Can you trust your AI?

An employee is promoted on the first of the month. The HR system records it that day. Payroll picks it up at the next cycle. Workforce planning is still running off a dataset from the previous quarter. Finance is tracking the same headcount in its own spreadsheet.

 

Every one of those systems is working exactly as designed. The business still cannot produce a single reliable answer to a simple question about what that person now costs.

 

Now ask an AI model to analyse workforce costs, or to recommend a compensation band for that role. It will give you an answer in seconds, well written and apparently confident. The difficulty is not with the model but with the information it has been asked to interpret.

 

Adoption is no longer the interesting question

The debate about whether businesses should be using AI is over. McKinsey’s most recent global survey found that 88 percent of organisations now report regular AI use in at least one business function, up from 78 percent a year earlier. What is more striking is the second number in that report: only 39 percent can attribute any measurable impact on profit to it.

 

That gap gets explained away as a technology problem, or a skills problem. In my experience it is usually neither. It is a foundation problem.

 

I have spent my career in high growth businesses, and one pattern repeats. Technology does not automatically simplify an organisation. Companies add systems as they scale, each one solving a real and specific problem, until the collection of solutions has quietly created a new one. HR runs on one platform. Finance runs on another. Payroll, recruitment, learning and compliance all have their own. Individually they work but collectively they make it surprisingly difficult to answer a basic question about the business.

 

AI does not resolve any of that; it scales it.

 

If your data is accurate, connected and governed, AI is genuinely powerful. It removes the "no joy" work, surfaces patterns nobody had time to look for, and gives people better information faster. If your data is fragmented, AI produces answers that look authoritative without being reliable. In payroll, compensation, compliance or workforce planning, that is a meaningful risk rather than an inconvenience.

 

What good foundations actually look like

The answer is not to slow down on AI but rather it is to build the conditions in which it can work properly.

 

That starts with information flowing between systems rather than being entered, reconciled and corrected by people who have better things to do. It means governance at the point data enters the organisation, rather than months later when two systems turn out to tell different stories and somebody has to work out which one to believe.

 

Then there is explainability, which I think is the part most businesses are underestimating. If a system tells a leader that attrition is rising, that leader should be able to see where the conclusion came from. If it recommends a compensation package, they should be able to trace the recommendation back to the underlying records. Without that traceability, you are not really making the decision; you are accepting one that has already been made for you.

 

The willingness to challenge

Explainability only gets you halfway. Being able to trace an answer is worthless if nobody in the room is willing to question it.

 

This is the part of AI adoption that is a leadership problem rather than a technical one. A well formatted output carries a lot of unearned authority. It is confident, it is fast, and it arrives without any of the hesitation a human analyst would show when they were unsure. Challenging it requires somebody to say, out loud and often in front of more senior people, that the numbers do not match what they are seeing on the ground.

 

Most organisations are not set up to make that easy. The junior person on the call is frequently the one closest to the data and the least likely to speak. If they have watched a previous challenge get waved away, they will not raise the next one, and the business will lose the early warning it needed most.

 

So the practical ask I make of my own teams is a simple one. If an output does not match what you are seeing in real life then say so before we act on it. AI surfaces information but we still need to embrace human judgment. Leaders have to model that themselves.

 

Where human judgment still belongs

The most valuable role for AI is to give people better information and more time in which to exercise their judgment.

 

That is particularly true in HR. People do not join the profession because they want to spend their days reconciling records across disconnected systems and chasing approvals. They join it to develop talent and to build somewhere people can do their best work. AI can take a great deal of that administrative load away, and the judgment stays where it belongs.

 

This is also why businesses should think carefully about the language they use internally. If the only conversation about AI is efficiency and headcount, employees will reasonably conclude it is a threat, and they will be slow to tell you when it is wrong. If the conversation is about removing the work nobody wanted in the first place, you get a very different response, and a much better one.

 

Someone is on the other end of the answer

Every output described here lands on a real employee. A pay band, a promotion, a restructure, a role that does or does not exist next quarter. That is a reasonable standard to hold the technology to, and a reasonable reason to insist on knowing where an answer came from.

 

Get the foundations right and AI will make you a much better decision-maker. But ultimately the decision still rests with the human. 

 

---

  

Wendy Harris is Head of EMEA at Rippling

 

Main image courtesy of iStockPhoto.com and bymuratdeniz