer today than tomorrow; he wanted something organized enough to produce consistent results. A trained predictive model gives the same output for the same inputs, every time.
Proven Accuracy
Before a predictive model goes anywhere near a decision, it’s back-tested: trained on history up to a point, asked to predict the period after, and scored against what really happened. You know its error rate by division, by month and by forecast horizon, and you can compare it directly against whatever spreadsheet or rule of thumb you use today. A chatbot can’t tell you how accurate its forecast of your business is, because it has never been measured against your outcomes.
Data At Real Scale
Useful predictions often depend on millions of data rows: every lead, every appointment, every sale, every project status change over several years, joined to weather, geography, capacity and marketing spend. GenAI works with what fits in a prompt or a retrieved / aligned set of documents. It isn’t built to learn statistical patterns across your full operational history and data estate.
Explainability You Can Check
A good predictive models can show which factors drove each prediction — seasonality, lead source, how quickly a customer was contacted, how much product was in inventory vs. backordered. Those drivers are calculated, not narrated. A language model will explain its reasoning fluently, but that explanation is generated text, not a reliable account of how the number was produced.
Ownership, Security, and Cost
When you build predictive models in Microsoft Fabric, they run where your data already lives. Training and scoring happen in your lakehouse / warehouse, on capacity you already pay for, under the security and governance you’ve already put in place. You own the model and retrain it on your schedule. Prompting an external AI service for every forecast means sending data out per query, paying per use, and depending on a vendor model that can change underneath you without notice.
None of this makes generative AI the wrong tool. It makes it the wrong tool for this job.
What Predictive Models Look Like in Fabric
One reason clients hesitate is that “machine learning” sounds like a separate science project. In Fabric, it isn’t. It extends the data platform you’ve already built:
- Features from gold data. Notebooks reshape trusted, governed tables into the inputs a model learns from, calculated as of each point in history so the model never sees the future (learns honestly).
- Training and tracking. Models are trained in Fabric notebooks and every run is recorded with MLflow — the parameters, the accuracy, the data snapshot it used. That’s what lets someone retrain it reliably a year from now.
- Scheduled scoring. A pipeline scores the latest data nightly or weekly and writes predictions back to gold tables, labelled with the date and model version that produced them.
- Reporting where people already work. Predictions flow through your semantic model into Power BI, shown alongside actuals, with ranges, under the same row-level security as everything else.
The output isn’t a one-off answer in a chat window. It’s a new, governed dataset — one your business has never had before — that refreshes itself and can be measured over time.
Where Generative AI Belongs
Don’t get me wrong – I’m certainly not arguing against GenAI. I’m enthusiastic about it, and most of my clients are (or should be) using more of it. It’s just strongest on problems that are fundamentally about language:
Unstructured Content
Summarizing sales calls, service notes, emails and survey comments; pulling themes and objections out of thousands of conversations
Document Work
Extracting fields from contracts, invoices and forms; checking documents against rules; comparing versions
Plain-Language Access To Data
Copilot in Power BI and Fabric Data Agents that let a manager ask “why is this region down this month?” and get pointed to the right figures
Narrative
Drafting monthly commentary, executive summaries and customer communications
Code
Writing, explaining, and debugging application source code
Notice that several of these sit on top of structured data rather than replacing the work of producing it. That’s the key insight
Better Together
The most valuable setups I see don’t choose one or the other. They layer them.
Predictive models produce the numbers. Generative AI helps people use them.
A demand model forecasts next season’s sales by region, with a range. An item return rate model flags which shipped SKUs are most at risk of return requests. Those predictions land in governed tables. Then GenAI makes them accessible: a regional sales manager asks in plain English why their forecast dropped, Copilot surfaces the drivers the model calculated, and a monthly summary explains what changed — grounded in numbers that were measured, not invented.
GenAI can also feed their predictive side. Turning call transcripts or free-text notes into structured signals — did the customer mention budget, timing, a competitor? — creates new features a predictive model can learn from. Each makes the other more useful.
