AI can generate endless possible options for you before you have even finished explaining the problem, to be honest. That is, of course, extremely useful. But it can also be very dangerous, because the quality of the answers we are being fed can make us forget who is actually responsible for the decision.
There are four decisions that AI can help inform, but should never be allowed to own. To understand why, we first need to discuss the difference between assistance and delegation.
Assistance is not delegation
Think of these in terms of actual roles in your business.
An assistant is someone who helps you. They analyse things, provide options, take care of some administration and give you information that helps you make a decision. That is assistance.
Delegation is where you give work to another person and expect them to own that piece of work and get it done. They take responsibility for it. With delegation, you expect a response of “done” rather than “here is some information that will help you get that done.”
When we are talking about an AI system, assistance means organising things, giving you information and providing tools for you to work with. Delegation would be telling an agent, “I want you to go and do this. Do not come back to ask if I approve. Just get it done.”
That distinction matters when we consider the following decisions.
Decision one: What problem is worth solving?
AI is very good at summarising meetings, documents, emails and customer feedback. It can produce precise and accurate summaries. It can identify patterns across massive amounts of data and information. It can review information and provide options and opportunities.
It can also compare alternatives. For example, you could ask whether you should buy product A or product B and have it explain the advantages and disadvantages of each. These are things AI is good at.
What is AI not good at?
It cannot properly understand how a problem or solution aligns with the purpose of your organisation or with your own personal purpose. It cannot decide which customers matter most.
In any business, you will have customers who are easygoing and high value. They may not need much handholding and might not even prefer it. You may also have a customer who is not particularly high value, or is somewhere in between, but requires a lot of work and assistance to maintain and make sure they sign again next month or next year.
There is a ranking you probably carry in your mind, or perhaps in your CRM, of who needs the most assistance and what that assistance should look like.
AI also cannot decide whether an opportunity is worth the hassle. There are salespeople who would say every opportunity is worth the hassle, of course, but that might not be true, particularly for a small business trying to scale quickly. You might not want to take on every opportunity. You might decide something is outside your area of expertise. Or you might decide, “We have never worked in this space before, but let’s try it.”
It cannot decide whether a problem should be solved at all. It might be a personnel problem that can be avoided by not hiring a particular person. Perhaps the business should not move into that market or geography. Maybe that decision makes life easier.
AI cannot make these decisions. Only an executive with the right context and information can make them.
Imagine that you have a CRM and ask AI to optimise the approval process. It could probably do a very good job. But an executive might look at the result and say, “Actually, several of these steps are not needed.” They might decide that the system being used for approvals is wrong, or that there should not be an approval process at all.
AI is usually better at answering the question you give it than determining whether you asked the right question. Remember that.
Decision two: What trade off or compromise is acceptable?
This can be a tough one.
Consider speed against accuracy. Depending on your business, your stage of growth and what you provide, you might decide that accuracy or quality is more important than speed. If you run a fast food restaurant, speed might be far more important than perfect quality or accuracy.
That decision needs to be made at an executive level and then communicated to the people doing the work. It is not something AI should decide.
The same is true of growth and privacy. Sometimes small businesses take liberties with privacy that a large enterprise could not. Their audit requirements are different. Their legislative requirements might also be different. This is definitely not something you leave to AI.
Do not ask AI whether a piece of information should be public. Do not ask it whether you need another layer of security. Do not let it decide whether a process should be automated or retain a direct personal touch.
Another common trade off is delivering quickly at the risk of maintainability. I have seen this many times during my years in software development.
Businesses with large applications often optimise for delivering features quickly. The CTO is responding to requests from the board or the user base and needs to keep delivering new features. Those features keep customers on board, prevent churn and give the marketing team something new to promote.
Maintenance, refactoring and small improvements that add up over time do not receive the same attention because their immediate value is harder to see. Important security maintenance can be delayed. CTOs need budgets for that work, but cannot always demonstrate an immediate return. The return is only visible when the application does not get hacked. That is important.
A trade off or compromise may involve a calculation, but the calculation is also a statement about the values and purpose of the business and where that business is at that moment.
Decision three: What does good enough mean?
Every day, as an executive in a business, you are answering the question of what is good enough.
You can continue working until something is near perfect but gain no additional value from it. At some point you have to say, “This is good. Ship it.”
Consider how many false positives are acceptable. On a production line, you can expect a certain number of flawed items that need to be discarded. You can continue improving that production line and reduce the number of flawed items. But eventually the cost of improving the process becomes greater than the money being lost on those imperfect products.
At that point, it might be easier to have someone check the items coming off the line and discard those that do not meet the standard.
AI will struggle to make that decision. If you instruct it that every paperclip must be a particular shape and colour, it will continue trying to improve the process until every paperclip matches. That might not be the best commercial decision.
You also need to understand the result of failure. What happens if you accept something as good enough?
If you are building a car, you cannot have brakes that are merely good enough. You cannot have a steering wheel that stays attached most of the time. The acceptable error rate depends on the consequence of being wrong.
AI cannot decide what those consequences mean for your business, your customers or the people affected by the result.
Decision four: Who owns the consequences?
This leads naturally to the final decision. Who takes responsibility?
When something goes wrong in a business, there is usually a clear chain of command that stops with the CEO. Ultimately, that is the person who owns responsibility for what goes wrong in the business.
If the business suffers financial harm, we can blame the accountants, the bookkeepers, the cashiers or even the CFO. But ultimately responsibility flows to the CEO.
If a customer is upset, the person in front of that customer needs to take some responsibility, and the issue then escalates through the business. Responsibility usually falls to the person who can do something about it or who was in a position to say, “This is not right. This is illegal. This does not work. This could cause further harm.”
Responsibility also includes deciding what level of performance is acceptable.
We cannot say, “AI made that decision, therefore I am not responsible.” It does not work. It is not going to work legally or ethically. It has not worked in the past and will not work in the future.
I am extremely sceptical that we will ever reach a point where a human can say, “I had nothing to do with that. It was the machine.” There will always be someone responsible. As humans, you and I need to make sure we own that responsibility.
AI can assist, but it cannot be accountable
A system, a computer program, and I include AI in this, can perform an action admirably. But it cannot accept responsibility for the outcome.
None of this means humans have to do everything manually or cannot get assistance from AI and other computer applications. In fact, we should be using them.
We simply need to know who owns responsibility at each point. We need to know how to reverse decisions and actions. We need to understand when an issue escalates from machine to human, how that escalation is triggered and where it goes. We need to know who is accountable and what policy is being followed. Make sure you have that policy in writing.
The four decisions come down to this. What is the actual problem that needs attention? What trade off or compromise is acceptable? What does good enough mean for you, your business, your users and your customers? Finally, who owns the result and takes responsibility at the end of the day?
It is not the machine.
You can use AI to widen your options, provide information and tools, challenge your assumptions and improve the information you have gathered. Do not use it to outsource your accountability and responsibility.
Even when organisations understand all of this, many still struggle to turn a successful AI demonstration into a working capability. That is why so many AI pilots die after the demonstration is done.
