Ideas lab
When machines join the team
Article
26/09/2025
Key takeaways
AI proofs-of-concept don’t fail because of the tech. They fail because organizations design them for show, not for scale.
Small, reliable wins create more momentum than big, flashy demos. Starting narrow builds confidence and paves the way for sustainable growth.
Ownership and economics matter as much as algorithms. Without clear accountability and cost discipline, even strong proofs-of-concept collapse in limbo.
By Ismail Amla, Senior Vice President, Kyndryl Consult
In August 2025, a relatively small MIT Media Lab Study claimed that 95% of all AI proofs-of-concept fail. That finding went viral, in part because the researchers boiled it down to AI hype.
I have my doubts about the core finding. The sample size was small, and also, most organizations don’t systematically track such metrics. Even still, the study highlights something important, which is that most executives still struggle to turn AI demos into working business tools.
Many organizations approach AI proofs-of-concept as polished showcases for leadership rather than practical systems. These efforts consume significant resources, depend on idealized data, and often fail when exposed to real-world use.
My team has worked on dozens of AI proofs-of-concept across all major verticals. We have learned the hard way how to convert them into business-ready systems. Reflecting on that experience, I’ve learned five rules that businesses should follow to radically improve their chances of success.
01
Embrace deliberate constraints
With AI hype, the instinct is to deploy the biggest model, process massive datasets, or build sprawling feature suites — but that path often leads to failure. The most successful proofs-of-concept start small and stay small: they run on manageable infrastructure, use datasets simple enough for human inspection, and tackle a scope narrow enough to describe in one sentence.
This deliberate, iterative approach often clashes with corporate culture, which often tends to reward ambition over restraint. Yet in AI, constrained wins compound into larger successes. A modest capability that works reliably inspires more confidence than a flashy demo that frequently breaks. The urge to showcase AI’s full potential can blind teams to the real value of incremental progress.
02
Design for production
The biggest reason proofs-of-concept fail is that teams often kick production too far down the road. But things like logging, monitoring, version control, and safety need to be built in from the start if you want a demo to grow into a real business system.
Production readiness is not a destination but a design philosophy. With that in mind, build systems that can handle real data volumes, unexpected inputs, and operational stress from day one. The additional overhead is minimal compared to the cost of rebuilding.
03
Prioritize reliability over innovation
AI proofs-of-concept succeed when they’re built on proven foundations. Use established infrastructure templates, integrate testing into existing workflows, and compare models with objective benchmarks — not gut feel. Assume today’s AI model families will keep improving, and design accordingly.
The real value in AI isn’t chasing novel algorithms; it’s framing business problems so AI can solve them reliably at scale. Solid, repeatable engineering practices beat flashy experiments. Standardized approaches cut complexity, lower maintenance costs, and raise the odds of deployment success. Keep innovation focused on problem-solving, not on reinventing infrastructure.
04
Solve genuine problems
Anchor every decision to measurable business pain points. Identify real users facing concrete problems they are willing to pay to resolve. If your proof-of-concept's primary value proposition centers on technological impressiveness rather than operational utility, you are building the wrong solution.
Skip the executive showpieces. Build simple applications that save people time and money in their daily work. The best proofs-of-concept tackle workflow inefficiencies or decision bottlenecks and fit naturally into existing processes. Start small, solve one real problem, and deliver immediate value — even a partial fix builds momentum.
05
Establish ownership and economics
Decide who owns the system before it goes live. Who will respond if it fails on a weekend? What service levels will you commit to? Who handles retraining models or approving compute costs?
From day one, track the unit economics — cost per request, per user, per workflow. Compare small and large models, or local versus cloud execution, using real data rather than projections. Don’t let proofs-of-concept drift between research and operations. Assign clear owners, responsibilities, and lifecycle plans upfront. Financial sustainability must be proven, not assumed.
As I look back on the successes and failures of many AI proofs-of-concept, too many were doomed from conception. Some were too ambitious. Others were simply too expensive, too disconnected from genuine business needs, or too optimized for demonstration rather than deployment. By inverting this approach, leaders can dramatically improve their odds of building applications that both ship and scale.
The difference between successful and failed AI proofs-of-concept rarely lies in algorithmic sophistication. It lies in mundane business decisions made during project planning — and these are things we can solve for.
