It’s a fact that companies are going all out with AI. But they’re especially eager to use it for customer-facing applications and are taking a very cautious approach to its use behind the scenes. I honestly think this should be handled exactly the other way around.
Usually, they start with the website chatbot, the phone support agent, and the agent who answers customer emails. And the internal systems—the ones employees themselves would use every day—are put off until later because “they don’t look impressive.” No one brings up in a committee meeting that their staff now writes reports in half the time; instead, they showcase the bot that talks to customers and appears in the photo.
The problem is that you’re choosing the option where a mistake is costly and putting off the one where a mistake is much more manageable.
Two different AIs working at the same company
It’s important to make a distinction here, because we tend to lump everything together, and they’re not the same thing. On the one hand, there’s AI for internal use: the assistant that helps your team write, summarize a two-hour meeting, prepare an analysis, or understand a contract.
Here, there’s always a person reviewing the results before they’re released anywhere. If it makes a mistake, someone catches it. If it invents a data point, someone spots it. The error is reversible and inexpensive, and on top of that, every correction teaches the person something about where the tool is falling short.
On the other hand, there’s AI for external use: the kind that talks to your customer, interacts with your systems, and carries out tasks without anyone watching. And here we’re no longer talking about assistants, but about agents—which is something entirely different. An assistant makes suggestions; an agent decides what to do next and does it.
The difference isn’t one of vocabulary, but of cost and risk. Deploying an agent where automation would have sufficed comes at a double cost: on the bill and in the form of surprises.
And as soon as you have two, they start calling each other
Let’s focus on this now because we’re going to see how complicated everything really gets.
A single agent is manageable. You know what they do, what permissions they have, and who to ask if something goes wrong, but the thing is, no one sticks to just one role: you set up the agent who handles email, and once that’s working, you set up the one who qualifies leads. And this is where things get complicated… since the second agent needs data that the first one has, you connect them. And suddenly, you have three agents talking to each other, and there’s no longer a single person in the company capable of sketching out on a napkin what happens when an email comes in.
That’s where the problems arise and really mess things up—believe me, I’m speaking from experience.
One of them is the issue of inherited permissions. This happens when one agent calls another, and the second one runs with the first agent’s permissions—or vice versa—or with its own permissions that no one remembers granting. No one decided that; it just happened as a result of stringing together pieces that made sense separately but, when combined, are a mess.
Another issue is traceability. When something goes wrong and three agents were involved, who was responsible? If none of them keeps a record of what they did and why, the honest answer is “we don’t know,” and this in itself poses a huge auditing problem.
And the third one—the quietest one— is money. One agent calls another, who calls yet another—and it becomes a loop if no one steps in to correct it (I recently told you about the danger this poses). Ultimately, this translates into money wasted for no reason, and it eventually becomes a business problem and a cost overrun that’s discovered at the end of the month.
But I need to make this very clear: starting to build agents without any governance between them is madness that ends up costing you headaches and money.
What Happened to Me One Saturday
And I’m not just saying this because I heard it from someone else. Do you remember when I told you about that Saturday when one of my agents made a decision on his own? A scheduled task meant to remind me to follow up on an issue woke up on its own and started running a migration against the production database. Nothing was lost (I was actually lucky and had a backup), but it was the kind of scare you never forget.
But what mattered wasn’t the technical glitch it was realizing that I didn’t know I had a task capable of doing that. I was in my own home, with my own agents, which I had set up myself. Imagine what could happen at a company with two hundred employees where each department has set up its own system on its own. It could be a real nightmare.
From these experiences, a rule emerged that is now non-negotiable in my home: no agent may access production materials without explicit, immediate authorization from a person identified by their first and last name.
"Governance cannot be delegated"
I recently told you about a conversation I had at MIT with Paul McDonagh-Smith , and there was something he said that strikes me as overwhelmingly logical: “Governance and accountability cannot be delegated.”
Paul added an image that really stuck with me: “Autonomy isn’t a switch; it’s a dial. You don’t turn it on or off you adjust it.” And here’s what I find most useful of all: when you turn up the autonomy dial, governance doesn’t increase proportionally it multiplies. An agent with the dial set to 30% is a configuration; the same agent at 90% is a matter for the board of directors.
That aligns with what we’re seeing with the European AI Regulation, which has been mandatory regarding transparency since August 2. But don’t just focus on the fines read it for the parts that save you money. Because the questions an auditor asks are exactly the same ones you should be asking yourselves before turning anything on. Who can stop this? What happens when the permission check fails, allows access, or blocks it? Where is it specified what this agent can and cannot do?
If you don’t have those three answers, you don’t have a system—you have an experiment in production.
So, where do you start?
Let’s get down to business
I’d start with internal training, not just because it’s inexpensive it’s also what teaches your people how to work with this system. Six months of an intern put to good use, and you’ll have a team that knows when the system is making things up a team that’s developed a knack for spotting it. Those are the people who’ll later be able to review what a customer-facing agent is doing.
If you start on the outside, you’re asking a team with no prior experience to oversee the very system that’s the most dangerous. It’s like learning to drive right on the highway.
Next, before setting up any agent-based processes, clearly distinguish between what an agent actually needs to do and what can be handled by automation. Extracting data from an invoice isn’t an agent; it’s a classifier: it doesn’t make any decisions—it always does the same thing. It’s much cheaper, more reliable, and easier to manage—and above all, you should ask yourselves this question: Does this part of the process decide what to do next? If the answer is no, it’s not an agent, so don’t use it.
And finally, if you’re going to have more than one, write down the rules before there are two. Who can call whom, with what permissions, what gets logged, how much each person can spend per month, and who has the red button. It’s just two pages—not a project—and writing them after you already have six agents up and running costs ten times as much. I’m telling you this from my own recent experience.
My thoughts on this matter
I believe that the order most companies are following is exactly the opposite of what’s best, and that the reason isn’t technical but rather for appearances’ sake. What’s on the outside is visible, what’s on the inside isn’t, and since everyone is watching what their neighbor is doing, everyone just jumps on the bandwagon…
My advice—and I’m going out on a limb here—is to make plenty of noise internally and keep it quiet externally until you have the guidelines written down. No one has ever lost a customer by taking three more months to launch a chatbot. Customers are lost when the chatbot promises a discount that doesn’t exist or when an agent sends an email with someone else’s information.
And keep in mind, I’m not saying you shouldn’t set up customer-facing agents—what I’m saying is that the order matters much more than it seems, and that the boring part (who responds, who stops the process, who pays) needs to be done sooner rather than later, because once it’s put off, it never gets done.
And in your case, could you tell me how many AI-powered automations you have up and running in your company, and who set them up?
Leave your comments below—I’d love to hear from you.
Have a fantastic week!
