Appen, which helps Amazon and Google train AI, is reeling

Mike Monegan saw the writing on the wall in January. He had trouble sleeping for weeks.

As vice president of product management at Australian artificial intelligence software vendor Appen, Monegan and several of his colleagues were doing their best to keep things afloat as tech behemoths slashed their spending on the company’s AI training data.

Five customers – Microsoft, Apple, Meta, Google and Amazon – accounted for 80% of Appen’s revenue, and this was considered the company’s moment to shine. Across the industry, companies are investing heavily in generative AI, ensuring they don’t get left behind in the sudden race to embed the latest big language model into all their projects.

Appen has a platform of nearly 1 million freelance workers in over 170 countries. In the past, networks of people have been used to train some of the world’s leading AI systems, working for a star-studded list of tech companies, including top customer names as well as Adobe, Salesforce and Nvidia.

But just as AI’s big moment was coming, Appen was losing business — and fast. Revenue fell 13% in 2022, a decline the company attributed in part to “challenging external operating and macro conditions”. Former employees, who asked not to be named for fear of retribution, told CNBC that the company’s current struggles to shift to generative AI reflect years of lax quality controls and a disjointed organizational structure.

In mid-December, Appen announced a change at the top. Armughan Ahmed, a 25-year veteran of the tech industry, will take over as CEO, replacing Mark Bryan, who led the company for the past seven years. Launching next month, Ahmed called generative AI “one of the most exciting advances” in the industry and noted that he was “excited to know that our team has already put the technology to work on our marketing content.”

Monegan wasn’t buying it. He told CNBC that after his first meeting with Ahmed, he began looking for another job. Monegan was watching Appen fall back and didn’t see Ahmed, whose LinkedIn profile says he’s in Seattle, presenting a realistic route.

Monegan left in March to help start his own company.

The numbers seem to prove him right.

Despite Appen’s enviable client list and its nearly 30-year history, the company’s struggles have intensified this year. Revenues fell 24% to $138.9 million in the first half of 2023, citing a “broad tech slowdown.” The company said its underlying loss widened to $34.2 million from $3.8 million a year earlier.

“Our data and services power the world’s leading AI models,” Ahmed said on last week’s earnings call. “However, our results are not satisfactory. They reflect ongoing global macroeconomic pressures and a continued slowdown in technology spending, particularly among our largest customers.”

In August 2020, Appen’s shares hit AU$42.44 ($27.08) on the Australian Securities Exchange, equating its market cap to $4.3 billion. Now, the stock is trading at around AU$1.52 for a market cap of around $150 million.

With troubled financials, the company is facing a string of executive departures. Helen Johnson, who was appointed finance chief in May, left after just seven weeks in the role. Marketing chief Fab Dolan, whose departure was announced on the earnings call, spent just two months in the position. The departure of Chief Product Officer Sujata Sagiraju was also announced recently.

“In an environment of change, we expect change,” an Appen representative told CNBC.

Elena Sagunova, global human resources director, in April, followed by Jane Cole, senior vice president of enterprise, in July and Jukka Korpi, senior manager of business development for the Europe, Middle East and Africa region, in August.

Still, Ahmed said on the earnings call that the company remains “laser-focused on resetting the business” as it provides data for generative AI models. He added that “benefits from our turnover have not yet materialized” and that “revenue growth is not compensating for the decline we are experiencing in the rest of the business.”

Appen’s previous work for tech companies has included projects such as evaluating the relevance of search results, helping AI assistants understand requests in different accents, classifying e-commerce images using AI and mapping the location of electric vehicle charging stations, according to public information. and interviews conducted by CNBC.

Appen has focused on search relevance for Adobe and translation services for Microsoft, as well as work providing training data for leading companies, security applications and automotive manufacturers.

Based on the data the customer needs, an Appen freelancer can label or categorize images or search results from a laptop, or use Appen’s mobile application to capture vehicle glass breaking or background noise.

During Appen’s growth years, manual collection of data was critical to the state of AI at the time. But today’s LLM has changed the game. The underlying models behind OpenAI’s ChatGPT and Google’s Bard explore the digital universe to provide sophisticated answers and advanced images in response to simple text questions.

To boost their LLMs, which are powered by state-of-the-art processors from Nvidia, companies are spending less on Appen and more on competing services that already specialize in generative AI.

Ahmed told CNBC in a statement that while the economy is taking its toll on the company and cutting back on higher consumer spending, “I believe our disciplined focus and the early progress we are making to turn the business around will enable us to capture value from the growing generative AI market.” And return Appen to growth.”

