The Trust Algorithm: Alice Sesay Pope on What Amazon Taught Her About AI and Trust

I've spent about three decades in Fortune 100 leadership across Amazon, Visa, Capital One, USAA, Microsoft, and Johnson & Johnson. Most recently I was Global VP of Device, Digital, and Alexa Support at Amazon, leading 10,000 associates and serving more than 200 million subscribers worldwide.

I started my career as a chemical engineer, and that shaped how I think. Engineers solve hard problems; they learn to find where a system breaks under pressure and design something that holds. That's what drew me in. Customer experience is where a company's promises either become real or fall apart.

Right now, I'm focused on Trust Algorithm: How Leaders Build Trust with Generative AI, the book I just wrote. It's about what leaders keep getting wrong with AI. They lose sight of the customer experience from the customer's point of view, and they fail to build a change management plan that keeps associates engaged and reskills them for what comes next. That's the conversation I want to be having because over the year, I have seen the good, bad and ugly with technology transformation.

I named it Trust Algorithm on purpose, because I wanted to take trust out of the realm of the soft and intangible and put it back where it belongs, which is in the realm of things a leader can actually build. For most of my career, trust was treated like a feeling. You either had it with a customer or you did not, and there was very little you could do about it on a Tuesday afternoon. That is wrong. Trust has inputs, and once you know the inputs, you can engineer the output.

In the book, the algorithm comes down to a few components that show up the same way every time. The first is transparency, being honest about what the technology can and cannot do, so the customer is never surprised. The second is reliability, doing the thing you said you would do, the same way, every time, so the customer can predict the outcome before it happens. The third is access, making sure a human is always one step away the moment confidence breaks. And running underneath all of it is accountability, a clear answer to who owns the outcome when the system acts on a customer’s behalf. Those components get applied through one lens, which is stakes. The higher the stakes for the customer, the more of each component you owe them.

What makes it a framework and not a philosophy is that those components do not change. The technology changes constantly. The variables do not. Digital banking, streaming, connected devices, and now generative AI, I have applied the same inputs to all of them and gotten a reliable result. A philosophy tells you to care about trust. A framework tells you exactly which levers to pull on Monday morning, in what order, and how to know whether it is working. That is the difference, and that is why a leader can pick it up and apply it to a problem I have never seen.

The earliest signal almost never showed up in the trust scores first. By the time satisfaction or net promoter dipped, the erosion had been underway for weeks. Those are lagging indicators. They tell you trust already broke, not that it's about to.

But here's the honest part. The numbers gave me the “what”. My frontline associates gave me the “why”, and they felt the shift before it was ever statistically significant. I learned to treat that as real data, not anecdote. Trust rarely collapses in a single event. It leaks, and, if you only watch the lagging scores, you miss the leak entirely. Finally, there were worst scenarios where you don’t even know something is broken until you start receiving customer complaints.

I watched it happen most clearly with deflection. When you are under pressure to move customers into self-service. The business case is compelling. Every contact deflected from a human saved money, and the business scorecard rewards it. On paper, it looks like a win.

But when you do so, you are not taking comprehensive view of your business, your business is focused on  the wrong thing because you fail to understand whether you solved a customer’s problems. A customer who gave up counted the same as one whose issue was resolved. The system couldn't tell the difference between success and surrender. In doing so, you are quietly training your best customers to stop trusting and the efficiency numbers applauded the whole time.

Catching it took putting deflection next to repeat contacts and churn. Then the trade-off was obvious. The customers deflected came back angrier, through costlier channels, or not at all. Real savings in the quarter, a liability over the year.

The lesson is simple. Efficiency metrics are seductive because they're clean and immediate, and the trust cost shows up later. Customers want an experience that is simple, fast, accurate that meets their needs.

The mistake leaders make is treating this as a channel decision when it is really a stakes decision. The question is not which interactions can be automated, because technically almost any of them can. The question is what the interaction means to the customer in that moment, and what they stand to lose if it goes wrong.

When you look at it that way, the line gets clearer. Automation should carry the routine, low-stakes moments where the customer mostly wants speed. Checking a balance, tracking an order, resetting a password. There, fast and accurate is the relationship. But when there is money, fear, or vulnerability involved, the customer is not just solving a problem, they are finding out whether you have their back. A fraud alert, a denied claim, a service failure that already cost them something. Those moments build or break trust, and you protect human access to them fiercely.

The test is simple. Ask what the customer stands to lose if the automated path fails them, and how easily they can reach a human when it does. If the cost of failure is high and the path to a person is hard, you are not saving money, you are quietly transferring risk onto your most vulnerable customers and calling it efficiency. When the savings point out one way and the relationship the other, weigh the savings you see this quarter against the trust you spend over the year, because the trust cost is just as real, it simply arrives later.

The principle underneath it all is that automation should remove friction, never remove the exit. A customer who knows a human is one step away will trust the automation more, not less. What erodes trust is not the technology itself - it's taking away the exit.

The first thing I would tell a leader who feels behind is that the pressure is real, but speed is not the same as progress. Moving fast on a broken process just helps you fail faster and at greater scale. So before you automate anything, start by understanding the experience you have today from the customer's point of view. Map where customers struggle, where they lose confidence, where they come back a second and third time. That tells you what is worth automating and what is just noise.

