AI Upskilling Pledges Audited: 23 Million Promised, No Job Outcomes Published

Amazon, Microsoft and Google have each made a public AI upskilling pledge, and together the three targets add up to 23 million people. Amazon says it hit its 2 million goal a year early. Microsoft publishes credential counts. Google has published no progress count toward its 1 million goal that we could find. None of the three publishes a completion rate or a job outcome for its AI pledge.

What the three AI upskilling pledges actually promised

A pledge is a public number with a date attached. It is useful because it can be checked later, and that is the only reason this audit is possible. Each of the three companies picked a different number, a different deadline and, crucially, a different unit of success.

This post is written for the person who owns delivery inside a team: the head of operations or learning who is deciding whether to point staff at a free vendor course. That reader needs to know what the headline numbers measure before borrowing them as a template for an internal programme.

Amazon AI Ready, November 2023

Amazon launched AI Ready on 20 November 2023 as "a new commitment designed to provide free AI skills training to 2 million people globally by 2025". The package bundled free AWS courses for adults, a programme for younger learners, and scholarships for a generative AI course on Udacity. The verb is "provide". The pledge counts what Amazon delivers, not what a learner finishes.

Microsoft Elevate, July 2025

Microsoft's version is larger and phrased more precisely. Brad Smith wrote on 9 July 2025 that "in the next two years, the Microsoft Elevate Academy will help 20 million people earn an in-demand AI skilling credential". Those credentials range "from foundational fluency to advanced technical training". It sits inside a commitment of more than $4 billion over five years. The unit here is a credential, which is a stricter count than a course view.

Google's AI Opportunity Fund, April 2024

Google.org announced a $75 million AI Opportunity Fund on 26 April 2024. It "will help over one million Americans learn essential AI skills" by funding workforce and education nonprofits. No deadline appears in the announcement. Google later folded the goal into a programme called AI Works for America in July 2025. In August 2025 it committed $1 billion over three years for AI training and tools in US higher education, as Reuters reported.

AI upskilling pledge against reported progress, company by company

The table below is the core of this audit. Each row quotes the company's own target wording, then the most recent progress figure we could open and verify, then what was published about completion and jobs. Where a cell says "none found", we searched the company's own newsrooms and impact pages and came back empty.

Read each row in its own unit. "Provide training", "earn a credential" and "learn skills" are not interchangeable, which is why the 23 million total in this post is a marker of scale and not a sum of like things.

Amazon's "ahead of schedule"

The claim that AI Ready was met came from Maureen Lonergan, who leads training and certification at AWS. She wrote in 2025 that Amazon "reached two million people at the end of 2024, one year ahead of schedule". The same article says social media and streaming were "a standout and preferred destination for AI training, helping us reach and train nearly one million people".

That second sentence carries the audit. Roughly half of the 2 million came through social and streaming channels, where a view and a lesson are hard to tell apart. The article names the most popular course as "Introduction to Generative AI, The Art of the Possible". It gives no completion rate and no count of certificates.

Microsoft's credential target is the only one of the three you can test against a run rate. In the year before Elevate launched, Microsoft reported 5 million people helped to earn an AI credential. Hitting 20 million within two years means averaging 10 million a year, so the programme has to double its prior annual pace. That is our arithmetic, not Microsoft's.

No global progress figure for Elevate's 20 million had appeared by late October 2026. The country releases are more specific. Microsoft's Indonesia team reported on 24 September 2026 that Elevate had "trained more than 2.4 million people across Indonesia with AI skills and issued more than 1.5 million credentials over the past two years".

How each AI training programme defines "trained"

None of the three companies publishes a definition of a trained person. That single gap decides how much any headline number is worth, because "trained" can mean anything from a two-minute video view to a proctored exam.

Microsoft's numbers make the spread visible, because it reports three counts side by side. In the year to mid 2025 it said it "reached 60 million people, trained 15 million, and helped 5 million earn an AI credential". On those figures, 1 in 4 people reached was counted as trained, and 1 in 3 trained people ended with a credential.

Seen through this funnel, Amazon's "provide training" and Google's "learn essential skills" both read closer to the top bar than the bottom one. Microsoft deserves credit for picking the strictest unit. It also has not said what a credential involves, beyond a range "from foundational fluency to advanced technical training", so a short badge and a certification exam count the same.

My position is simple. I would not count a streamed video or a social post as training for any internal purpose. If a person cannot show you a completed task that used the skill, they were reached, not trained. The definitional problem is the same one that inflates internal programmes, which the analysis of use-case counts as a vanity metric covers from the deployment side.

The older skills pledges show the same reporting pattern

AI pledges are recent, so there is limited history to judge them on. The same companies made large skills pledges before generative AI, and those have run their course. They show how the reporting tends to end.

Cloud and digital skills: targets beaten, outcomes absent

In December 2020, AWS committed to free cloud skills training for 29 million people by 2025. Amazon later reported it had "helped more than 31 million learners across 200 countries and territories build their cloud skills", more than a year early. That is 107% of target, with "learner" undefined.

Microsoft's 2020 Global Skills Initiative aimed to bring digital skills to 25 million people by the end of that year. Microsoft's European policy blog later said the goal was passed with over 30 million people reached globally. Microsoft's April 2026 Australia release adds a pattern: it said its earlier pledges of 300,000 and 1 million Australians were both "achieved ahead of schedule". It then tripled the target to 3 million by the end of 2028.

