How to avoid burning through AI investment before it pays off

Executive overview

Most companies are automating existing workflows with AI rather than reimagining them — this produces incremental gains at high cost, not the 10x returns executives expect. The cost of building and running AI products is orders of magnitude higher than web development, so the ROI window is long and the burn risk is real.

The winning approach combines a bold product vision with disciplined weekly execution, metered capital deployment, and deep technical talent — not prompt engineers relabelled as AI engineers.

Obsess over the customer's problem, not the AI itself; the technology is irrelevant if it doesn't materially change someone's life.

Dream big, but ship in weeks

• AI investment only pays back if the vision is transformational, not incremental — automating existing workflows rarely justifies the cost.
• The iPhone's early apps replicated old workflows (e.g., a taxi-call app ignoring the onboard GPS); Uber came from connecting two existing tools in a new way.
• Being AI-first means tearing down existing structures, not bolting AI onto them.
• The right question: can AI eliminate a process entirely, not just speed it up?
• A long-term vision is necessary, but must be validated by continuous weekly improvement signals.
• If week-on-week metrics aren't improving, the one-year outcome won't either.

Focus on the product, not the AI label

• Customers don't care whether a product uses AI — they care whether it solves their problem.
• Companies that make the greatest use of AI rarely advertise it (Amazon, Apple).
• The correct KPIs: 50% workflow reduction, 20% cost decrease, higher customer satisfaction — not "AI adoption."
• Internal users are the best first test market; their frustrations surface the best use cases.
• Asking "what frustrated you this week?" is more productive than asking "what should we build with AI?"
• Don't ask customers what they want — identify the problems that will still exist years from now.

Burning star syndrome: managing AI CAPEX

• Burning star syndrome is the pattern of burning capital so fast on AI infrastructure that the company collapses before ROI arrives.
• Big tech is committing ~$600B in CAPEX in 2025; startups and enterprises building on top of that must absorb those costs without a guaranteed return.
• AI revenue cycles are long — Salesforce confirmed Agentforce won't add revenue in 2025, even after heavy investment.
• Analogy: the brightest stars burn out soonest, becoming black holes or supernovae. A brown dwarf generates nothing. Aim for a sun — steady, long-burning, capable of sustaining life.
• FOMO and FUD (fear, uncertainty, doubt) are driving executives into use cases whose value conversion doesn't justify the spend.
• The antidote is metered funding: deploy capital in stages tied to demonstrated proof points, the same way startups move from pre-seed to Series A.
• Don't bet everything on a few large initiatives — take multiple smaller bets and let evidence, not conviction, determine what scales.

Validating use cases: desirable, viable, doable

• A grand vision can fail simply because current AI technology can't deliver it — feasibility must be tested early, not assumed.
• Three filters for every use case: desirable (does it solve a real problem?), viable (is the economics sound?), doable (can it be built with available models without R&D)?
• Unlike other products, AI ideas may be technically impossible off the shelf today, even if they're theoretically sound.
• Ship to real users (internal first) early and regularly; gen AI products get used in ways no one anticipated because of the "ask me anything" surface.
• Use weekly shipping to generate the signal that the product is moving toward value — not just toward completion.
• Growth-linked use cases (new customers, higher engagement) are more defensible than pure cost-cutting plays.

Avoiding the ITification of AI

• Treating AI like an IT project — assembling a team, specifying requirements, waiting for delivery — systematically underestimates the skill gap.
• Prompt engineering fluency is not the same as data science expertise; enterprises that conflate them consistently fall short of production-grade accuracy.
• Accuracy is non-linear: going from 0% to 50% is easy; going from 50% to 80–90% (where customer value begins) requires deep expertise.
• Analogy: the Large Hadron Collider was needed to approach the speed of light; similarly, the final accuracy gains in complex AI require specialist capability, not generalist effort.
• A practical threshold: at least one-third of the AI team should be deep experts, not IT professionals recently relabelled.
• Build a culture of problem-solving and rapid technique adoption — this is the differentiator between AI leaders and AI laggards.

Three priorities for corporate AI leaders

1. Dream big, deliver in weeks. Simplify the end user's life radically — cut entropy, don't just add features.
2. Manage CAPEX carefully. Ensure the vision has a credible adoption path and that funding is sized to last until ROI arrives.
3. Build the right team and culture. Deep expertise, a problem-solving mindset, and the ability to adopt new techniques quickly without rerunning entire pipelines.