Begin with the high-volume, low-stakes interactions where speed is the value and the cost of an occasional miss is small. That is where generative AI earns trust quickly and gives you room to learn. The guardrails and escalation paths you build there become the foundation you rely on when you move to harder ground.

Do not start by automating anything that touches a customer at their most vulnerable time. The denied claim, the fraud event, the moment something has already gone wrong. Those moments matter, and they are the worst place to learn your model is not ready. Automate those last, and only after you have proven the technology where the stakes are lower.

Bring your associates along on the automation journey. The leaders who skip the change management end up with a workforce that quietly resists the very tools they deployed, and customers feel that resistance long before any metric catches it. 

The biggest lesson from those earlier shifts is that customers do not actually resist new technology. They resist losing control and the uncertainty that comes with new technology. When digital banking arrived, the fear was never the app itself, it was wondering whether your money was still safe when you could no longer see a teller. When streaming and connected devices scaled, the anxiety was not the feature, it was whether the thing would work when you needed it and whether someone would help if it did not. Technology has changed every time. The underlying human question never did. Can I trust this, and will you be there when it fails.

Most leaders are learning the hard way with AI right now. They treat each wave as a new problem and obsess over capability, when the truth is that trust is built the same way, it always has been. You earn it by being transparent about what technology can and cannot do, by being reliable enough that customers can predict the outcome, and by making it easy to reach a human the moment confidence breaks. Those principles did not change because technology got smarter.

The second lesson is that trust is in the failure moments, not the happy path. Any technology looks trustworthy when everything works. What customers remember is what happened the one time it did not.

The reskilling almost never happened in the right sequence. That is the honest part, and it is exactly why I write about it. I write about reskilling because I know what goes wrong when it comes last instead of first. The technology lands, the work changes that same week, and the training, if it comes at all, comes months behind. By then the damage is done. Associates have watched their jobs change without ever being told what they were changing into, and they have quietly disengaged. Once people decide the change is being done to them rather than with them, no training module wins them back.

Done right, the sequence is reversed. The reskilling starts before the tools arrive, not after, and it starts with honesty. The first move is to name the change out loud and early. You tell associates plainly that the routine work they do today is going to be handled by the system, and their job is going to change, because they can already see it coming and the fastest way to lose people is to insult their intelligence. Then, just as plainly, you show them what it is changing into, which is almost always a more valuable job and not a smaller one. The role moves from handling the volume to handling the judgment, the escalations, the moments when a customer is upset or vulnerable and a human still has to step in. That is harder work, and it is worth more, but only if people can see it coming.

The concrete part is building the path, not just describing it. That means mapping the new skills the future role actually requires, the judgment, the empathy, the ability to supervise the system and catch it when it gets something wrong, and then giving people a real way to get there. Time carved out to learn, not a course squeezed in after a full shift. Coaching, not just content. A career ladder attached to the new skills so the effort pays off in someone’s own progression. Bring associates into the design early, because the people doing the work always know where the system will break before any executive does, and being asked turns them into owners of the change instead of casualties of it.

On keeping people engaged, the honest answer is that you do it by being honest. Associates do not resist change because they are afraid of technology. They resist because they are afraid of being discarded by it. When you show them a future where they are still needed, where the new role is more human and not less, and you put real investment behind getting them there, most people lean in. The leaders who lose their workforce are the ones who automate quietly and hope nobody notices the sequence was wrong. The workforce always notices, and the customer feels it right after.

When AI acts rather than just answers, the whole risk equation changes, because a wrong answer disappoints a customer, but a wrong action costs them something. A misinformed response can be corrected in the next sentence. A payment moved, a claim closed, an account changed, those have consequences that impact the customer, and the trust you lose when the system acts wrongly is far harder to rebuild than the trust you lose when it simply answers wrongly.

So, a few things become non-negotiable. The first is that every action the system takes must be reversible or contained. Before you let AI act, you must know how a customer undoes it, how quickly, and how easily they can reach a human to help them do it. If you cannot answer that, you are not ready to give the system full authority. The second is that authority must be matched to stakes. An agent can reschedule a delivery on its own. It should not be closing a claim or moving money without a human in the loop, because the cost of being wrong in those moments is borne by the customer, not the company. The third is accountability. When the system acts, you must know who is accountable for that action, because a customer harmed by an automated decision does not want to hear that the algorithm did it.

Where most organizations will underestimate the risk over the next 12 to 18 months is in the gap between what Agentic AI can do and what it should be trusted to do unsupervised. The capability is arriving faster than the governance, and the pressure to deploy is enormous. Leaders will grant these systems broad authority because they perform beautifully in the demo and across the routine cases, and they will discover the edge cases in production, on real customers, at scale. The failure will not look dramatic. It will be thousands of small wrong actions compounding quietly before anyone connects them to the trust erosion showing up downstream.

The risk leaders will underestimate is the one inside their own organization. When the system starts acting, the role of your associates’ changes from handling the work to supervising it, and that is a genuinely different job that requires real reskilling. The organizations that move fast and skip that step will find their people unprepared to catch the system when it acts wrongly, which is exactly the moment that matters most. The technology acting on a customer's behalf is only as trustworthy as the human still accountable for it.