Shadow AI: when banning AI makes it riskier

Carlos: Eleven at night. A vice-president at a bank, a regulated bank, has a board meeting at eight in the morning, and the deck isn't finished. At the office, every AI tool is blocked. So she forwards the board pack to her personal email, opens it on her laptop at home, and pastes it into her personal AI account. Twenty minutes later the summary is done, and it's good. She goes to bed.

Mite: And the bank will never know.

Carlos: The bank will never know. I made that story up. But tell me you didn't recognise it.

Carlos: Welcome to Inteligencia de Negocios 20/20. I'm Carlos Miranda Levy, Coordinator of Impact, Innovation and Intelligence.

Mite: And I'm Mite Nishio. I've worked in technology and digital transformation in the Dominican Republic for more than three decades.

Fabiola: I'm Fabiola Herrera, Digital Transformation, AI & Cybersecurity Advisor. I've spent more than thirty years in digital transformation, payment systems, cybersecurity and financial regulation.

Carlos: We called this show «20/20» because that's the job: to see clearly. Not what the policy says is happening, but what's actually happening. So for episode one I want to start with a provocation. Mite, I think most AI policies in regulated organisations are fiction.

Mite: I'd say it a little differently. The policy describes what the organisation wants to happen. Practice shows what people need. When those two are far apart, there's a gap, and the gap is where the risk lives.

Carlos: And that gap has a name. Shadow AI.

Mite: Let me define it first, because the term gets used loosely. Shadow AI is using AI tools for work without the organisation knowing, approving or controlling it: personal accounts, personal devices, free apps, browser extensions. As a pattern it's not new. We had shadow IT for years: personal storage, chat groups for client work, people's own phones.

Carlos: So what's different now?

Mite: What the tool does with what you give it. A shared folder stores your document. An AI tool reads it, processes it and, depending on the plan and its terms, may keep it.

Carlos: Now let me question the picture everyone has in their head. When you say «Shadow AI», people imagine the intern, the junior analyst.

Mite: But it goes all the way up.

Carlos: All the way up: managers, directors, VPs, presidents. And who holds the most sensitive information in any organisation? The board pack, the merger file, the regulator's letter. The same people with the least day-to-day oversight of their own tools.

Mite: And in practice it looks very ordinary: emailing a file to yourself, a phone photo of a screen, a contract pasted into a chat window, a personal coding assistant working on company code at home. None of it feels like a security incident while you're doing it.

Carlos: How widespread is it?

Carlos: So here's my question, and I want to challenge the logic, not anyone's decision. Why does Shadow AI grow exactly where AI is banned hardest?

Mite: Because a ban closes the channel but not the need. People see what these tools can do. They see others moving faster, and the gap feels personal. If the institution closes the official door, the demand doesn't disappear. It finds another way in.

Carlos: Prohibition doesn't remove AI from the organisation. It just moves it out of sight. And I'll say something I've said before, because I believe it more every month: «If we try to make change look like the past or respond to the expectations and structures of the past, it's not true change, is it?» A ban is an attempt to make the future look like the past.

Mite: And the sectors that ban the most are the regulated ones: finance, health, government, telecommunications, legal. So the shadow is deepest exactly where the data is most sensitive.

Carlos: Mite, you spent years inside regulated institutions, in banking and in telecoms. Why is the first reflex always «no»?

Mite: It's an understandable reflex. A regulated institution is built to see risk first; that's its responsibility. Saying no protects you in the short term. The problem is that "no" also carries risk, only you don't see it in any report. And there's a difference between digitising and transforming: you can block a tool, but you can't block a change in how people work.

Carlos: So «no» feels like zero risk.

Mite: It's a risk you've decided not to measure.

Carlos: Fabiola, you sat on the other side of the table at the Central Bank. What does a regulator see that we don't?

Fabiola: A regulator answers for the trust people place in the whole system, so the first question is what happens when it fails. But a rule nobody can follow protects no one. People are people, and they will do what they need to do. The job is to write rules for people as they are, and to believe they can do better.

Carlos: Then let's measure it.

Mite: Let's be concrete, because a board needs a list, not a feeling. First: the data leaves the perimeter. Client data, financials and trade secrets end up on devices and accounts the organisation doesn't control, can't wipe and can't audit.

