CIOs Need a Better Way to Account for AI Work in IT

For the last two years, CIOs have had a simple goal for AI in IT: get it into the organization, let teams experiment and see where it could improve service delivery.

IT organizations bought licenses, ran pilots and added AI features across already bloated software stacks. That made sense while teams were still figuring out what AI could do. Now that spending sits inside real budgets, and CFOs want to know what the company is getting for its money.

The problem is that many of the numbers CIOs use to measure AI do not answer that question. CIOs can show how many employees use AI, how often they use it and how much time they say it saves them. Finance wants to know whether the company spent less money, got more work done or avoided hiring more people.

Saving an employee five hours a week sounds valuable, but the company still needs to know what happened to those five hours. If the employee handles more work, or the team can support a growing workload without hiring another person, the company can see the financial result. If nothing else changes, the time savings may never show up in the budget.

As companies spend more on AI, CIOs need to get much more specific about the work AI does and what that work costs.

Time Saved Does Not Necessarily Mean Money Saved

A lot of AI ROI calculations start with time as a criteria – i.e., if AI saves an employee 30 minutes a day, a company can multiply those minutes across thousands of employees and come up with a very large number.

But business processes rarely work that neatly. Work moves from one person to another, waits for approvals and sits in queues. If AI saves ten minutes on one part of a process and the work then waits six hours for someone else, the company has not gained ten minutes of useful capacity.

For CIOs, the more useful question is whether AI changes the performance of the service itself. If access requests move through the queue faster, fewer tickets need human handling or the cost per request falls, IT can measure the result using the same operational benchmarks it already tracks.

Aviva is a good example. The insurer used more than 80 AI models as part of changes to its claims operation and reported cutting liability assessment time by 23 days and complaints by 65 percent. Those numbers tell management what changed in the operation and give finance something it can measure.

CIOs need to apply the same thinking to their AI investments. Before expanding a pilot, they should know what part of the business they expect to change and how much work AI needs to take on for that to happen.

The Work AI Completes Is a Better Place to Start

Consider a common IT access request. A copilot might summarize the ticket and draft a response, while an IT agent checks company policy, verifies the requester, gets approval, changes the permission and closes the request.

The copilot may save the employee a few minutes, but most of the work still belongs to the employee.

Now consider an AI system that can check the request against company policy, collect missing information, route the request for approval and make the change once it receives that approval. IT can see how many requests the AI system completes, how often an employee needs to step in and how much human work each request still requires.

That gives CIOs a much clearer way to account for AI. Instead of estimating how many minutes the technology saved, they can measure how much work it completed and what it cost to complete that work.

The same approach can apply to employee onboarding, internal support and incident response. The important part is giving AI responsibility for work that the company already understands and knows how to measure.

But You Need to Count the Human Work Too

An AI system may start 1,000 onboarding requests and require an employee to step in on 600 of them. Calling all 1,000 automated would give management the wrong picture of how much work the AI actually handled.

Those interventions cost money. If an agent works on a request for several minutes and then hands it to an employee who spends another twenty minutes finishing it, those twenty minutes belong in the cost of the process.

That means IT needs a record of what happened during each job. It needs to know what the AI did, which systems it accessed, where it needed approval, when an employee stepped in and whether the work actually got completed. Those records can feed traditional service desk measures such as mean time to resolution, while adding measures that matter specifically for AI, including intervention rates and cost per completed task.

Security teams need much of the same information when something goes wrong. If an AI system changes the wrong permission or takes an action it should not have taken, they need to reconstruct what happened and understand why.

Keeping that record gives CIOs a way to measure both sides of the investment. They can see the cost of running the AI and the human labor still required to supervise and finish its work.

How Much Responsibility AI Gets Affects the Economics

CIOs also have to decide how much authority they are willing to give AI.

An AI system with limited permissions may need an employee to approve many of its actions. That keeps a person involved, but it also means the company continues paying for more human labor.

Giving the system more authority may allow it to complete more work on its own, but that also gives IT more risk to manage. A system that can create an account or change a permission can cause more damage when it makes a mistake than one that can only recommend what an employee should do. Cost can also rise unexpectedly when an agent gets stuck in a loop, retries the same task or burns through compute because an instruction was too vague. Those failures belong in the economics too.

Companies will draw that line differently depending on the work involved. What matters for the ROI calculation is knowing where the line sits and how much human involvement it creates.

If an AI system can safely complete 90 percent of a process, its economics may look very different from a system that needs an employee every other step.

Start With Work That Already Costs Money

CIOs planning their next round of AI spending should start with work that already consumes meaningful time or headcount.

Access requests, employee onboarding, incident response and internal support all give IT something concrete to measure. The company already knows roughly how many requests it handles, how much employee time they require and where work tends to get stuck.

From there, IT can determine which parts AI can handle safely and where an employee still needs to get involved. It can then decide what it expects the investment to accomplish.

Maybe the team wants to handle 30 percent more tickets without hiring. Maybe it wants to lower the cost of each access request. Maybe it wants employees to spend less time on routine support so they can handle work that currently sits in a backlog. Once IT knows the full cost of completing a task with AI, including human intervention, it can set a target cost for that task and decide whether automation is actually worth it.

Once AI starts doing the work, CIOs can compare what they expected with what actually happened. They can see how much work AI completed, how often people had to step in and what the whole process cost.

That gives CIOs a much simpler way to talk about AI spending with finance. Instead of trying to put a dollar value on thousands of estimated minutes saved, they can show what work AI handled, how much human labor it still required and whether the company spent less money to get that work done.