Where AI Reduces Cost

In 2025, Thomson Reuters asked professionals how much time AI would give back to them. Legal professionals expected to free up nearly 240 hours a year each, up from 200 in the previous year’s survey, and the report put the value at an average of 19,000 dollars per professional¹. Those are professionals’ expectations, not measured results.

Take that number into a boardroom, as a simple illustration. Imagine a firm with 200 lawyers: it multiplies 19,000 by 200 and arrives at 3.8 million. Has the firm’s cost fallen by 3.8 million?

It has not, and the reasons are worth slowing down for. The lawyers are salaried, so the payroll is the same whether they spend those hours drafting, thinking or going home earlier. In the United States, the American Bar Association’s ethics guidance adds a twist that is particular to professional firms: lawyers who bill by the hour may bill for the time they actually worked, not for the time a tool saved them². For an hourly firm, faster work can lower revenue before it lowers any cost. The 3.8 million is a measure of capacity. Whether any of it becomes a lower bill depends on decisions nobody has made yet.

That gap, between the hours a technology releases and the money an organization stops spending, is where many AI cost cases succeed or fail.

The core idea

AI reduces cost when it lowers the resources needed to produce the same business outcome, or a better one, at acceptable quality and risk. Each part of that sentence does work. Resources means money that actually leaves the business: salaries and contractors, outside services, infrastructure, rework. The same outcome means you have not quietly produced less. Acceptable quality and risk means the saving does not return later as write-offs, complaints or regulatory trouble.

The practical test follows from it. A cost reduction is real when you can name the line of the budget that falls, and by how much, after counting what the AI itself costs. If you cannot name the line, you are looking at something else: capacity, or avoided growth, or a hope. Those can be valuable too, and this chapter shows how to report them honestly. How to turn released time into realized capacity is a subject of its own, taken up in Productivity vs Realized Capacity later in this module. The question here is narrower: when does the bill change?

Where cost actually sits

Before asking where AI saves money, it helps to be clear about where money goes. The cost of almost any process can be read as six lines.

When people hear that AI saves cost, they tend to picture the first line and a smaller team. That line is the hardest to move. Salaries are paid in whole people, while AI usually saves time in scattered minutes, and decisions about roles take months and carry legal and human obligations. The other lines are often easier. A contract with an outside provider can be cut when its volume falls. Rework stops costing money the moment it stops happening. A failure that is predicted and prevented never reaches the downtime line at all.

The mix differs in every business, but the discipline is the same: name the line the saving lands on before you name the number.

Three mechanisms

AI changes those lines through three mechanisms, and keeping them apart makes a cost case much easier to judge.

Automation means AI performs a task that used to need a person: classifying documents, capturing invoice data, routing requests. The strongest cases have high volume, repetitive work and a cost that can be counted per transaction. Automation is also the mechanism most often oversold, because a task that AI can do is not necessarily one it is worth paying AI to do. MIT researchers who modeled the full cost of building and running computer-vision systems found that, at the costs of the time, only 23 percent of the wages paid for exposed vision tasks would have been attractive to automate³. Costs are falling, so that share will rise, but the principle holds: automation has to beat the full cost of the work it replaces, not just its salary line.

Augmentation means people stay in charge and AI helps them work faster or better. In a controlled experiment with 453 professionals doing realistic writing tasks, ChatGPT cut the average time taken by 40 percent and raised the graded quality by 18 percent⁴. Notice what that shows and what it does not. It shows that the same work can be done in 60 percent of the time, on set tasks. It does not show that anyone’s cost fell. Augmentation usually creates capacity first, and the arithmetic of speed needs care: “30 percent faster” frees about 23 percent of the time, not 30, because the work now takes 1 divided by 1.3 of the old time.

Optimization means AI improves how work and resources are organized rather than doing the work itself. A better demand forecast means less stock written off. A maintenance prediction means a repair scheduled on a quiet shift instead of a breakdown on a busy one. When it works, optimization often produces the cleanest savings, because it lands on lines that are easy to see: inventory, energy, downtime. These are capabilities until you have measured them: the saving has to show against the stock, energy or downtime your own operation had before.

Errors are a cost line

The fourth link in the chain deserves its own section, because it is where large AI savings can hide. Quality professionals have long divided the cost of quality into four categories: prevention, appraisal, internal failure and external failure⁵. Read through that lens, much of what AI can do is to move spending from the expensive end of the list to the cheap end.

An error found by a customer costs the most: the refund, the repeat contact, the write-off, sometimes the claim. An error fixed internally costs less. An error prevented costs least. An AI check that catches a missing field before a case is submitted can save more downstream work than it costs to run, and that value shows up on budget lines that are already measured: rework hours, credit notes, complaints.

The lens also works in reverse. AI can add failure cost. A system that drafts fluently and errs confidently can push mistakes further down the line, where they are dearer to fix. That is why a credible AI cost case keeps an appraisal line, such as human review or sampling, and measures failure costs before and after rather than assuming they fall.

When a saving reaches the budget

Now back to the hours. A useful way to think about released time is to picture an office floor that empties because teams have changed how they work.

An empty floor does not lower the rent. The landlord sends the same bill. You can hand back the lease, and the rent falls: a direct cost reduction. You can fill the floor with work that earns, and output rises while the rent stays the same: capacity value. Or you can leave it empty and describe it as a saving, in which case nothing has changed except the story. There is a fourth move. If you had planned to lease another floor next year and no longer need it, that is avoided cost: real value that never shows as a lower bill.

