Beyond prompts and context

Commonwealth Bank of Australia cut 45 customer service roles last year, replacing them with an AI voice bot that was supposed to reduce call volumes by 2,000 per week. Call volumes went up. Wait times exploded. Managers had to start taking calls themselves. Within weeks, CBA publicly admitted the "error" and rehired everyone.

IBM replaced thousands of HR roles with AI, then quietly started hiring humans back after realizing they'd lost the institutional knowledge and judgment that made those functions actually work.

Klarna's AI customer service agent saved the company $60 million — then the CEO went on Bloomberg to explain why they were hiring the humans back. The AI optimized for fast resolutions. The company needed judgment, not just speed.

Three different companies. Three different industries. Same pattern: the AI was technically working, but it was optimizing for the wrong thing.

This isn't unusual. It's becoming the norm. 95% of generative AI pilots fail to deliver measurable business impact (MIT). 42% of companies abandoned most AI initiatives in 2025, double the prior year (S&P Global). 74% of companies globally have yet to show tangible value from AI (BCG). And 55% of employers who laid off workers for AI now regret it (Forrester).

These aren't numbers about bad technology. The models are extraordinarily capable. These are numbers about organizations that haven't figured out how to connect AI capability to organizational purpose.

At Ripio, we've spent twelve years building crypto infrastructure across Latin America. I've been coding since I was eight. I build with AI daily — not as a CEO who delegates this to a committee, but as someone who ships things with it. And what I've learned over the past year is that the gap between AI potential and AI value isn't a technology problem. It's an organizational design problem.

Here's how I think about it.

Most companies are working on one layer of AI. They need three.

Layer 1: Prompt Engineering — telling AI what to do. This is the most familiar layer. You type an instruction, the AI follows it. The quality depends on how well you write the prompt. It's individual, session-based — one person, one task, one conversation. Useful, but the ceiling is whatever one person can extract from one interaction. It's like giving a new hire a single instruction with zero context about the company.

Layer 2: Context Engineering — telling AI what to know. This is where most of the industry sits today. Instead of just giving AI an instruction, you connect it to your company's knowledge — documents, customer data, policies, product information. The AI can now give much more informed answers. Context engineering is necessary. But as CBA, IBM, and Klarna all showed us, it's not enough. An AI with access to all your company's information can still optimize for the wrong objective.

Layer 3: Intent Engineering — telling AI what to want. This is the layer almost nobody is building for, and it's the one that matters most.

Intent engineering is the practice of encoding your organization's actual goals, values, tradeoff rules, and decision boundaries into your AI systems. Not as vague instructions, but as structured rules the AI can act on.

Here's the difference: context gives AI knowledge — "here's our refund policy." Intent gives AI judgment — "when a long-term customer is frustrated, prioritize the relationship over the policy, up to this threshold, then escalate."

When a new employee joins your company, they absorb these judgment calls over months. They watch how managers make decisions. They learn which values the company actually prioritizes when tradeoffs arise. They internalize unwritten rules. AI can't do any of that. It needs this organizational wisdom made explicit from day one.

This is what was missing in every one of those cautionary tales. They had great prompts. They had comprehensive context. They didn't have intent.

So what are we actually doing about it?

At Ripio, we've identified three pillars we need to build. We're early, and this is hard work. But I want to share the framework because I think most companies are only working on one of these — if any.

First: unified context infrastructure. Organizational knowledge — product information, compliance rules, customer data, operational playbooks — lives in dozens of systems across teams, markets, and languages. If you deploy AI tools without connecting them to a single, well-organized source of truth, each tool operates with a partial and potentially contradictory picture of who you are. We're building a governed, current, cross-departmental knowledge layer that our AI systems can reliably draw from.

Second: coherent AI workflow architecture. Right now across most companies, different people use different AI tools in different ways. One person's clever AI workflow is invisible to everyone else. Nothing is standardized, transferable, or measurable. We're mapping our workflows and classifying them — which can be fully AI-handled, which work best as AI-assisted human work, which must stay human. Then building shared tools and processes around those decisions, evolving as capabilities improve.

Third: organizational alignment — intent engineering in practice. For our critical workflows, we're defining explicit decision boundaries (what AI can decide alone vs. what needs a human), tradeoff rules (when competing goals conflict, which wins and under what conditions), and escalation triggers (what situations always get flagged for human review, no matter how confident the AI is).

This third pillar is the hardest. It requires making tacit knowledge explicit. It means sitting down with experienced team members and extracting the judgment calls that live in their heads — the ones they've developed over years of navigating real situations. It's uncomfortable work, because a lot of organizational wisdom has never been written down.

What's given me conviction this matters is what I'm seeing inside our own company.

Someone on our team rebuilt a complete internal tool in two days using AI — a tool we'd been asking to have built for two years. One of our developers, working with AI coding agents, completed nearly an entire sprint's worth of team tickets solo. Not by working harder — by working differently. We've seen non-developers on our customer experience team build sophisticated internal tools — thousands of lines of functional code — because AI lowered the barrier from "years of training" to "clear thinking about the problem." Two squads are close to completing their entire annual roadmaps before the end of Q1. We keep discovering unplanned work to take on because teams have extra capacity they didn't expect.

The productivity unlocks are real. But they're also precisely why intent engineering becomes urgent. When AI makes execution dramatically faster, you need to be even more precise about what you're executing toward. Speed without direction is just faster drift.

In fintech, this isn't abstract.

An AI that optimizes client onboarding for speed instead of compliance rigor creates regulatory risk. An AI handling cross-border operations that doesn't understand regulatory boundaries between jurisdictions creates real exposure. An AI managing institutional relationships that treats every interaction as a ticket to close — rather than a relationship to build — erodes the trust that took years to earn.

Only 21% of organizations have a mature model for governing autonomous AI agents (Deloitte). In fintech, that's not a statistic. That's a warning.

The race isn't about who has the smartest model. It's about who has built the organizational infrastructure for AI to operate with a clear understanding of what the company is actually trying to accomplish.

We're not done. We're not even close to be done. But I believe the companies that get this right — that invest in context, workflow design, and especially intent — will be the ones that capture real, durable value from AI. Everyone else will keep cycling through pilots that impress in demos and disappoint in production.

The technology is ready. The question is whether our organizations are.

One more thing. Building this kind of company — one that's redesigning itself around AI at the infrastructure level, not just bolting it on — requires people who are comfortable operating in territory that doesn't have a playbook yet. If you're the kind of person who gets excited about that, and you want to work at the intersection of crypto and AI in Latin America, reach out. Ripio is always looking for people who'd rather build the frontier than read about it.

What's the hardest part of making AI actually work in your company? I'm genuinely curious whether others are hitting the same wall we did — and how you're thinking about the intent layer. Let's compare notes.

Originally published on X.