Skip to content
EN DE

When AI

Your stakeholder wants an “AI-powered” workflow. Your engineering team can build it. The question nobody asks: Does this problem even need AI?

As a PM, your most important skill isn’t deploying AI — it’s recognizing when you shouldn’t. Because the most expensive AI projects aren’t the ones that fail. They’re the ones that solve a problem that didn’t exist.

Not every problem is an AI problem. The decisive question: Is the problem deterministic or probabilistic?

DeterministicProbabilistic
LogicClear, stable, documentableAmbiguous, context-dependent
RulesFinite if-then chainsToo many exceptions for rules
OutputOne “correct” answerNo single “correct” answer
ExampleTax calculation, validationSentiment analysis, recommendations

When rule-based is enough: Logic is clear and stable, decisions must be auditable, processes are repeatable.

When AI/ML makes sense: No single “correct” answer, system improves with more data, rules have too many exceptions.

Reality check: Most enterprise problems are hybrids. Your login flow is deterministic. Your fraud detection is probabilistic. Both in the same product.

The numbers are sobering:

  • 64% of users would disable features or leave products when AI features are forced on them (ZDNET/Aberdeen)
  • Only 5-30% of AI implementations deliver measurable P&L impact — depending on study and definition (MIT Sloan Management Review 2025, McKinsey Global AI Survey 2024)
  • Only 1 in 4 AI projects delivers promised ROI, only 16% scale enterprise-wide (IBM, 2025)

The problem is rarely the technology. The problem is that no clear user problem existed before someone put “AI” on the roadmap.

  • Humane AI Pin & Rabbit R1: “A solution looking for a problem” (Logitech CEO). Hardware products without a clear use case.
  • McDonald’s AI Drive-Thru (IBM): Shut down after 3 years. Viral TikTok video: 260 McNuggets ordered because the system interpreted background noise as orders.
  • United Healthcare: 90% error rate in a specific algorithm (nH Predict) for post-acute care denials (STAT News Investigation).
  • Google AI Overviews: “Put glue on pizza” recommendations. Deployment without adequate error tolerance definition.

5 check questions before putting AI on your roadmap:

#QuestionNo = Warning Sign
1Is the problem probabilistic?A rule-based engine probably suffices
2Does the system improve with more data?A static solution is more efficient
3Is there enough quality training data?Cold-start problem, high upfront investment
4Is the error tolerance acceptable?High-stakes domain without human review = risk
5Does expected value exceed AI-specific costs?Variable LLM costs in fixed subscriptions = unit economics problem

All 5 yes? AI is a valid approach. One or more no? Evaluate alternatives before committing to AI.

You’re a PM at a B2B SaaS for invoice processing. The VP Product wants “AI-powered Invoice Processing” on the roadmap. Currently you use template-based OCR with fixed rules.

The situation:

  • 120,000 invoices/month, 85% follow 4 standard formats
  • Current rule-based accuracy: 94% for standard formats, 45% for non-standard
  • AI model in eval: 96% for standard, 82% for non-standard
  • AI cost: $0.03 per invoice (LLM-based) vs. $0.001 per invoice (rule-based)
  • 15% of invoices (non-standard) cause 60% of manual effort

Options:

  1. Switch everything to AI: 96%/82% accuracy, $3,600/month
  2. Keep everything rule-based: 94%/45% accuracy, $120/month
  3. Hybrid: Rule-based for standard formats, AI only for non-standard
How would you decide?

The best decision: Option 3 — Hybrid.

Why:

  • Standard formats: Rule-based delivers 94% at 30x lower cost. The 2% improvement from AI doesn’t justify a 30x price increase
  • Non-standard formats: AI improves from 45% to 82% — that’s the real lever. These 15% of invoices cause 60% of manual work
  • Hybrid cost: ~$660/month (rule-based for 102k standard + AI for 18k non-standard) instead of $3,600

The PM mistakes in this scenario:

  • “Bolt-on AI”: Slapping AI on top without rethinking the workflow. Rule-based works for 85% of cases — why replace it?
  • Models as strategy: The model is a commodity. Your competitive advantage lies in the data pipeline and validation rules.
  • Ignoring unit economics: $0.03 per invoice sounds cheap — until you multiply by 120k invoices per month.
  • AI is a solution, not a feature. Start with the problem, not the technology. If you can’t articulate which user problem AI solves, you don’t have one.
  • Most problems are hybrids. Deterministic where possible, probabilistic where necessary. This saves cost and reduces errors.
  • Define error tolerance before deployment. McDonald’s, United Healthcare, and Google showed what happens when you don’t.
  • Do the unit economics math. Variable LLM costs inside fixed-price subscriptions are a business model risk, not a technical detail.

Sources: ZDNET/Aberdeen Consumer AI Survey (2024), MIT Sloan “AI Implementation Strategies” (2025), IBM Global AI Adoption Index (2025), Logitech CEO Bracken Darrell on Humane/Rabbit (2024), McDonald’s/IBM Drive-Thru Pilot Post-Mortem (2024)

Part of AI Learning — free courses from prompt to production. Jan on LinkedIn