Before You Create Value With AI, Stop Destroying It
Created on 2026-02-24 14:39
Published on 2026-02-24 14:42
In June 2023, a New York attorney named Steven Schwartz filed a court brief citing six previous judicial decisions. The brief looked thorough. Professional. Exactly what competent legal work should look like.
None of the six cases existed.
Schwartz had used ChatGPT to research the brief. The AI generated six fabricated citations, complete with realistic case names, docket numbers, and judicial reasoning. He didn’t verify any of them. When opposing counsel couldn’t find the cases, the judge ordered Schwartz to produce copies. He went back to ChatGPT and asked if the cases were real.
ChatGPT confirmed they were.
They weren’t.
Schwartz and his colleague were sanctioned, fined $5,000, and publicly humiliated in a ruling that made international headlines. His firm’s reputation was destroyed. His client’s case was compromised.
The technology did exactly what it was asked to do. It generated plausible legal research at speed.
Nobody checked whether the output was real.
The Question Nobody Is Asking
Every AI conversation I’m hearing right now, across financial services, legal, consulting, government, every industry deploying agents, centres on value creation. How do we use agents to create value? Where’s the ROI? What’s the business case?
Those are the wrong questions to start with.
The right question: where are your agents already destroying value, and do you even know it?
MIT’s Project NANDA research, published in July 2025, tracked over 300 AI implementations and found that 95% of organisations are getting zero return from their AI investments. Not low returns. Zero. The finding that should reframe every AI conversation in your boardroom: “This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.”
The 95% aren’t failing because they chose the wrong model. They’re failing because they’re deploying agents against the wrong kind of problem. Or deploying agents against the right kind of problem without the human layer that makes the output trustworthy.
In industries where regulatory, reputational, and fiduciary consequences are severe, deploying without that layer doesn’t just waste money. It destroys value faster than it could ever create it.
“Hard” Is Not One Thing
Here’s the insight that reframed how I think about AI deployment.
We’ve been treating difficulty as a single dimension. A problem is hard. Let’s throw AI at it. But “hard” has at least seven distinct types. AI agents are genuinely good at two of them right now.
The other five? Deploying agents against those problems doesn’t just fail. It actively destroys value. And the destruction compounds faster than the creation because AI is an accelerator. An accelerator does not care which direction you’re travelling.
For each of these seven types, ask yourself: where is my organisation deploying agents right now, and which kind of problem are they actually solving?
1. Effort Problems
Not intellectually difficult. Just massive. Auditing 3,000 vendor contracts. Processing thousands of insurance claims. KYC document review across a large client portfolio. Any competent person could handle any individual piece. The challenge is sustained attention across a huge surface area without dropping detail.
This is exactly what agentic AI was built for. High volume, straightforward thinking at each step, immediately measurable ROI. This is your entry point.
But only with human auditors in the loop. Never agents alone.
I learned what skeleton-crew thinking costs early in my career at HSBC. The bank’s data processing operation was running on what management called “skeleton headcount.” The decision was purely cost-driven. Then a double payment was processed to a company called Impressed Metal. It became the largest unrecovered loss in the bank’s history.
Skeleton headcount because the spreadsheet said you could. That’s the same logic behind deploying effort-solving agents without verification layers. The spreadsheet says you can cut the humans. The spreadsheet doesn’t account for the catastrophe that arrives the one time the system gets it wrong.
The rule: Effort problems are the entry point. Agents plus auditors. Never agents alone.
2. Reasoning Problems
Holding multiple variables in your head, following a chain of logic through branches and dependencies, arriving at a non-obvious conclusion. Multi-jurisdiction tax optimisation where interaction effects across a dozen countries create genuinely hard combinatorial space. Determining whether a financial instrument triggers reporting obligations under multiple regulatory frameworks simultaneously. Tracing a chain of transactions across entities to reason about structural risk.
Frontier models have made dramatic leaps here. But the consequences of a wrong answer are severe. You need humans who can assess whether the reasoning is actually sound, not just plausible-sounding.
And reasoning problems are rarer than most leaders think. The tax attorney spends perhaps 10% of her week on the genuine multi-jurisdiction puzzle. The other 90% is client management, document gathering, coordination, and navigating ambiguity about what the client actually wants. The reasoning slice is high-value. But it’s narrow.
The rule: Reasoning problems are the high-value play. But they demand the strongest verification you have. And they’re a smaller slice of the work than you’d expect.
3. Coordination Problems
Getting six teams aligned on a shared decision when each has different priorities. Managing information flow during M&A due diligence across legal, financial, regulatory, and operational workstreams. Agent teams are beginning to handle information routing and dependency flagging. But this is early.
Coordination failures compound non-linearly. If an agent routes a compliance question to the wrong team or auto-escalates a client issue to the wrong authority, the downstream damage multiplies in ways that don’t surface on a dashboard until the consequences have already materialised.
The rule: Governance is not the Department of No. It is the Department of How. Deploy coordination agents without clear governance, and you’ve built an accelerator for chaos.
4. Emotional Intelligence Problems
Delivering bad news to a client whose portfolio just dropped 18% and who is already anxious about succession planning. Reading a room and knowing that someone’s silence means opposition, not agreement. Navigating a negotiation where the stated concern is pricing but the real concern is control.
AI does not solve this reliably. And this is where I see the most value being destroyed with the least awareness.