Every one of these was beaten. Every one was counted in reach or learners. I take the consistency as a finding in itself: pledges designed around reach are pledges designed to be met.

Amazon Upskilling 2025: the employee pledge

The more interesting comparison is Amazon's pledge to its own staff. In July 2019 Amazon pledged to upskill 100,000 US employees for in-demand jobs by 2025, at a cost of $700 million. It later described the pledge as more than $1.2 billion to upskill 300,000 employees.

Amazon now reports that "over 425,000 Amazon employees in the U.S. participated in a skills training program since 2019." Note the verb. The pledge said upskill for in-demand jobs. The progress report says participated. We found no Amazon statement that reports how many participants moved into a higher-paid role, which was the stated point of the pledge.

That matters more than it looks. Amazon is the one company here that employs the trainees, holds their pay records and could publish an outcome cleanly. If an employer pledge with full data access reports participation, a public pledge with no data access will not do better.

What workforce training evidence says actually changes outcomes

Labour economists have run randomised trials of job training for decades. The evidence is uneven, and it points the same way: training changes earnings when it is tied to a specific job, and rarely otherwise. This section sets out three bodies of evidence, from strongest effect to weakest.

Sectoral programmes: the strongest result

Lawrence Katz, Jonathan Roth, Richard Hendra and Kelsey Schaberg re-analysed randomised evaluations of sector-focused programmes for low-wage workers. These programmes combine upfront screening, occupational and soft-skills training, and support services, all aimed at one industry. The authors found they "generate substantial and persistent earnings gains (11 to 40 percent) following training completion."

The mechanism is the part a pledge cannot copy. The paper says the gains come from "getting participants into higher-wage jobs in higher-earning industries and occupations not just by raising employment". The training worked because it was attached to an employer and a job category. A course with no destination has nothing to attach to.

Public training funds and open online courses

The US Department of Labor's national experiment on its Workforce Investment Act programmes ran across 28 randomly selected local areas between 2011 and 2013. Mathematica found that job seekers offered individualised staff assistance earned about $7,100, or 20%, more over 30 months. Access to WIA-funded training on top of that did not raise earnings within the same 30 months.

Open online courses are closer to what the AI pledges deliver. Justin Reich and José Ruipérez-Valiente analysed MIT and Harvard courses on edX from 2012 to 2018 in Science in January 2019. Completion rates did not improve over six years, and the share of first-time learners who came back the next year fell from 38% in 2012-13 to 7% in 2016-17. Free and self-paced reaches many people. It holds very few.

Generative AI training and earnings: the early evidence

The obvious objection is that AI skills are different. Generative AI tools might raise productivity fast enough that even light training pays off in wages. The best large-scale test so far does not support that.

Anders Humlum and Emilie Vestergaard linked two large Danish surveys of chatbot adoption to administrative employer and employee records. Their 2025 NBER working paper found null effects on earnings and recorded hours, and ruled out effects larger than 2% two years after adoption. They report that employers commonly encouraged use and ran training initiatives. Average time savings were around 3%.

That result does not say AI training is useless. It says adoption plus general encouragement did not reach the payslip within two years. The authors point to modest productivity gains and weak pass-through to wages. For the self-reported side of the same question, see the breakdown of where AI time savings actually go.

Put next to the sectoral evidence, the lesson is consistent. A generic skill delivered at scale has, so far, produced small measured effects. A specific skill tied to a specific role is where the large effects have been found.

Where this audit of AI upskilling pledges is weakest

Three objections deserve a fair hearing, and the first one is strong.

The pledges were never meant to be job programmes

None of the three AI pledges promised a job. Amazon promised training provided, Microsoft promised credentials, Google promised skills learned. Holding them to an employment standard grades them on a test they did not sit. A free introductory course that makes 2 million people less anxious about a technology may be worth funding on its own terms, and nobody runs a randomised trial on curiosity.

My answer is that the pledges borrow the language of workforce outcomes. Microsoft calls its credentials "in-demand". Amazon framed AI Ready around jobs and pay when it launched. When a company uses employment as the justification, the absence of employment data is a fair thing to point out.

Two further limits apply. We could not find every progress report, and a figure published in a regional release or a sustainability report may exist that we missed. The research evidence also comes from different populations: low-wage US job seekers, edX learners and Danish workers. None of it tests a corporate AI pledge directly, because nobody has published the data that would allow it.

How to use free vendor AI training programmes for your team

The free courses are not the problem. They are a reasonable first exposure, and they cost your budget nothing. The problem is reporting their completion as if it were capability. If you want training that changes output, build around the features the evidence rewards.

The last row is the honest concession. Sectoral programmes are expensive and narrow. Free courses reach millions for almost nothing per head. The right design for your team uses the cheap layer for exposure and spends real effort on the narrow layer.

Inside a company, the sectoral logic translates directly. Pick one workflow, train the people who own it, and measure output on that workflow before and after. The design detail, including the five-hour threshold, is in the piece on AI training programmes that changed behaviour. If you are in the EU, the free courses may also help with the AI literacy duty, though the evidence an Article 4 file actually needs is more than a completion list. Rollouts that skip this step tend to show up as licences bought and seats left idle.

Frequently asked questions

Where to start this week

Ask one question of every training number in your own reporting: which bar of the funnel is it counted on? If the answer is reach or enrolment, relabel it before it goes into a board pack.

Then choose a single workflow, one team and one output measure. Record the measure now, before anyone takes a course. In 90 days you will have something none of the three pledges has published: a before and after for the people you trained.

Related reading