Ahmed said on the earnings call that consumers are interested in certain types of data that are more difficult to obtain. For Appen, this means finding experts in specific types of information that can power generative AI systems. This means it needs to expand its workforce while at the same time finding ways to conserve cash.

Appen had $55 million in cash as of June 30, thanks to proceeds from a $38 million equity raise. Before the new infusion, cash had been declining, from $48 million at the end of 2021 to $23.4 million a year later.

Even before the generative AI transition, pay was a sticking point for Appen’s data labellers. In 2019, Google said its contractors would have to pay their workers $15 an hour. Appen did not meet that requirement, according to public letters written by some workers.

In January, after months of organizing, raises went into effect for Appen freelancers working on the Bard chatbot and other Google products. Rates rose to $14 and $14.50 an hour.

That was not the end of the story. In May, Appen was accused of squeezing freelancers focused on generative AI, allocating strict time limits to time-consuming tasks such as evaluating a complex answer for accuracy. A worker, Ed Stackhouse, wrote a letter Two senators expressed concern about the dangers of such limited working conditions.

“The fact that raters are exploited leads to a defective and ultimately more dangerous product,” he wrote. “Raters are not given time to deliver and test perfect AI models under the Average Estimate Time (AET) model, which they are paid for,” a practice that “prompts raters to check only a handful of facts before submitting work,” he added.

In June, Appen faced charges from the US National Labor Relations Board after it fired six freelancers who allegedly spoke out about their frustrations with workplace conditions. The workers were later rehired.

Appen employees who have spoken to CNBC on behalf of the company in recent months said the rapidly changing AI environment poses challenges. Eric Vogt, vice president of solutions at Appen, told CNBC in May that the sector is in a state of flux.

“There’s a lot of uncertainty, a lot of tentativeness to experiment, and new startups trying new things,” Vogt said. “How to make new use cases a reality usually means acquiring unusual data – sometimes astronomical volumes of data, or very rare resource types. It requires expertise with a wide range of different capabilities.”

For recent projects, Vogt said Appen needed help from doctors, lawyers and people with experience using the project tracking software Jira.

“People who you might not think of as gig workers, we got to engage with these experts for these expert systems in a way that there wasn’t a huge demand before,” Vogt said.

Kim Stagg, Appen’s vice president of product, said the work required for generative AI services was different from what the company had previously required.

“A lot of the work we’ve done has been around the relevance of big engine discoveries — a lot of them are more, ‘Is this a hot dog,’ than ‘Is this a good discovery,'” Stagg said. “With generative AI, we see a different demand.”

One focus Stag highlighted was the need to find “what we would call really good quality creative people” or those who are particularly good at language. “And there are other domain experts: sports, hobbies, medical.”

However, former employees expressed deep doubts about Appen’s ability to succeed due to its chaotic state and executive reshuffle. Part of the problem, he says, is organizational structure.

Appen was divided into a global business unit and an enterprise business unit, which at one time consisted of about five clients and over 250 clients, respectively. Each had separate teams and limited communication between them, which led to internal inefficiencies, former employees said. A former manager said it felt like two separate companies. Appen said last quarter the company consolidated its global and enterprise business units.

The company’s declining share price suggests that investors don’t see the company’s business offerings transferring into the generative AI space.

Lisa Braden-Harder, who served as Appen’s CEO until 2015, echoed that sentiment, telling CNBC that “data-labeling is completely different” than how data collection works in the ChatGPT world.

“It’s not clear to me that their previous experience with data labeling is now a competitive advantage,” she said.

Former Appen employees say the company has faced quality control issues in recent years, which have hampered its ability to provide valuable training data for AI models. For example, a former department manager said people would annotate rows of data using automated tools instead of the manual data labeling required for accuracy, which clients thought they were buying.

Consumer expectations for “clean data sets” are often not met, the person said, leaving them open to competitors such as Labelbox and Scale AI. When the manager started at the company, the enterprise business unit had over 250 clients. In 18 months Dr, That number was reduced to less than 100.

Appen told CNBC that he “won 89 new clients” in the first half of the year.

Monegan recalled that many customer relationships were “hanging by a thread”.

After the earnings report, Canaccord Genuity analysts more than halved their price target on Appen to AU$1.56. One concern cited by analysts is the 34% cut in spending by Appen’s top customer, a number that Appen would not confirm or deny.

A more existential problem, analysts note, revolves around Appen’s efforts to win business and plans to cut costs by 31% in fiscal 2023.

“That looks like a brutal level of cost cutting,” he wrote, as the company tries to stabilize “the core revenue base while growing the business around generative AI.”

See: AI is the next trend when it comes to smartphones, says a market research firm