This AI Pilot Was Stuck at 2.5% Click-Through. An Early Failure Helped Increase It by 25%. 

One of the most useful AI lessons I have learned as a consultant came from a project that initially did not work.

A mid-sized regional retailer had asked me to help it use generative AI to personalise customer marketing. The technology worked in the narrow sense: it produced emails quickly and at scale. The business result was another matter.

Click-through rates remained stuck at 2.5 percent. The messages sounded generic. Worse, the AI had drifted away from the company’s voice.

That experience reinforced something I now encourage leaders to establish before automating an important workflow: an AI failure budget.

A failure budget defines how much error, rework, customer exposure, and operational disruption an organisation is prepared to tolerate while an AI workflow learns. It forces leaders to decide what can safely go wrong before they become captivated by what AI can theoretically automate.

At the retailer, we did not respond to the disappointing pilot by buying a different model or declaring AI a failure. We examined what had actually gone wrong.

The customer segmentation was too broad, so the AI had too little context to produce genuinely relevant messages. Its writing was also substantially more formal than the company’s established brand voice.

Those lessons changed the workflow. We expanded the customer segmentation from 10 groups to 25 and recalibrated the writing style. Within three months, click-through rates increased 25 percent, conversion rates rose 15 percent, and sales attributable to email eventually increased 8 percent.

What mattered most to me was not the improvement itself.

It was that the early failure was cheap, visible, and reversible.

That is what a good failure budget buys you.

Automation Scales Mistakes as Fast as It Scales Productivity

Many organisations do the opposite. They automate first and discover the boundaries afterward.

That becomes particularly dangerous when an AI system moves from helping an employee make a decision to making or executing the decision itself.

Research on AI workflow design increasingly points in the same direction. The companies getting more value from AI are not simply giving employees better tools. They are redesigning workflows around where AI adds value and where human judgement still matters.

The Hidden Cost of Unclear AI Boundaries

I saw the cost of getting that boundary wrong in another engagement.

A healthcare association had begun using AI to generate first drafts of policy briefs. On paper, this seemed like an ideal use case. AI could accelerate research-heavy knowledge work while allowing staff to spend more time on strategy.

Instead, volunteer physicians found themselves correcting inaccurate clinical references and irrelevant regulatory examples. Committee meetings that should have focused on policy strategy became line-editing sessions.

The AI had saved time for one group by quietly transferring work to another.

That is an overlooked form of AI adoption risk. Research from BetterUp Labs and Stanford’s Social Media Lab has documented a related phenomenon they call “workslop”: polished-looking AI output that creates extra work for whoever receives it.

In our case, the technology itself was not the fundamental problem. Staff had felt pressure to demonstrate that they were using AI, but the organisation had not established sufficiently clear quality standards or involved the volunteer experts whose judgement ultimately determined whether a policy brief was useful.

We changed the approach. I facilitated a structured review of the workflow, followed by co-creation sessions involving staff and volunteer experts.

Instead of relying on a generic AI process, the team built an advocacy assistant around the association’s own legislative archives, position statements, fact sheets, and successful regulatory comments.

The larger lesson was that a failure budget must measure more than hallucination rates.

Before Automating, Ask Where the Error Lands

Before automating a workflow, I now encourage leaders to ask where an error lands.

Will an error create five minutes of internal rework, send a confusing message to a customer, expose confidential information, or directly influence a hiring, financial, or regulatory decision?

That is AI workflow risk in practical terms.

The acceptable failure rate should shrink as the consequences of failure grow.

For Asian startups, this discipline may be especially valuable because speed is often the competitive advantage. A founder can move from a working prototype to hundreds or thousands of automated interactions remarkably quickly.

Unfortunately, automation scales mistakes at the same speed it scales productivity.

The regional evidence already suggests why caution at the workflow level matters. NewInAsia recently reported that six of seven Vietnamese fintech companies it examined remained stuck in pilot mode, with data quality, leadership, and operational readiness emerging as bigger barriers than enthusiasm for the technology.

That startup AI adoption pattern should sound familiar to founders everywhere: getting a model to perform a task is easier than building an organisation that can rely on the result.

The More Autonomous AI Becomes, the Clearer the Stop Rules Must Be

Asia’s policymakers are moving toward a similar philosophy.

The ASEAN framework for responsible AI scaling emphasises governance, risk management, and appropriate human involvement rather than treating innovation and oversight as competing objectives.

Singapore has gone further as AI systems become more autonomous. Its 2026 governance framework for agentic AI stresses traceability, control, and meaningful human oversight.

That principle matters even to a five-person startup.

The moment an agent can communicate with customers, alter records, spend money, or trigger another system, somebody needs to know exactly when the machine should stop, and a person should take over.

This is why I would make a failure budget part of AI startup governance from the beginning rather than adding it after a company becomes large enough to have a compliance department.

For every automated workflow, define what good performance looks like, what kinds of mistakes are tolerable, who reviews exceptions, and what threshold pauses the automation.

Record failures instead of hiding them. Look for patterns. Then expand the system’s autonomy only after the evidence justifies it.

That is a more useful AI automation strategy than asking whether the latest model can complete a task without supervision.

Final Thoughts: A Failure Budget Is Really a Learning Budget

The hardest part of AI adoption has never been persuading a machine to produce an answer. It is persuading humans to use the machine in a way that improves the whole system of work.

That human challenge is central to the psychology of AI adoption I study and advise organisations on.

Founders understandably want AI systems that succeed. They should also design for failure.

A failure budget does not mean accepting poor performance. It means deciding in advance where experimentation is safe, how much risk the organisation can absorb, and what evidence is required before automation receives more autonomy.

The companies that learn fastest will not be those whose AI never makes mistakes.

They will be the ones that decide, in advance, which mistakes they can afford to make, learn from them while the consequences are still small, and refuse to automate beyond the evidence.

---

Adapted from The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).

Dr. Gleb Tsipursky is a behavioural scientist and CEO of AI consultancy Disaster Avoidance Experts. Dubbed the “Office Whisperer” by The New York Times, he advises technology-focused leaders on AI adoption, workplace transformation, engagement, and innovation. He is the author of eight books, including The Psychology of AI Adoption at Work: From Resistance to Results.

Read the Chinese article here.