Carlos: Second: there's no audit trail. If a regulator asks «did any client data go to an external AI service?», the honest answer is «we don't know».

Mite: Third: terms nobody read. Consumer plans and enterprise plans can treat your data very differently: how long it's kept, whether it's used for training.

Carlos: Fourth: regulatory exposure. Data protection, banking secrecy, sector confidentiality, cross-border transfers.

Fabiola: Exposure is about trust before it's about fines. I put it this way once: “Without trust there is no adoption. And without adoption there is no transformation.”

Mite: Fifth: quality and accountability. An output nobody checked feeds a real decision, and nobody knows AI was involved. If the summary was wrong, who's responsible?

Carlos: And sixth, the one nobody puts on the risk register: the organisation learns nothing. Our VP got her deck done in twenty minutes, and the bank will never find out that the process could take twenty minutes for everyone. Every gain stays private: unmeasured, unrepeatable, invisible.

Mite: That's the part I'd underline for any executive. It's not only a security problem, it's a knowledge problem. Information is being processed every night, and none of it becomes organisational knowledge.

Carlos: All the exposure, none of the learning.

Mite: And notice that all six share one thing: you can't manage what you don't measure. Today most organisations couldn't put a single number on any of them.

Carlos: And I'd add a seventh that no dashboard will show you: the scenario where your competitor stops hiding it and starts governing it. What does your market look like in two years if they learn and you don't?

Mite: That one's harder to measure.

Carlos: That's exactly why nobody's looking at it.

Carlos: Now I want to provoke you, Mite, because this is where my own thinking has moved this year. Look at what our VP actually did. She used AI to do the same thing she always did, only faster.

Mite: Same process, new tool.

Carlos: And my position, since February, is this: AI will replace us. Yes, it can augment us. But if we insist on doing what we did before, only faster, cheaper, more comfortably or better, we will inevitably be replaced. The road forward is to reinvent ourselves. So the shadow user is speeding up exactly the work that AI will end up doing instead of her.

Mite: I share the diagnosis about the process. Changing the tool without changing the process and the culture brings very little. Where I'd be more cautious is the word «replace». I'd separate tasks from people: many tasks will be automated, and what happens to the people depends a lot on the decisions organisations and countries make now, especially in training.

Carlos: And that's exactly why I say reinvent, not resist. My grandfather was a blacksmith, so famous that people came from other towns on horseback to have him shoe their horses. Then the automobile arrived and transformed his trade. Shoeing horses faster was never going to be the answer.

Mite: So the question for the bank isn't «how do we stop people summarising decks?».

Carlos: It's «what can our people do now that wasn't possible a year ago?». That's the question a ban never lets you ask.

Mite: And this is where we usually part ways, you and I. You're the economist, so you go to scenarios: what if, what could be, the moonshot. I'm the engineer, so I ask how we'd know it's working. A reinvention without indicators is a hypothesis.

Carlos: And an indicator without a scenario only measures the past. Measure any blacksmith's forge at its peak and every number says «keep going», right up until the automobile arrives.

Mite: Fair. Then we need both: the scenario to know where to go, and the indicator to know if we're getting there.

Fabiola: You two can argue about the scenario and the indicator. I'll ask who it's for. People are people: tired, afraid, wanting to get home, so they take the shortcut. But I have always believed we can make people better, and that is what all of this should be for.

Carlos: That's why this show has three hosts.

Carlos: So, ban it or allow it?

Mite: We've just seen that banning it fails. Allowing everything fails too, with no data rules, no standards and no accountability. That's simply Shadow AI with permission.

Carlos: So the answer is neither, or both: allow it, and govern it, done right. My own principle here is the one I've lived by for years. NEVER HELP: Engage, Enable, Inspire, Empower and Connect. Don't «help» your people use AI from above. Engage them, enable them, inspire them, empower them, and connect what they learn.

Mite: And let's make it practical. I'd put it in steps.

Mite: One: discover before you decide. Find out what people actually use, not with a hunt, which only pushes things further into the shadow, but with something closer to an amnesty: «Tell us what you use and for what. We want to learn.»

Carlos: That's «engage». And that's where you'll find your best use cases, already proven by your own people.

