Intent Engineering: Why 95% of Enterprise AI Fails and What the Top 5% Do Instead
Apr 05, 2026
|
Key Takeaways ▶ 95% of generative AI pilots fail to deliver measurable business impact (MIT) - not because the tech is bad, but because the intent is missing ▶ Prompt engineering tells AI what to do. Context engineering tells AI what to know. Intent engineering tells AI what to want. ▶ Klarna saved $60m then started rehiring - their AI optimised for speed, not the outcomes that actually mattered ▶ The fix is composable infrastructure deployed against specific business targets, not more pilots |
|
95% of generative AI pilots fail to deliver measurable business impact |
40% of agentic AI projects will be cancelled by 2027 |
3.3% of Microsoft Copilot users actually pay for it despite 90% Fortune 500 adoption |
Klarna replaced 853 employees with AI. The system handled 2.3 million conversations in its first month. Resolution times dropped from eleven minutes to two. The company saved $60 million.
Then the CEO admitted it had broken something more important than it fixed, and started rehiring humans.
This is the most instructive AI failure of the past year. Not because the technology didn’t work. It worked brilliantly. The failure was that nobody defined what “working” actually meant.
The three stages of AI deployment
The industry has moved through two distinct phases of AI maturity, and most organisations are stuck between them.
|
Stage 1 Prompt Engineering Tells AI what to do. Crafting instructions, tweaking settings, hoping for the right output. Useful for individuals. Useless at enterprise scale. |
Stage 2 Context Engineering Tells AI what to know. Retrieval systems, knowledge bases, memory architectures. Where most serious organisations are today. The infrastructure is real. The investment is significant. |
Stage 3 Intent Engineering Tells AI what to want. Encodes organisational goals, trade-offs, and decision boundaries so agents optimise for actual business outcomes. Almost nobody is building this yet. |
Without intent engineering, you get Klarna. A system that optimised for speed and volume because those were the metrics it could see. Customer loyalty, relationship quality, lifetime value: invisible to the agent, so ignored by the agent.
The Klarna warning
Klarna’s story isn’t unusual. It’s just the most visible example of a pattern playing out across every industry.
|
95% of generative AI pilots fail to deliver measurable business impact. Not because the models are bad. Because the implementations optimise for the wrong things. Source: MIT NANDA Initiative, “The GenAI Divide: State of AI in Business 2025” |
Gartner predicts over 40% of agentic AI projects will be cancelled by 2027. The reasons are consistent: escalating costs, unclear business value, and what Gartner calls “agent washing” - vendors rebranding existing products with agentic buzzwords while delivering nothing new. Of thousands of vendors claiming agentic AI capabilities, Gartner estimates only about 130 are genuine.
Microsoft Copilot tells the same story from a different angle. 90% of Fortune 500 companies have adopted it. Just 3.3% of users actually pay for it. Broad adoption, minimal real usage. The tool is available. The intent - the reason to use it for meaningful work - is missing.
The common thread: organisations have solved “can AI do this task?” They haven’t solved “can AI do this task in ways that serve our actual goals?”
What intent-driven deployment looks like
There’s a counter-example worth examining.
A European industrial distributor had a specific problem: customer churn was invisible. The sales team couldn’t see which accounts were drifting until they’d already left. Revenue was leaking, not because the product was wrong or the service was bad, but because the signals were buried in order book data that nobody had time to analyse.
|
Intent-driven deployment in action The AI deployment was built around one intent: surface at-risk revenue before it’s lost. Not “deploy a chatbot.” Not “explore use cases.” One operational target. One measurable outcome. EUR 45 million in at-risk revenue identified in six weeks. The system knew what “good” looked like because someone specified it in business terms - revenue at risk - not technical metrics. |
This is what PE operating partners understand instinctively. They call it the value creation plan. Every acquisition has one. It specifies exactly where value will be created, by what mechanism, on what timeline. When AI is deployed against that plan - against a specific operational target - the intent gap closes.
When AI is deployed without it - when the instruction is “explore AI opportunities” or “build an AI strategy” - you get scattershot pilots. You get Klarna.
The three missing layers
Nate B Jones, whose analysis prompted this piece, identifies three architectural gaps that explain why most enterprise AI fails:
|
Gap 1 Unified context infrastructure Most organisations have data everywhere: ERPs, CRMs, email, spreadsheets, the heads of experienced employees. But it’s not organised coherently. AI systems can’t reason across domains because the knowledge isn’t connected. Industrial businesses are particularly exposed. Decades of data trapped in SAP, in tribal knowledge, in manual processes. |
|
Gap 2 Coherent workflow architecture Agents get deployed in isolation: one for customer service, another for reporting, a third for document processing. They don’t share decision frameworks. They don’t coordinate. Each optimises locally, and the global outcome suffers. This is Klarna’s exact failure pattern: a customer service agent optimising for resolution speed while the business needed it to optimise for retention. |
|
Gap 3 Organisational alignment frameworks No mechanism exists to translate what the company values - customer relationships, revenue protection, operational safety - into constraints that autonomous systems can follow. The values stay in the boardroom. The agents operate without them. |
The organisations getting AI right are building these layers, whether they call it “intent engineering” or not. They start with the business outcome. They connect the data. They deploy against a specific target. And they measure success in business terms, not technical ones.
The infrastructure problem nobody is solving
There’s a deeper issue underneath the three missing layers. Most AI deployments are point solutions. A chatbot here. An analytics dashboard there. A document processing tool somewhere else. They don’t learn from each other. They don’t compound. When a better model arrives, you rebuild.
