Match the tool to the problem! Stop defaulting to “GenAI”…

Three enterprise use cases. Five delivery paths. One decision grid.

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027. [1] Of the thousands of vendors now marketing “agentic” solutions, Gartner estimates only around 130 are the real thing. [2] Neither number is the actual problem. The actual problem is more basic: enterprise teams keep applying the wrong class of AI to the problem in front of them.

Don’t confuse the three kinds of AI

This post is about agentic AI: systems that plan, act across tools, and hand off to humans at gates. It is not about GenAI, which is the linguistic engine inside an agent but not the agent itself.

It is also about classical AI (Machine Learning and Optimization): the mathematical methods that still solve most enterprise optimization and forecasting problems better than any language model ever will. If your problem is linear programming or time-series regression, a transformer is not your friend. Math is.

Here are three real use cases from manufacturing and CPG with five delivery paths each:

The five candidate paths

For every use case below, an enterprise buyer in 2026 realistically chooses between:

1. Classical math on-prem or on a hyperscaler: Gurobi, Pyomo, OR-Tools, scikit-learn, LightGBM on managed compute
2. Cloud-native AI/ML platform: Amazon SageMaker, Google Vertex AI, Azure AI Foundry
3. Claude Code as the orchestration engine: agentic coding and tool-use at build and runtime
4. Enterprise SaaS suite: a CRM, OT, or logistics platform with embedded AI workflows
5. Microsoft Copilot: the M365 productivity assistant, with Copilot Studio for light agent workflows

Case 1 — Parts allocation across factories (Procurement)

The problem: allocate constrained parts to competing factories to maximize throughput and minimize expedite cost. Constraints: supplier lead times, contractual minimums, plant-level capacity, dual-sourcing rules. This is mixed-integer programming (MILP), a deterministic, audit-ready solver. It runs nightly. It does not hallucinate and it is highly explainable on its solution path.

Case 2 — Demand forecasting

The problem: hierarchical demand forecasting across SKUs, channels, and geographies. Inputs: POS data, promotions, weather, macro signals. Outputs: weekly forecasts that feed S&OP. This is time-series forecastig, i.e. classical machine learning. This is NOT a GenAI problem. Never will be.

Case 3 — CPG quality complaint with legal exposure

Now the archetype changes.

A consumer contacts a CPG brand claiming a product quality issue. They mention injury. They threaten escalation. Under FSMA in the US and Regulation 178/2002 in the EU, certain complaint categories trigger mandatory notification clocks. The average direct cost of a food recall is approximately $10 million. Five percent of CPG companies incur over $100 million in total impact. Roughly 80% of recall costs arrive long after the recall itself is closed. [3][4] This is not a math problem. It is a multi-system orchestration problem with legal gates, cross-functional handoffs, and hard compliance deadlines.

This is what agentic AI is actually for!!!

The full cycle

The agent handles intake, triage, and data gathering autonomously. Every legal, regulatory, and external-communication step is human-in-the-loop. By design, not by exception. See the workflow diagram below.

Stage 1 parses the complaint and extracts the product batch, injury signal, and severity markers. Stage 2 enriches by querying MES, QMS, ERP, and supplier cold-chain data. Stage 3 is the first human gate: if the enriched picture crosses a regulatory threshold, legal and QA review is mandatory before the case advances. Stage 4 generates the customer response and opens a CAPA if the issue is systemic: the agent drafts, the human approves. Stage 5 is the hardest gate: regulatory notification to FDA, EFSA, or the relevant local authority is signed by a named human, with full audit trail attached.

The end state is not an agent that replaces the quality and legal teams. It is an agent that compresses hours into minutes, so humans focus only on decisions that carry regulatory or reputational weight.

How to know what to pick and when?

Classical AI is a fit for mathematical problems — GenAI is a chatbot — Agentic AI is an orchestration that can handle complex workflows

Reality check

Three things worth saying out loud:

1. For classical cases like allocation and forecasting, stay classical at runtime. Pyomo, LightGBM, and OR-Tools run for free on your own compute, don’t pay SaaS license fees or agent-platform overhead for what is fundamentally linear algebra. Where Claude Code still earns its place is as a build accelerator: 10 to 14 days from problem statement to working prototype, with the math running classical underneath.
2. Claude Code appears unusual on this grid. It is not a product category in the same sense as the others: it is a delivery and orchestration pattern. What it measurably changes is time-to-value. Ten to fourteen days from problem statement to working prototype is no longer an outlier; for any AI investment with real P&L exposure, that compression is the single biggest lever available in 2026. To me there’s is no tool on the market that provides the same performance and accuracy and speed.
3. Agent-washing is real, and it is getting worse. Most enterprise AI roadmaps still default to “add GenAI” as if it were a universal solvent. It is not. The discipline required is the same discipline we have always applied in enterprise technology: match the architecture to the problem, budget for integration, make governance non-optional, and refuse to confuse a productivity assistant with a production system.

If you are sponsoring an AI investment this year, run every use case through one question: what class of AI does this actually need? Most of the time, the answer will not be what the vendor is selling.

Sources

[1] Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025. (https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)

[2] Gartner, ibid. Estimate of approximately 130 genuinely agentic vendors among the thousands marketing the label. (https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)

[3] Food Marketing Institute and Grocery Manufacturers Association, widely cited in CPG industry reporting. Average direct cost of a food recall: approximately $10 million. (https://globalfoodsafetyresource.com/wp-content/uploads/2014/08/www.gmaonline.org_file-manager_images_gmapublications_Capturing_Recall_Costs_GMA_Whitepaper_FINAL.pdf)

[4] GMA survey: 5% of surveyed companies report over $100 million in total direct and indirect recall costs. Lockton estimate, widely cited across industry reporting, attributes roughly 80% of total recall cost to post-recall long-tail impact (lost sales, reputation, contract cancellations). (https://www.rentokil.com/blog/industry-insights/cost-of-product-recalls)