How To Decide Where To Start
Your immediate needs should drive the first step. Here’s the guidance I give clients.
Start With Predictive Models If…
- You plan, staff or budget around numbers that are currently produced in spreadsheets by individuals
- Different people produce different forecasts, and leadership doesn’t know which to trust
- Your forecasts rely on conversion rates or factors that are reset once a year and go stale in between
- A wrong number has a real cost: missed hiring windows, idle crews, excess inventory, a budget that starts the year behind
- You already have a reasonably clean, governed lakehouse with several years of history
Start With GenAI If…
- Your biggest pain is volume of unstructured content — documents, calls, emails — that people read and re-key by hand
- Your users can’t find or interpret the reports you already have
- You need quick wins to build momentum while longer data work is under way
Fix The Foundation First…
- Your core data isn’t yet centralized, trusted or consistently defined
Both kinds of AI are only as good as the data underneath them. A predictive model trained on inconsistent definitions will learn the inconsistency. A chatbot pointed at ungoverned data will confidently repeat it. If this is where you are, the most valuable AI investment you can make is a sound lakehouse
Practical Steps For A First Predictive Model
If predictive modeling is the right starting point, keep the first one small and measurable:
- Write a clear problem statement. Name the decision the prediction will support and who makes it. “Forecast monthly sales by region, six months ahead, to set hiring plans” beats “use AI on our sales data.”
- Establish the baseline. Rebuilt today’s spreadsheet logic first. It proves the data is right and gives you the benchmark the predictive model has to beat.
- Separate belief from evidence. Every business has tribal knowledge about what drives results. Treat those beliefs as hypotheses for the model to test, not facts to hard-code.
- Validate in two directions. Back-test on a recent year you hold out, then run the model alongside reality for a few months before anyone relies on it.
- Plan for ownership. Document each model with a “model card” and a retraining runbook so your own team can operate it after launch.
A focused first model that measurably beats the spreadsheet does more for adoption than an ambitious one nobody trusts.
The Bottom Line
“Why build predictive models when I can just ask AI?” is the right question to ask, as long as you follow it with a second one: what kind of answer do I need?
If you need language — summaries, extraction, drafting, explanation — generative AI is often the fastest, most effective tool available. If you need a number you’ll plan, staff and budget around, you need a model trained on your own history, tested against your own outcomes, and running inside the platform where your data already lives. For most organizations, the best path is both, in the right order.
If you’re weighing where to begin, I’m always happy to talk it through. The goal is the same either way: getting the most out of the data you already have.
Frequently Asked Questions
No. A generative tool writes a plausible answer based on what you put in the prompt. A predictive model is trained on your full historical data and tested against outcomes it never saw. That testing what gives you a measured error rate you can plan around. Both are AI, but they’re built for different jobs
It depends on the question, but a good rule of thumb is several years of consistent history. That’s enough to capture seasonality and at least a few business cycles. Consistency matters more than volume. Three years of cleanly defined data usually beats ten years where definition changed along the way.
A focused first model on an existing, governed lakehouse typically takes weeks, not months. The bigger variable is data readiness. If your core data still needs to be centralized and cleaned up, that foundation work comes first and will take longer than the model itself.
Not necessarily. Fabric runs models in the same environment as the rest of your data platform, and every training run is tracked with MLflow. With a model card and a retraining runbook, a data or analytics team that’s comfortable with Fabric notebooks can usually handle retraining and monitoring.
Measure both the same way. Rebuild your current spreadsheet logic as a baseline, then back-test the model and the baseline against a held-out period and compare their error rates. Before anyone relies on the model, run it alongside reality for a few months to confirm it holds up.
Yes, and that’s often where the most value is. Predictive models produce governed, measured numbers. Generative AI tools like Copilot in Power BI help people explore and explain those numbers in plain language. Generative AI can also turn unstructured content, such as call notes and emails, into structured signals that improve predictive models.