What organisations actually need is composable, vendor-agnostic architecture. Infrastructure that treats AI as a strategic investment, not an IT project.
|
What composable AI infrastructure means ▶ Model-agnostic. The best models swap in and out as AI advances. No vendor lock-in. No rebuilding when something better arrives. ▶ Start narrow, expand fast. One high-impact area first - revenue protection, operational efficiency, knowledge capture - then expand. Each new deployment builds on the last. ▶ Compounding intelligence. The platform embeds the company’s key processes, its unique ways of working, and all intelligence gathered since inception. The knowledge compounds. ▶ Progressive autonomy. From surfacing insights to predicting the next best action, to delivering the right data to the right people at the right time. More tasks automated over time. |
This is the infrastructure layer that enables intent engineering at scale. Without it, every AI deployment is isolated. With it, each one makes the next more valuable.
The firms still buying point solutions are building a collection of tools. The firms building composable infrastructure are building a competitive advantage that compounds through the hold.
The firms that get this right
85% of PE firms are pushing portfolio companies toward AI adoption. 70% plan to increase AI investment by 25% or more in the next 18 months.
Most are doing it wrong. Running scattershot pilots. Buying point solutions. Exploring use cases without the infrastructure to connect them.
The firms getting it right start with intent - a specific operational outcome - and build on composable infrastructure that learns the business over time. Deploy against one target. Prove value. Expand to the next. The intelligence compounds. The automation deepens. What started as a single revenue protection tool becomes an operational nervous system.
The gap between “spending on AI” and “generating returns from AI” is where the next wave of competitive advantage lives. The firms that close it will be the ones that invested in infrastructure, not just tools.
What’s the specific operational outcome your AI deployment is measured against? If the answer is vague - “efficiency” or “innovation” or “exploring use cases” - the intent layer is missing. And without it, you’re building the next Klarna.
|
If this resonates - let’s have a conversation We deploy OEX - our Operational Excellence platform - as composable, vendor-agnostic infrastructure for industrial businesses. Start with one operational target. Expand across the business. The platform learns your processes, embeds your intelligence, and compounds in value over time. Working solutions in weeks. If you want to see what that looks like for a specific business, let’s talk. |
Sources & References
| Klarna: Q3 2025 Earnings, Fortune, Bloomberg - 853 FTEs replaced, 2.3m conversations, $60m savings, CEO reversed course and started rehiring |
| MIT NANDA Initiative: The GenAI Divide: State of AI in Business 2025 - 95% of generative AI pilots fail to deliver measurable business impact |
| Gartner: Agentic AI Forecast, June 2025 - 40%+ of agentic AI projects cancelled by 2027; ~130 genuine agentic vendors out of thousands |
| The Register: Microsoft Copilot Research, February 2026 - 90% Fortune 500 adoption, 3.3% paid uptake |
| CLA Connect: 85% of PE firms pushing portfolio companies toward AI adoption |
| Bain & Company: 70% of PE leaders increasing AI investment by 25%+ in the next 18 months |
| G3NR8: ERIKS deployment - EUR 45m at-risk revenue identified in six weeks |
Frequently Asked Questions
|
Why do 95% of AI pilots fail? MIT research found that 95% of generative AI pilots fail to deliver measurable business impact. The failure is not in the technology but in the implementation: generic tools deployed against generic objectives, pilots proliferating without clear connections to the outcomes that matter, and no mechanism to translate organisational goals into constraints that AI systems can follow. |
|
What is intent engineering? Intent engineering is the discipline of encoding organisational goals, trade-offs, and decision boundaries into AI systems so autonomous agents optimise for actual business outcomes. It sits above prompt engineering (telling AI what to do) and context engineering (telling AI what to know). Intent engineering tells AI what to want. Without it, systems optimise for measurable proxies rather than the outcomes that actually matter. |
|
What happened with Klarna’s AI deployment? Klarna replaced 853 employees with AI, handling 2.3 million conversations in the first month with resolution times dropping from eleven minutes to two and saving $60 million. The CEO then admitted it had broken something more important than it fixed and started rehiring humans. The system optimised for speed and volume because those were the only metrics it could see. Customer loyalty, relationship quality, and lifetime value were invisible to the agent. |
|
What is the difference between prompt, context, and intent engineering? Prompt engineering tells AI what to do - crafting instructions and tweaking settings. Context engineering tells AI what to know - building retrieval systems, knowledge bases, and memory architectures. Intent engineering tells AI what to want - encoding organisational goals so autonomous agents optimise for actual business outcomes. Most organisations are stuck between prompt and context engineering, with almost nobody building the intent layer. |
|
Why are 40% of agentic AI projects expected to be cancelled? Gartner predicts over 40% of agentic AI projects will be cancelled by 2027 due to escalating costs, unclear business value, and “agent washing” - vendors rebranding existing products with agentic buzzwords. Of thousands of vendors claiming agentic AI capabilities, Gartner estimates only about 130 are genuine. The root cause is deploying agents without defining the business intent they should serve. |
|
What does intent-driven AI deployment look like in practice? At a European industrial distributor, the deployment was built around one intent: surface at-risk revenue before it is lost. One operational target. One measurable outcome. EUR 45 million in at-risk revenue identified in six weeks. The intent was defined before a single line of code was written. The system knew what “good” looked like in business terms, not technical metrics. That is the difference between a working deployment and another failed pilot. |
Inspired by Nate B Jones, “Prompt Engineering Is Dead. Context Engineering Is Dying. What Comes Next Changes Everything.”
G3NR8 builds operational AI capability for PE firms and their portfolio companies. OEX - our Operational Excellence platform - deploys in weeks and compounds through the hold.
Keep informed with the newsletter for PE operating partners and the portfolio companies they back.
Get operational insights and trends, AI frameworks, resources and real deployment stories.
