Sold Separately

The details matter. Microsoft’s package changes include storage and security features as well as AI; Business Premium’s list price stayed flat. And Copilot Chat was already available without an extra charge, alongside a more extensive, paid Microsoft 365 Copilot offering. The price increase alone tells you neither how much you’re paying for AI nor which Copilot capabilities you have.

Adobe made a related choice in 2023, adding Firefly and generative credits to Creative Cloud while raising prices on selected plans. In either case, AI can enter your software budget through a renewal rather than a deliberate purchase. That gives you access before you’ve made a case for using it.

The trap is letting the higher bill make that case for you. If you’re keeping the subscription because you still need its other features, the money is committed whether anyone uses the assistant or not. Staff time remains yours to commit. Spending more of it to justify a bill you’ve already accepted can turn one unwanted expense into two.

Suppose your team writes customer quotes, then waits for a manager to approve the price. An assistant might shorten the drafting. If every quote still waits overnight for approval, customers get no earlier answer. If the draft also introduces an unauthorized discount, the approver gains a correction to make. A larger engine in a van parked outside a locked loading bay buys a louder idle.

Less writing could still be a worthwhile saving. But then the claim is less staff effort, not faster customer service. Those are different outcomes, and the distinction tells you what to measure. If approval is the delay you want to remove, letting staff approve routine quotes within agreed prices and terms may help more than changing how the sentences get written. Discounts and unusual commitments can still go to the manager.

Assistance can produce a measurable result. In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a staggered rollout that took place mainly in 2020 and 2021. Their data covered 5,172 customer-support agents at one enterprise software company. The assistant suggested replies during customer chats and supplied links to relevant internal documentation. Employees could edit or ignore its suggestions and remained responsible for the conversation.

The researchers estimated a 15 percent increase in issues resolved per hour, on average. Less experienced and lower-skilled workers improved in both speed and quality; the most experienced and highest-skilled workers saw small speed gains and small quality declines. The system had been trained on support conversations for that setting. Its results give you no forecast for a bundled Copilot seat, but they do show that a useful outcome can come from helping someone do an existing job.

The employees’ final say didn’t prevent that quality decline. Human review belongs inside the comparison, with its effort and errors counted. Our reading is that the useful lesson to borrow is the completed job: the company could count resolved issues rather than generated replies. You can use that distinction without buying the same system or rebuilding an entire department.

For the quoting example, compare the current process with an AI-assisted version on comparable requests, using the people who normally do the work and data your business permits the tool to handle. Count the effort through to an approved, accurate quote, including gathering source information, checking prices and correcting mistakes. Record the drafter’s and approver’s time separately so a saving for one can’t conceal extra work for the other. Track the customer’s wait separately too. A faster first draft earns only the saving that survives those later steps.

An assistant included in a subscription you’d keep anyway doesn’t have to repay the whole renewal. A modest saving can justify using it. But using it still has to earn back the setup and training effort, along with any extra license or usage charges.

For repeated work, subtract the AI-assisted process’s recurring staff hours per completed quote from the current process’s, including the upkeep of prices and instructions in both. If the difference is positive, divide setup and training hours by hours saved per quote and round up. The result is the number of quotes needed to recover the initial effort, assuming the saving holds. This measures time payback; extra charges and the value of each person’s time still need their own judgment.

The person who owns quoting can run that comparison with whoever manages your software and data access. If they can judge the errors and configure the tool safely, they can make the call internally. A reusable template may be enough when quotes follow a standard form; its upkeep and checking are a fair comparison for the assistant’s. If checking consumes the drafting benefit, keeping the existing process is a valid result.

Ask the person drafting the quotes what work they stopped doing. Then ask the person approving them what work they gained.