In 2023, the National Eating Disorders Association replaced their human helpline counsellors with an AI chatbot called Tessa. Within days, Tessa was recommending calorie restriction and weight monitoring to people suffering from eating disorders. People in clinical distress. The effort component worked perfectly: queries handled at scale, around the clock, at a fraction of the cost. But the problem was never effort. It was emotional intelligence. Every person calling that helpline needed someone who could read between the lines, calibrate tone, and recognise when a “routine” question was actually a crisis.
The chatbot caused active harm to the people the organisation existed to serve.
This is particularly consequential across APAC markets, where business is fundamentally relationship-first. Where trust precedes transactions. Where a client’s decision to stay or leave is personal. Deploying AI that can’t read emotional context in cultures where indirect communication carries more meaning than direct statements is a recipe for quiet, compounding damage.
The rule: Every dollar saved by automating effort is worthless if a tone-deaf AI interaction destroys a relationship worth multiples of that saving in recurring revenue.
5. Judgment and Courage Problems
Deciding to exit a profitable but ethically questionable product line. Saying no to a large client whose onboarding would stretch compliance past the breaking point. Flagging a risk that nobody wants to hear about because it would reduce short-term revenue.
AI can produce the analysis. Model the scenarios. Quantify the risks. Present the options.
But the decision itself, the willingness to accept career risk, reputational cost, or short-term pain for long-term institutional integrity, remains entirely human.
I think about 2008. The financial crisis was not a reasoning failure. Michael Burry at Scion Capital saw the subprime collapse coming years in advance. His analysis was correct. Any competent analyst could have reached the same conclusion.
The hard part was holding the position while investors threatened to pull their money, while his board pressured him to close, while every peer and every headline told him he was wrong. He endured years of career-threatening pressure to act on what the data clearly showed.
The bottleneck was courage. Not computation.
The rule: Any organisation that believes AI removes the need for human courage in decision-making is building a system that will fail at the moment it matters most.
6. Domain Expertise Problems
The veteran M&A attorney doesn’t evaluate deals better because she’s smarter. She evaluates them better because she’s closed 300 deals and has internalised which representations and warranties actually get litigated versus which ones are boilerplate nobody enforces. The senior trader recognises a market pattern from having lived through three cycles. The compliance officer knows that a specific regulator always asks about certain issues during audits, even though it’s not in any published guidance.
This is the most valuable and most fragile asset in any organisation. It walks out the door every time a senior professional retires, gets poached, or burns out.
Your AI models are trained on generic knowledge. Every competitor has access to the same training data. That’s not a moat. That’s a commodity. What competitors cannot buy is what I call your Context Graph: the accumulated record of why your best people make the decisions they make.
If you’re not capturing those decision traces right now, you’re watching your most valuable competitive asset evaporate in real time. Pair the 30-year veteran who has the expertise but won’t touch AI with the high-agency junior hire who knows the tools but lacks the lived experience. Together, they digitise your Context Graph. Neither can do it alone.
The rule: Data is a commodity. Context is a moat.
7. Ambiguity Problems
The client says they want “better reporting” but actually wants their boss to stop questioning their numbers. Defining your AI strategy when market signals are contradictory. Figuring out whether a regulatory change is a compliance burden or a competitive opportunity, when the answer depends entirely on how you choose to interpret it.
AI can explore options. It cannot resolve ambiguity where the real challenge is determining what the question actually is.
This is where the most expensive AI failures will come from. Not hallucinations. Not misclassified documents. A perfectly specified wrong strategy, executed at scale, fast enough that the damage compounds before anyone notices. AI agents will flawlessly execute exactly what you specify. Even if you specified the wrong thing.
The rule: Someone has to resolve the ambiguity before the agent touches it. That’s not optional. That’s the whole job.
The 90/10 Reality
Even in roles where AI-tractable problems exist, the tractable slice is typically about 10% of the actual work. The other 90% is emotional intelligence, ambiguity, judgment, coordination, and domain expertise.
The organisations that understand this ratio will deploy agents where they create value and protect the human capabilities where they must.
The organisations that don’t will automate the 10% and destroy the 90%.
Five Questions Before Your Next Deployment
Before another agent goes live, run a value destruction audit.
One. Are you deploying agents against effort problems with skeleton-crew oversight? Where are the human auditors?
Two. Are you putting AI in client-facing contexts that require emotional intelligence? What is the cost when it gets the tone catastrophically wrong?
Three. Are you using agents to execute strategy that nobody has properly specified because the underlying ambiguity hasn’t been resolved?
Four. Are you letting domain expertise walk out the door while investing in AI models trained on the same generic data every competitor uses?
Five. Do you know which of your problems are reasoning problems, which are effort problems, and which are courage problems? Or are you treating “hard” as one thing?
If you can’t answer those questions with specifics, you are almost certainly destroying value right now. You just can’t see it yet. Because the destruction shows up in eroded trust, regulatory near-misses, and departed expertise. Not on dashboards.
The Steering Wheel
Steven Schwartz didn’t need better AI. He needed the judgment to verify what the AI gave him. The National Eating Disorders Association didn’t need a smarter chatbot. They needed someone to recognise that human distress is not an effort problem. Michael Burry didn’t need a better model. He needed the courage to hold the position.
MIT’s interviews with 17 procurement and IT leaders established consensus: a strategic positioning window is closing between mid-2026 and early 2027. The organisations that use that window to build the human verification layer, to capture their Context Graph, to classify their problems before deploying agents against them, will have AI systems that create real value.
The organisations that skip that step will be faster at destroying it.
AI is the engine. You are the steering wheel.
The question is whether you know what kind of road you’re on.
The AIR APAC Readiness Scorecard maps where your organisation stands across six dimensions of AI readiness, including whether your current deployments are creating or destroying value. It takes 15 minutes.