The Four Bands of AI

How to Evaluate Agentic-AI Solutions and Future-Proofing Your Stack

There’s a diagram circulating that attempts to explain “Agentic AI” using concentric rings. It’s useful. The inner ring is traditional machine learning — pattern detection, forecasting, recommendations. The next ring out is generative AI — the ChatGPTs and Claudes that produce text and code. Then AI agents that execute discrete tasks. And finally, at the outer ring, agentic AI that automates entire processes.

Most people look at this diagram and see a progression. Newer is better. Outer is more advanced. That’s not quite right.

What the diagram actually shows is a stack. And understanding where a vendor operates in that stack tells you what problems they can actually solve for you.

The Four Bands

Band 1: Machine Learning (ML). This is the foundation — neural networks, regression models, classification algorithms. The stuff that’s been around for decades, super-powered in recent years. Netflix recommendations. Stripe’s fraud detection. Amazon’s demand forecasting. These systems don’t generate prose. They calculate. They find patterns humans can’t see and make predictions humans can’t make.

Band 2: Generative AI. Large language models. ChatGPT, Claude, Gemini. These systems are remarkable at language — writing, summarizing, explaining, coding. They power everything from marketing copy tools like Jasper to image generators like Midjourney. They’re also famously unreliable at math.

Band 3: AI Agents. Systems that can take actions, not just produce outputs. Microsoft Copilot updates your CRM. Intercom’s Fin resolves support tickets without escalation. These agents execute tasks autonomously within defined boundaries — retrieving data, updating records, sending notifications.

Band 4: Agentic AI. Full process automation. Multiple agents coordinating. Complex workflows with decision points, error handling, and human oversight. GitHub Copilot writes 40% of code on the platform. Harvey AI automates end-to-end legal workflows. These systems don’t just assist — they own entire processes.

What This Means For Vendor Evaluation

When you’re evaluating AI vendors, the first question should be: which bands do they actually operate in?

Most vendors today are Band 2 companies with Band 4 marketing. They’ve built a chat interface on top of an LLM and call it “decision intelligence” or “agentic AI.” The interface is real. The underlying capability often isn’t.

Here’s how to tell the difference.

Band 2 vendors can summarize your data, generate reports, answer questions about documents, and create content. If you ask them to analyze something quantitative — pricing optimization, demand forecasting, risk assessment — they’ll produce confident, articulate answers. Those answers may be wrong. LLMs don’t calculate; they are masters of language and visualization.

Band 3 vendors can do everything Band 2 does, plus take actions in your systems. They integrate with your CRM, ERP, or data warehouse. They can retrieve information, update records, trigger workflows. The test: can they do something, or just say something?

Band 4 vendors can orchestrate multi-step processes with multiple agents, handle exceptions, maintain state across sessions, and deliver complete outcomes rather than intermediate outputs. The test: do they replace a process, or just assist with tasks within a process?

Band 1 capability is the hardest to evaluate and the most important for analytical use cases. Ask: what models run under the hood? Where does the quantitative analysis actually happen? If the answer is “the LLM figures it out,” be skeptical.

The Domain Divide

Not every problem requires all four bands. This is important.

Language-native domains — legal, content creation, customer support, sales communication — can be well-served by Band 2-4 solutions. The core complexity is textual. LLMs are natively suited for it.

Harvey AI is a good example. They’ve built a highly successful legal AI platform using generative AI and agentic workflows. They don’t need Band 1 ML models because legal work is fundamentally about interpreting and producing language. The architecture fits the domain.

Analytics-dependent domains — pricing, demand planning, site selection, financial forecasting, supply chain optimization — have different requirements. The core complexity is quantitative. You need actual calculations, not generated text that sounds like calculations.

This is where the band question becomes critical.

If you’re buying AI for pricing optimization and the vendor can’t explain what statistical models power their recommendations, you’re buying a chat-wrapper. (Gen-AI on top of a Dashboard) If you’re buying AI for demand forecasting and there’s no ML infrastructure behind the natural language interface, you’re buying a chat wrapper.

Chat wrappers can describe analytics. They cannot perform them.

What We See in the Field

After 40+ years advising restaurant operators, we’ve watched technology cycles come and go. The pattern is consistent: operators get sold on capability that doesn’t match reality, then spend years recovering from the wrong choice.

