5 steps to check an AI investment's ROI

Welcome back.

If you're considering a larger AI budget or hiring another AI-savvy person, you need to think about two things.

First, what should the business get out of it? A faster team is a good start, but more speed doesn't automatically mean more useful work.

Second, can the rest of the workflow keep up? AI might speed up one person's part of the job while the review and follow-up stay just as slow.

So today I’ll cover how I'd work through both, using a made-up onboarding team to show the math.

IN THIS EDITION

Anthropic filed to go public on October 26. Public investors tend to push for higher margins, so I’d expect pricing and plans to change over time, though nothing has been announced yet.

Anthropic · OCT 26

GPT-6.1 Terra is 50% cheaper per token and built for routine work like reading invoices. If you test it, compare the cost of a finished, checked result, since cheaper tokens don’t help if someone spends the savings fixing mistakes.

OpenAI · OCT 26

Unilever says its agents cut the first step of order processing from 15–20 minutes to about 90 seconds across 11 pilot markets. AI handles routine orders, and people get a clear queue of those that need judgment.

Microsoft/Unilever · OCT 26

Last week, I spoke with a leader whose team had put real work into adopting AI. They ran hackathons, people helped colleagues get comfortable with it, and the team had plenty of tools to choose from. Naturally, their AI spending had gone up.

And it was paying off in one clear way. The team was noticeably faster and producing a lot more than before.

The only problem was that when finance asked what the company was getting in return for that spending, he couldn't point to anything specific. He could see the team doing more work, but he couldn't connect that extra work to tangible business results.

He explained that while his team could build twice as much, it didn’t necessarily mean they'd build twice as many good things.

Usually, teams pick the first project carefully because time is limited. Once AI frees up more capacity, the extra project often gets much less thought and scrutiny, so it's less likely to produce anything the business would notice.

That is why you need a clear view of what you expect that investment to improve before considering more AI budget or headcount.

Here’s how I’d work through it:

Before thinking about tools or hires, pick a real decision you're facing right now. Maybe you're deciding whether to bring on another customer success person, or you're about to renew an expensive AI contract and want to know if it's worth it.

Then write down the specific result the business needs. Skip phrases like "boost productivity" and name the work itself.

For example, "help the onboarding team be more productive" is too vague to measure. "Next quarter, we need to onboard 30 customers a month, up from 20, while keeping the same standard of service" gives you a number someone in finance could check.

Once the goal is that specific, you have something concrete to work backward from. You can start figuring out how much work it really takes to onboard those extra customers, which parts AI could realistically help with, and whether another hire is still the right answer.

Once you know the result you need, look at all the work behind it. Onboarding a customer isn't only the setup call. Someone has to prepare beforehand, review the work, answer follow-up questions, and check in once the customer is live.

Add up the time each step takes per customer. In our example, say it comes to 6 hours per customer, and the team has 140 hours a month for onboarding. At 20 customers a month, that's 120 hours, which fits.

At 30 customers, it's 180 hours, which doesn't fit.

If AI only speeds up the setup call, the review and follow-up still take just as long, so the team still can't handle 30 customers.

Now go through those steps and mark where AI could realistically save time. Be painfully honest about what it can't do.

AI might draft the prep notes and answer common follow-up questions, but someone still has to review the work and talk to the customer.

Then estimate the new time per customer. In our example, say AI brings onboarding down from 6 hours to 4.

Don't forget the work AI adds. Someone has to:

Say that takes 10 hours a month.

If AI really does cut onboarding to 4 hours, the current team can take on 30 customers a month without another hire. That gives you a clear answer to take to finance, because AI spending is tied to onboarding 10 more customers each month.

The estimate has to hold up, though.

If AI only brings onboarding down to 5 hours, the team needs 160 hours, and you're back to needing more people.

That small difference changes the whole decision, so the next step is to test it.

Before you approve the budget or post the job, check whether the 4-hour estimate holds up.

Run the AI-assisted process with a small group of new customers and track how long each onboarding actually takes, including the review and follow-up.

Keep an eye on quality too. If onboarding gets faster but customers come back with more questions or take longer to get going, you've just pushed the time cost later in the process.

If the test shows 4 hours, you can make the case for the AI spending with real numbers. If it shows closer to 5 or 6, you know the team needs another person, and you found out before spending on the wrong thing.

You can use the same process for any AI decision, not just onboarding. More budget or another hire won't multiply your team's output on its own. But once you know the result you need and you've tested whether AI gets you there, you can show finance exactly what the spending will get the company.

I'd hold off on the new tool for now. A context layer pulls from what you already have, so if old pages still mention the discount, it'll just find them faster.

I’d start with pricing. Make one page the official source, assign an owner, and archive the old pages. Then ask ChatGPT the questions your reps ask. If it gets them right, you’re saving yourself a hefty software purchase.

If the right answer is still split between your CRM and Notion, that's when a context layer earns its keep. Someone still has to decide what's official, though.

I’ve spent over a decade in AI, helped 300+ teams put it to work, and taught 70,000+ people through AI For Anyone. Hit reply and ask me anything - each week I’ll pick one question and answer it here.

That’s it for this week’s issue of Not A Bot (formerly known as AI Ready, formerly known as Not A Bot). Thanks for reading.

Want to try this on your own team? Reply "worksheet," and I'll send you the one-page version of these 5 steps to fill in before your next AI decision.

Until next time,
Haroon