EssayAI Adoption and Business Decisions
You May Have Been Doing the Math on AI Wrong from the Start
AI’s value cannot be reduced to headcount savings. A useful business case starts with the bottleneck and accounts for capacity, safeguards, and the cost of errors.
Original Chinese version
On this page
A business leader I know told me his company was preparing to deploy AI for customer support.
I asked what problem the business was trying to solve and how they planned to use AI. He said they had not worked through the details yet, but the initial goal was clear: reduce the team from twenty people to five.
Had they tested the system? Not yet. It had not even launched.
My first reaction was not to object to saving money. It was that the calculation was being done in the wrong order. The system had not proved what problem it could solve, yet headcount had already been written into the answer.
If you set the staffing-reduction target first and only then look for work AI can do, your deployment strategy, team behavior, and acceptance criteria will all start drifting in the wrong direction.
Anyone who has run a business can understand the impulse. Headcount is visible. Payroll is easy to calculate. It is also easy to explain in a meeting. But the easier something is to count, the easier it is to mistake it for the whole value.
The moment you begin the AI calculation with “How many people can we cut?”, you have already made the business case too narrow.
Headcount Is the Easiest Number to Calculate—and the Easiest Way to Think Too Narrowly
Suppose an AI customer-support system goes live and handles 90 percent of routine inquiries. Where should the time and energy released across the original twenty-person team go? That is not a technical question. It is a management choice.
There are at least three options on the table.
The first is to use AI to absorb growth. Suppose demand has increased by 40 percent this year. Under the old staffing model, the company would have needed to hire eight more people. Now it does not. AI absorbs the additional volume. The value does not appear as jobs cut this quarter. It appears as future hiring the company no longer needs—and a real reduction in future spending.
The second is to move people into more valuable work. Perhaps all twenty employees were previously buried in FAQs. Nobody had time to study customer behavior, conduct proactive follow-ups, or build an early-warning process for high-value customers at risk of leaving. Once AI handles the baseline volume, six people can move into those areas. The team stays the same size, but the work moves up a level. The accounts may show no simple line for “money saved,” yet renewal rates, average spend per customer, and customer satisfaction can all improve.
The third is to reduce staffing moderately after validation. Suppose service-quality monitoring has run for three months, the human-escalation rate is within an acceptable range, and fallback processes for exceptional cases have proved reliable. At that point, the team might be adjusted from twenty people to ten, saving the cost of ten positions. That is an outcome, not a starting point.
These options are not mutually exclusive. If the business is still growing, use AI to absorb incremental demand first. If valuable work has gone unattended for too long, move people into it. If the business is stable, incremental demand is limited, and quality and safeguards have been thoroughly validated, then consider reducing the staffing level.
Headcount math has a legitimate place. But it is only the most visible line in the AI ledger. Put every destination for the released capacity on the table. Then use the state of the business and the evidence from validation to decide which path to take first—and when to switch.
The calculation needs several more lines because AI capacity does not behave like conventional equipment.
AI Does Not Deliver to a Fixed Specification Just Because You Bought It
Buy an automated packaging machine and you have a reasonably clear basis for calculation: thirty units per minute, a defect rate below one in a thousand, and labor savings equal to three positions a year. The return on investment is straightforward because the machine’s capacity is relatively fixed and its quality standards are easier to measure.
AI is not that kind of equipment. A large language model does not follow fixed rules step by step to produce one uniquely correct answer. With complex questions and edge cases, reliability has to be demonstrated through testing. It cannot be guaranteed by a product specification.
An AI customer-support system may answer one hundred return-and-exchange questions perfectly today, then give an inappropriate response to an unusual complaint tomorrow. You do not know when that “tomorrow” will arrive. But you cannot assume it never will.
That is why an AI business case needs more lines. Do not calculate only the human hours saved. Include the cost of evaluation and monitoring, the cost of human review and escalation, and the customer-loss risk created by incorrect responses. Once those lines are included, the equation is no longer simply “AI investment minus payroll saved.” It looks more like a trade: operating cost in exchange for elastic capacity.
And that elasticity is precisely why AI can do things fixed equipment cannot. When volume suddenly increases, capacity can expand without waiting through a hiring cycle. When demand changes, redirecting the system is faster than rebuilding a production line.
AI may reduce unit cost, and it may expand the capacity of knowledge work. It changes more than headcount. It changes the cost, quality, and output structure of an entire part of the business. Counting only headcount is incomplete. Counting only what AI creates is incomplete too.
Customer support is simply the easiest example to see. Replace it with sales, content, research and development, or a back-office function, and managers still face the same question: where should the capacity released by AI ultimately flow?
A Better Way to Calculate AI’s Value
The next time someone says, “Let’s add AI here. It will save headcount,” do not nod too quickly. Ask three questions.
First, where is the bottleneck in this part of the business? Do not begin with “What can we save?” Begin with “What is holding the work back?” In customer support, the bottleneck may not be too many employees. It may be slow response times, inconsistent service quality, or a lack of attention to high-value customers.
Second, what will you do with the capacity AI releases? The question is not simply “We saved labor—now what?” Consider every destination: absorbing growth, moving people to higher-value work, or reducing staffing moderately after validation. If you cannot answer this question, the AI business case is not finished.
Third, what would a complete calculation include? Keep three ledgers. The investment ledger records spending on models, systems, data, integration, and ongoing operations. The value ledger records cost savings, avoided hiring, released capacity, revenue growth, and quality improvement. The risk-and-safeguard ledger records human review, exception handling, incorrect responses, and the risk of losing customers. Put the three together, and you have the real AI business case.
Managers already know how to run the numbers. AI has not invented a new kind of arithmetic. It has simply added a few more lines worth calculating.
Management Advice
Of course the AI calculation should include cost reduction. Just do not begin with “How many people can we cut?” Find the bottleneck first. Then decide whether the released capacity should drive revenue growth, improve quality, or ultimately reduce staffing. Get the sequence wrong, and the numbers may look better while the business gets worse.
When a business leader asks you to report AI cost savings, do not hand over only a headcount-reduction table. Put the human hours saved, the additional demand the system can absorb, the required human safeguards, and the cost of errors on the same page. That is an AI business case someone can actually use to make a decision.
Next time, we will use management thinking to crack open another AI question.