AI is following the same pattern, but faster and with higher stakes.

We’re seeing restaurant groups sign six-figure contracts with “AI-powered analytics” vendors that, in practice, add a chat interface on top of their existing BI dashboards. The operator asks: “Which stores should get a price increase?” The system generates a confident answer. That answer is based on pattern matching, not on deep machine learning analysis of transaction data.

The consequences show up in the P&L six months later.

Contrast this with operators who’ve implemented full-stack solutions. Starbucks built Deep Brew on a foundation of demand forecasting and site selection models they’d been developing for a decade. The chat interface came last, not first. Their results: 30% ROI improvement, 15% increase in customer engagement, and site selection accuracy materially better than competitors still using traditional methods.

Domino’s followed the same playbook. Band 1 demand forecasting and delivery optimization. Domino’s stock has outperformed Google, Amazon, and Apple since its 2004 IPO.

The lesson: in analytics-dependent operations, the sequence matters. Foundation first, interface second. As the pace of AI innovation accelerates, the companies that set the data and analysis foundation will find rapid proliferation of innovation in Band 3 & Band 4 - with the right stack.

At SignalFlare.ai, we’re finding that entire processes that once required teams of people to complete are starting to take teams of agents. But there’s one caveat: the up-front data work and attention to infrastructure is non-negotiable. While I respect any organization’s decision to ‘build it yourself’, in the new world you need a realistic understanding of what that means. “Good data” isn’t just about data quality — it’s also about dynamic ontologies, data unification and the right agent framework to be ready for the agentic revolution.

The Future-Proofing Question

Technology stacks are expensive to change. The vendor you choose today will shape your capabilities for years. So the question isn’t just “what can this vendor do now?” It’s “what will they be able to do as the technology evolves?”

Here’s the problem: architecture determines destiny.

A vendor that built a chat interface on top of an LLM is a Band 2 company. They can add agent capabilities (Band 3) and workflow orchestration (Band 4) over time. What they cannot easily add is Band 1 — the ML models that perform actual quantitative analysis.

Think of it like baking. You can frost a cake after it comes out of the oven. You can add layers. What you can’t do is change the fundamental recipe. A strawberry shortcake cannot become a black forest cake. The ingredients were set when the batter was mixed.

Vendors that started with ML foundations and added generative interfaces can operate across all four bands. Vendors that started with generative interfaces face structural limitations in analytics-dependent use cases.

The future-proofing question: where did this vendor start, and what does their architecture actually support?

Questions To Ask Vendors

For any AI vendor:

• Which of the four bands do you operate in today?

• What’s your roadmap across the bands?

• Where did your technical architecture originate — ML or LLM?

For analytics and decision support:

• What quantitative models power your recommendations?

• When I ask for a forecast or optimization, what’s actually calculating the answer?

• Can you show me the methodology, not just the output?

• How do you validate accuracy?

For agentic capabilities:

• What processes can you fully automate end-to-end?

• How do you handle exceptions and edge cases?

• What human oversight is built into the workflow?

• Can you maintain context across sessions and decisions?

For data integration:

• How do you connect to our existing systems?

• What happens to our data — where is it processed, stored, used for training?

• How do you handle data quality issues?

• What’s your semantic layer — how do you understand what our data means?

Restaurant-Specific Considerations

Restaurant operations present unique challenges that expose the band limitations quickly.

Menu pricing requires price elasticity modeling. An LLM can tell you what price elasticity means. It cannot calculate yours from your transaction data. That requires Band 1 machine learning models on your actual sales by item, daypart, location, competitive context & guest sentiment analysis. Band 2, 3 and 4 now operationalizes the science to simulate, distribute, contextualize, implement and educate to make implementations seamless and successful.

Labor scheduling requires demand forecasting at 15-minute intervals. Weather, local events, seasonality, and promotional calendars all factor in. Chipotle’s implementation of AI-driven scheduling cut their time-to-hire by 75% and improved application completion rates from 50% to 85%. That’s Band 1 prediction models powering Band 3 automation.

Store performance diagnosis requires isolating controllable factors from market conditions. When a store underperforms, you need to know whether it’s execution (fixable), trade area decline (strategic decision), or competitive intrusion (tactical response). That’s multi-factor analysis, not text generation.

For each of these, ask your vendor: what’s actually doing the math?

