This article originally appeared on HBR.org (subscription may be required).

Whether they are shopping for spectacles or a sofa, consumers have no shortage of augmented reality apps to assist them these days. Following the viral success of Pokémon Go in 2016, hordes of retailers have embraced the technology.

Nintendo’s cellphone-enabled treasure hunt was the first big showcase for AR’s innovative blend of real-world and computer-generated images. Now the retail sector is using AR to sell items as diverse as furniture, cosmetics, home improvement products, and fashion.

Executives like the way that AR could help make online shopping feel as good as—or better than—shopping in person. Sephora’s app does this brilliantly, for instance, by allowing consumers to use their phone cameras to “try on” makeup virtually. Retail leaders also reckon AR will make their physical stores more engaging—and their salespeople more productive.

But AR’s promise doesn’t make it a great investment for every retailer. Executives must make painful trade-offs among myriad investment options, and they have reason to be suspicious of tempting gadgets. Bets on the likes of 3-D TVs, scan-and-go checkouts, and facial-recognition software have tested their patience. Such innovations may not prove worthless, but they are certainly worth less than tech dreamers imagined.

The recent collapse and relaunch of Blippar, a prominent European AR start-up that created AR apps for consumer goods and retail customers such as Covent Garden, Net-a-Porter, and McDonald’s have added to worries that, for all its promise, AR might struggle to reach the mainstream. Forrester recently reported that new venture capital funding for AR in 2018 was $1.69 billion, less than half the $3.58 billion raised in 2017. “We believe such a dramatic pullback is in direct response to expensive and underwhelming results from early adopters of XR,” Forrester said, advising companies considering AR to “proceed with an abundance of caution.”

So, how should retailers determine the appropriate role for AR in their businesses? By answering four questions that could apply to almost any technology decision.

Will our customers value this (more than a price cut)? Customers have a hard time telling us what they will want. We haven’t met any customers who were asking for Pokémon Go. Conventional wisdom says that clothing shoppers want fashion advice, top-notch service, and experiential ambience, yet we watch Amazon and discounters such as T.J.Maxx gain market share with lower prices (and a “root through the piles” treasure hunt at the latter’s stores).

Whether retailers make or buy them, AR apps cost real money—anywhere from $300,000 to $30 million for development costs alone. Are customers willing to pay for that or would they rather have lower prices? The answer depends on whether your targets are aggressive adopters of technology, the app enhances your brand, and the purchase and usage of your product is sufficiently complex to justify the use of AR.

Furniture apps such as Ikea Place use AR to ease a notorious source of pain for shoppers—namely, the difficulty of predicting what a couch, bed, or table will look like when brought home. Will it fit into the available space? Will it go with the existing furniture, carpets, and walls? That’s a perfect problem for AR to solve. Consumers really suffer when they buy the wrong furniture online: They might lose eight weeks waiting for delivery only then to be forced into the nightmare that engulfs those trying to return these bulky items. The value to the consumer is high relative to the cost of the innovation.

Does the technology have value to a wide range of customers? When you’re thinking about AR’s value to customers, don’t stop with consumers. It turns out that AR can be useful in education and training simulations, helping field representatives perform maintenance and repairs, and testing complex store designs and tricky user experiences.

Often these uses are more profitable than consumer apps and are the right places to begin building AR capabilities. For instance, stock pickers at a Dutch e-commerce warehouse operator worked 15% faster when they were equipped with Google Glass; orders were pushed directly to the AR-enabled spectacles, accelerating a process that had relied on the retrieval of physical print-outs. Start with the most profitable applications and move on to the tougher ones.

Can the math work? Even if the math turns out to be wrong, it’s worth laying out how a technology such as AR is supposed to improve profits. Is it supposed to improve sales (the number of customers that visit each year, the frequency of shopping visits per year, the percentage of shopping visits that create purchases, the number of categories shopped, the number of items purchased per category, the average unit revenue per item)? Is it supposed to reduce costs (labor, materials, distribution, marketing)? Is it supposed to reduce inventory levels or capital expenses?

Limit the intangible benefits. Don’t just “imagine the PR power”; quantify the improvement in marketing expenses. With these kinds of estimates in hand, it is far easier to test initial assumptions, compare actual results with early estimates, improve investment proposals over time, and identify better ways of solving customer problems.

And don’t forget to look for creative sources of funding. Technology vendors often are willing to subsidize AR projects for learning and publicity purposes. Merchandise vendors may be willing to pay to have their products featured in the apps.

Where does this belong on our technology backlog? Let’s face it: The technology systems of most retailers are dreadful. As a result, they become the choking points for almost every important innovation that retailers need to succeed. Yet the number of projects heaped onto this creaking infrastructure is growing fast. This is why the biggest question of all is how you should prioritize and sequence AR on your technology to-do list.

Given constraints on budgets and the hiring of tech experts, the results are often disastrous when retail executives add projects to this to-do list. Delays ripple across the existing backlog. Meanwhile, customer needs evolve, and nimbler competitors charge ahead, making many of these stalled projects obsolete.

Retail executives can’t see technology projects piling up the way they can see inventory stacked in warehouses or backrooms. But technology projects are every bit as expensive and perishable as physical inventories. Retailers need to stop starting innovation projects and start finishing them.

Agile methods often sequence technology projects according to their cost of delay. In other words, what would it cost to delay this project by one month? These costs can be a matter of life and death for core technology projects in retail such as e-commerce websites and apps, the integration of online and offline shopping, advanced analytics, and call center improvements.

And the costs of delaying an AR project? Let’s just say that we haven’t seen any retailers die because their AR project was delayed by a year. Nor have we seen languishing retailers leap to leadership based on their AR apps.

As a tool, AR likely will get more powerful. It helps that billions of people will always have an AR-enabled gadget in their pocket or handbag. But the appropriate role for it will vary significantly by retail sector and by the health of a retailer’s core technologies. So while AR should be on many retailers’ test-and-learn list, it should not delay the advancement and completion of more important tech projects that will determine whether these companies follow the path of Amazon or that of Toys “R” Us.

Darrell K. Rigby is a partner in the Boston office of Bain & Company. He heads the firm’s global innovation practice. He is the author of Winning in Turbulence. Mikey Vu is a partner in the Chicago office of Bain & Company. He is a leader in the firm’s Retail and Digital practices. Asit Goel is a partner in the San Francisco office of Bain & Company. He is a leader in the firm’s global technology practice.