Released hours behave the same way. The metaphor also shows where they differ from a floor. A floor is one block of space that can be handed back in one piece. Hours released by AI are usually scattered: ten minutes here, twenty there, across dozens of people. Scattered time rarely becomes a removed cost. It has to be gathered into usable capacity first, which is why the decision about released time needs an owner and a plan.

Unit cost, net of what AI costs

The cleanest way to keep a cost case honest is to stop counting hours and count unit cost: the total cost of a process divided by the number of cases, documents or transactions it handles in the same period. Cloud finance teams use the same idea, called unit economics, to judge whether rising spend is buying proportionately more business⁶. If a document review costs 20 per document before AI and 10 after, with every cost counted, the saving is real whatever the hours say.

“Every cost counted” is where many cases go wrong. The visible price of AI is a small part of what the budget pays.

The full picture of AI cost is the subject of The Economics of AI and, in depth, Understanding AI Total Cost of Ownership. For a cost case, two habits are enough. First, put the AI’s running costs into the unit cost, including the human review you keep for quality, and show one-off setup costs separately. Second, test the unit cost at next year’s volume, not the pilot’s. In a traditional process, cost tends to grow with headcount as volume grows. AI can flatten that curve, but usage fees grow with volume too. The prize is often a flatter cost curve rather than a lower bill today.

Cheaper is not the same as better

A dangerous AI cost strategy fits in three words: make it cheaper. The right goal is the required outcome at a lower total cost, with quality and risk held where they need to be.

A widely reported example comes from the payments company Klarna. In February 2024 it announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its service chats, doing the work of 700 full-time agents. It estimated a 40 million dollar profit improvement for the year and said customer satisfaction was on par with human agents⁷. Fifteen months later, its chief executive told Bloomberg that cost had weighed too heavily in how the service was organized, and that “what you end up having is lower quality”. The company began recruiting people again so that customers could always reach a human⁸.

The lesson is not that automation was the wrong choice; the company kept its assistant and added people back. The lesson is that a case judged mainly on cost will not show what is being traded away until customers show it for you. How AI changes what customers experience is the subject of the next chapter. For a cost case, the rule is simple: every saving carries a quality measure and a risk measure beside it, owned by someone who is not paid on the saving.

Reduction and avoidance are different lines

Not all cost value is a cut in today’s spend. Some of it is spend that never happens. Consider two illustrative cases. In the first, a function costs 100 million this year and 90 million next year: a cost reduction of 10 million that anyone can see in the accounts. In the second, the forecast says growth will take the function from 100 million to 120 million, and with AI it reaches 105 million.

The second case is worth 15 million of avoided cost, and it may well be the bigger prize. But spend still rose by 5 million. A finance team that was promised a saving will see a higher number and stop trusting the next AI case it receives. Avoidance also rests on a forecast, and forecasts can be inflated. Report reduction and avoidance on separate lines, write down the baseline and the growth assumption behind every avoidance figure, and check them against actual volumes each quarter.

Story: a law firm’s review bill

The following case is a composite of a mid-sized US law firm, with illustrative numbers. It shows the discipline of this chapter applied to the industry from the opening.

Before. The firm’s corporate group handles due diligence for repeat clients on fixed fees, so the cost of reviewing documents falls on the firm, not the client. Every deal meant a first-pass review of thousands of contracts. The firm sent about 60,000 documents a year to an outside review provider at 20 per document: 1.2 million a year, on the external-services line.

After. An AI system now does the first pass, extracts the key clauses and sorts documents by risk. The firm’s own paralegals check every flagged document and a sample of the rest. When the partners did the arithmetic, they counted everything: 180,000 a year for the platform and usage, 360,000 of paralegal checking time, costed at full rate because it came out of other work, and 60,000 for quality sampling.

Running cost fell to 600,000, or 10 per document instead of 20. That is a gross saving of 600,000 a year. Setup and training cost 200,000 in the first year, so the first-year net saving was 400,000. In the second month, sampling caught the system missing one type of assignment clause across a batch of leases. The fix took a week, and it is why the sampling line stays in the budget.

The partners could defend the saving because they counted every cost, put it on the line where it landed, and checked quality beside it.

What this means for leaders

The chapter comes down to four habits. Ask which budget line falls, and refuse a saving that cannot name one. Ask for unit cost before and after, with the AI’s own running costs inside it and setup shown separately. Keep a quality measure and a risk measure beside every saving, owned by someone other than the person who claims it. And report reduction, capacity and avoidance on separate lines, so that finance can trust each of them.

Check yourself

1. If AI halves the time a salaried employee spends on a task, the task’s cost has halved.MythFact
2. The fastest AI savings often land on budget lines other than payroll.MythFact
3. Any task AI can technically do is worth automating.MythFact
4. Preventing errors can be worth more than the AI costs to run.MythFact
5. Spend that rises less than forecast can be reported as a saving.MythFact
6. A credible cost case includes human review in the AI’s running cost.MythFact

Reflection: find the line

What comes next

Revenue and cost are the two sides of the financial ledger, and this chapter and the last have shown how AI can move each. But value is not only financial, and the Klarna case showed how a cost case that ignores customers can give its savings back. The next chapter, AI and Customer Experience, asks how AI changes what customers experience, and how to tell an improvement from a cheaper way of disappointing them.