The Build vs. Buy Consideration

Some organizations are considering building their own AI capabilities. The band framework is useful here too.

Band 2 is increasingly commoditized. The APIs from OpenAI, Anthropic, and Google make it relatively straightforward to build chat interfaces and content generation features. If that’s all you need, building may be viable. However, even if your only goal is to leverage LLM’s for teams, pay careful attention to enterprise security, privacy and persistent memory considerations. “Free” versions don’t guard your private data, most LLM’s do not come with persistent team memory - and the whole point of that the algorithms should get smarter over time. Without memory you’ve missed the point.

Band 3 requires integration expertise. Connecting LLMs to your systems, managing authentication, handling errors, building reliable agents — this is harder than it looks. Most organizations underestimate the engineering required.

Band 4 requires orchestration infrastructure. Multi-agent coordination, workflow state management, human-in-the-loop processes, observability and debugging — this is genuinely complex. Few organizations should build this from scratch.

Band 1 requires domain expertise. Building ML models that actually work for your domain requires data science capability, training data, validation methodology, and ongoing model management. This is years of work, not a project.

The organizations best positioned to build are those with existing Band 1 capabilities (data science teams, ML infrastructure) who want to add generative interfaces. The organizations that should buy are those seeking Band 1 capabilities they don’t have internally.

Red Flags

Watch for these warning signs when evaluating vendors:

“Our AI handles everything.” No vendor operates expertly across all bands as a ‘walled garden’. If they claim to, ask specifically how. The best solutions will are building ecosystems for multi-partner data sourcing and agent proliferation.

Inability to explain the quantitative methodology. If you’re buying analytical capabilities and the vendor can’t explain what models produce the numbers, the LLM might select on-the-fly. That’s not a stable solution.

Band confusion. Vendors describing Band 2 capabilities (chat, summarization) as Band 4 (process automation). The language is often intentionally blurred.

No integration story. Agentic AI requires connecting to your systems. If the vendor can’t explain specifically how they’ll integrate, it might be slideware. Or if the solution ‘can learn any new data integration’ - remember, you will pay consumption credits to teach the model how to become a data onboarding analyst. It might be the most expensive ‘cheap’ data onboard you’ve ever attempted.

No restaurant references. If a vendor is selling analytics to restaurants but can’t provide references from operators who’ve validated accuracy against actual results, proceed with caution. This industry has specific data patterns, seasonality, and margin structures that generic solutions miss.

A Practical Framework

When planning your AI stack, consider this framework:

1. Map your use cases to bands. What do you actually need? Content generation is Band 2. Task automation is Band 3. Process automation is Band 4. Quantitative analysis requires Band 1.

2. Assess your internal capabilities. Do you have ML expertise? Data engineering? If yes, you can build on that foundation. If no, you need vendors who bring Band 1 capabilities.

3. Evaluate vendors against the full stack. Don’t just ask what they do — ask how they do it and where their architecture limits them.

4. Plan for evolution. The AI landscape is moving fast. Choose vendors whose architecture can grow with the technology, not vendors locked into a single band.

5. Validate before committing. Run pilots. Test on your data. Verify that the impressive demo translates to your reality.

At SignalFlare, we’ve spent many years building ML models for restaurant and retail analytics — demand forecasting, price elasticity, media attribution, menu optimization. When we added generative AI and agentic capabilities through our Navigator platform, we weren’t starting from scratch. We were adding an interface layer to an existing analytical foundation. That architecture allows us to operate across all four bands.

Not every vendor needs this. But for analytics-dependent use cases, understanding where a vendor sits in the stack is the difference between buying capability and buying marketing.

The Bottom Line

The AI landscape is confusing. Not by design, but some vendors benefit from ambiguity. Buzzwords blur distinctions.

The four-band framework cuts through the noise. It gives you a vocabulary for evaluation. It surfaces the architectural questions that matter. And it helps you distinguish between vendors who can solve your actual problems and vendors who can only talk about solving them.

The technology is evolving fast. The vendors who will matter in three years are those building across the full stack today. Choose accordingly.

---

Mike Lukianoff is co-founder of SignalFlare.ai, a decision intelligence company serving the restaurant and retail industries. www.signalflare.ai

Fred LeFranc is co-founder of Results Thru Strategy, a consultancy serving the restaurant, retail industries and technology companies. www.resultsthrustrategy.com