Feature The tech industry's enthusiasm for artificial intelligence software - a conveniently amorphous term - has yet to generate much of an economic windfall.

An estimated $1 trillion of capital expenditure commitments in the years ahead to develop and deploy AI "has little to show for it so far," according to a June report [PDF] from investment firm Goldman Sachs.

The Economist earlier this month said AI technology so far has had "almost no economic impact."

And The Information recently reported that OpenAI could lose $5 billion this year on model training and staff, potentially putting it at risk of running out of funds within twelve months unless it can raise more.

Could the party be over before it has really begun? Is anyone getting a return on their AI investment?

It's only been since November 30, 2022, when OpenAI's ChatGPT was released, that the funding frenzy, which peaked in 2021, and associated frothy predictions set the expectation of a bonanza.

Shortly thereafter, Amazon, Google, and Microsoft made a show of investing huge sums (or adding to existing bets) in orgs such as Anthropic, OpenAI, and Hugging Face. And AI was everywhere, at least as measured by web queries.

Despite the hype, the impact of AI is less evident among businesses. In March, the US Census Bureau released data based on a sample taken from 164,500 companies about how AI is affecting the American economy.

"Recent reports suggest ChatGPT has about 180.5 million users as of March 2024," the report [PDF] says. "However, it remains unclear how many businesses are currently using AI."

Not many, it seems. As of February, just 5.4 percent of US businesses reported using AI. That's up from 3.7 percent in September last year, with the expectation that figure will reach 6.6 percent this autumn.

In some industries, AI is more popular than others. Usage has reached 18 percent of firms in the information sector, while only 1.4 percent of companies are doing so in the construction and agriculture sector.

The Census Bureau report says that businesses are using AI for marketing automation, virtual agents and chatbots, natural language processing, and data/text analytics.

Speed

Clearview Consulting Group is doing just that with AnswerRocket's GenAI Analytics Platform.

Clearview, based in Atlanta, Georgia, works with retailers to analyze their data and provide product and brand recommendations. It provides guidance on what products should be stocked and ordered, what products competitors order, and data on brand popularity by region.

Allen Welch, managing director, told The Register his firm tried out AI a few years ago but wasn't ready for it. Then about a year ago, the firm connected with AnswerRock and developed the platform they're now launching in the retail space.

"It's applicable for retailers," Welch explained, "it's also applicable for manufacturers, but ultimately, we're using it ourselves.

"We do a lot of testing on products. Think Consumer Reports, but it's not published. That information is used by our clients, usually the retailers, to make a more informed buying decision about what products to put on their shelves and what mix of products they should have, as far as opening, mid-, high-price point products, private brand versus national brand, and so forth."

These would literally take us days sometimes to build, depending on the size of the category

That requires analyzing lots of data about what folks want, need, are asking for, and are complaining about. Much of this data is available online in the form of user comments and product ratings.

AnswerRocket's analytics software has allowed Clearview to go through these comments and understand them in what Welch estimates is five percent of the time of a manual review.

In addition, Clearview develops a competitive analysis for retailers that covers what's on the shelf and what's planned, and what rivals are doing.

"These would literally take us days sometimes to build, depending on the size of the category, and just how many products are on the shelf," said Welch. "With our new platform that we launched, we can actually do that ourselves in a matter of minutes."

"There's an immediate windfall right there just in your efficiency … just being able to turn through this data so much faster than others are," said Welch.

Asked to quantify the productivity gain, Welch estimated 85 to 90 percent.

Welch said the use of AI not only has made Clearview's data more complete, it has opened up capabilities that have helped its customers.

"Our customers can say, 'I want to see gloves. Show me only gardening gloves made out of such and such material that are waterproof.' And [the AI software] will immediately filter thousands of records of gloves down to the handful that meet that criteria. So [not only is it] making [our process] faster, but it's actually making it much more simple for the end user."

Asked what feedback Clearview had received from customers, Welch said, "My favorite one - and I wrote it down and stuck it on my white board at the office - it was, let's just say, one of the largest retailers in the world. [The customer] said, ''This is industry-wide, game-changing. And of course, we want to be part of it.' So they were one of our early adopters."

Not bad for an AI project that took about a year from conception to deployment, at a cost Welch says was around $750,000 to $800,000 in total.

Frances Karamouzis, distinguished VP analyst at IT consultancy Gartner, told The Register that there's a huge spectrum of AI initiatives and calculating ROI can be complicated.

One reason for that is that AI implementations vary widely. They may be front office applications (marketing, customer service), back office applications (admin, HR, legal, IT), core functions (research, supply chain, logistics), or related to products and service (product development).

We're actually recommending to CFOs not to bother calculating the ROI

With something like Microsoft Copilot, AI assistance for the Windows giant's productivity apps, calculating ROI is a bit like trying to measure the benefit of Word or Excel, Karamouzis said. It could help accelerate some tasks more than others, and on some days and not others.

"The benefits might not always be in the form of a pure profitability or pure financial number," she said in reference to ubiquitous tools like Copilot. "We're actually recommending to CFOs not to bother calculating the ROI, because you're not really going to be able to demonstrate that on your financials."

And for other sorts of AI initiatives, there's a lot of variation.

"There are some that are achieving the ROI and some where you can't measure the ROI," she explained. "And then there are some, of course, that are not, for a myriad of reasons of how often clients are not truly understanding how to estimate or manage the costs and things like that."

Karamouzis said, "We think cost is one of the greatest near-term threats [to AI adoption] because people are really not taking into account comprehensive lists of all the costs. And they're not calculating them out in a multi-year period."

According to Gartner, more than half of organizations abandon AI projects due to underestimation and miscalculation of costs.

Budgets

Inability to calculate ROI is just one of the troublespots, Karamouzis explained. Another is that AI can potentially be very expensive. Weekly, bi-weekly, or monthly budget reviews won't do.

"You actually have to look at this daily," she explained, "simply because we've seen projects where these inference costs or compute costs can go awry by 500 percent to 1,000 percent in a single day, depending on how people are [designing the system architecture]."

Karamouzis argues that traditional IT practices, where the C-suite got involved at the beginning (designing the strategy) and at the end (measuring the ROI), no longer apply. "This is the fundamental problem with AI," she explained. "You actually have to be dynamic, iterative, and transparent throughout the whole thing. And you need both C-level people and operational people to make this work ... The C-suite has to make choices to fail fast and fail cheaply."

It actually can do a lot of the things that are promised

Adding AI to a company thus is not just a subscription fee or an API integration. It's an organizational effort, one that may not be easy to measure. And without the buy-in and involvement of leadership, the result can be unrealistic mandates to boost productivity or revenue that just end up alienating employees.

For Clearview, which has used generative AI to turn its traditional consulting business into a software-as-a-service offering that supports direct client access, some operational reconfiguration was necessary.

"We really had to restructure because we were changing our service offering," Welch said. "AI itself didn't require that we change. I mean, we certainly had put in some tactical and procedural changes in place, you know, but we really restructured a whole new organization within our organization to deal with this software as a service."

But given the necessary groundwork, Karamouzis sees promise.

"The technology is not vaporware," she said. "It actually can do a lot of the things that are promised. And people are only scratching the surface of the capability. It's just going to take a while to go through the learning curve."

"....So over the long term, as a journey, we think pretty positively that this is going to really change lots and lots of industries and change a lot of business functions and even change, obviously, the way we work." ®

https://www.theregister.com//2024/07/30/ai_has_yet_to_pay/