Mite: Two: sanctioned tools, with enterprise data terms, so the safe path is also the easy one. If the approved tool is harder to use than the free one on a phone, people will use the phone.

Carlos: «Enable». The sanctioned path has to be easier than the shadow path. That's the whole game.

Mite: Three: clear data classes, explained on one page: public, internal, confidential, restricted. A person should be able to decide in five seconds what can go where.

Mite: Four: permission-aware access. The assistant reads only what each person can already open. It doesn't become a back door to the whole company.

Mite: Five: a person decides and signs. AI drafts and analyses, and a human answers for the decision. This matters more as we move from assistants that answer questions to agents that execute workflows. Those need supervision, access controls and clear responsibility.

Carlos: Six: work through the committees you already have: risk, security, data protection. Don't build a parallel AI bureaucracy.

Mite: And seven: train people and measure. Set a baseline before the pilot. Then time per task, error rate, rework, adoption, satisfaction. If you can't show a number after ninety days, you have an experiment, not an adoption.

Carlos: And alongside the dashboard, I'd keep one page that isn't a number: three scenarios for what the organisation could become if this works. The baseline tells you where you are. The scenarios tell you what's worth measuring next.

Mite: As long as each scenario eventually gets its indicator.

Carlos: Deal.

Carlos: And on training, I'll push one step further, because I care about this deeply: it's about the learning experience, not the teaching experience. Don't design a course about AI. Design the conditions for your people to learn by doing, safely, and then let them teach each other. That's «inspire», «empower» and «connect».

Carlos: Mite, where does this land for the Dominican Republic and Latin America?

Mite: We have an extra reason to get this right. If our organisations only consume AI from outside, through personal accounts and with no strategy, we also lose the chance to build our own talent and capabilities. And SMEs need especially simple, affordable rules. They don't have large compliance departments. Governance has to fit their size.

Carlos: A short break, and I'll tell you plainly what it is: a word from our own house. ibizai is part of the CEMI ecosystem that Mite and I work in.

Carlos: If this episode sounded like your organisation, ibizai's AI Adoption and Transformation Strategies starts where we just left off: diagnose, transform, scale. It begins with a readiness assessment across eight dimensions, with a free online self-assessment to start. Then comes a 90-day programme, week by week, in four phases: assessment and discovery, strategy design, implementation and pilots, and scale and optimise. Governance is built in: permissions reviewed before every data connection, work done within your existing risk, security and data committees, and a person who decides and signs. Results are measured in each engagement, never promised. ibizai.ai, under Services, AI Adoption.

Carlos: Let's close with something useful for each person listening. Mite, you go first.

Mite: Presidents and boards: assume Shadow AI already exists in your organisation. Don't ask for a ban. Ask for the map: who is using what, for what, with which data.

Fabiola: Get into the classroom yourselves. I came to understand that a leader has to be as prepared as the team. Or more. I started on punched cards, and I'm still learning.

Carlos: CIOs and CISOs: make the sanctioned path the easiest path in the building. And dream a little: ask what the organisation could do with AI that it couldn't imagine before.

Mite: Then put a number next to the dream.

Fabiola: Regulators: clear guidance that enables governed use reduces more risk than silence or a blanket prohibition. Institutions need to know what «done right» looks like. And remember: technology changes every year; the judgment to govern it does not, and that includes AI.

Carlos: And if you're the person forwarding files to yourself at eleven at night: you already know what this technology can do. Take that to your organisation. Go to your boss tomorrow and say «here's what I've been doing, and here's what it saved». You might be the champion they've been missing.

Mite: And everyone, ask a second question. Not only «how does AI make this faster?», but «is this process still the right one?»

Carlos: Or, in my words: «Do not artificially truncate, limit, mutilate innovation. Be part of it, accompany it.» Govern it, and reinvent with it.

Carlos: That's our first episode of Inteligencia de Negocios 20/20. If it helped, share it with the person in your organisation who needs to hear it. You know who it is.

Mite: Every episode comes out in Spanish and in English, with the full transcript, at podcast.cemi.ai. Subscribe wherever you listen.

Carlos: Next time: (topic to be decided)

Mite: Until then, see clearly.

Carlos: 20/20.