
‘A million calls an hour': Israel relying on Microsoft cloud for expansive surveillance of Palestinians
Meeting at Microsoft's headquarters near Seattle, a former chicken farm turned hi-tech campus, the spymaster, Yossi Sariel, won Nadella's support for a plan that would grant Unit 8200 access to a customised and segregated area within Microsoft's Azure cloud platform.
Armed with Azure's near-limitless storage capacity, Unit 8200 began building a powerful new mass surveillance tool: a sweeping and intrusive system that collects and stores recordings of millions of mobile phone calls made each day by Palestinians in Gaza and the West Bank.
Revealed here for the first time in an investigation by the Guardian with the Israeli-Palestinian publication +972 Magazine and Hebrew-language outlet Local Call, the cloud-based system – which first became operational in 2022 – enables Unit 8200 to store a giant trove of calls daily for extended periods of time.
Microsoft claims Nadella was unaware of what kind of data Unit 8200 planned to store in Azure. But a cache of leaked Microsoft documents and interviews with 11 sources from the company and Israeli military intelligence reveals how Azure has been used by Unit 8200 to store this expansive archive of everyday Palestinian communications.
If you have something to share about this story, you can contact Harry Davies and Yuval Abraham using one of the following methods.
Secure Messaging in the Guardian app
The Guardian app has a tool to send tips about stories. Messages are end to end encrypted and concealed within the routine activity that every Guardian mobile app performs. This prevents an observer from knowing that you are communicating with us at all, let alone what is being said.
If you don't already have the Guardian app, download it (iOS/Android) and go to the menu. Select 'Secure Messaging'.
To send a message to Harry and Yuval please choose the 'UK Investigations' team.
Signal Messenger
You can message Harry using the Signal Messenger app. Use the 'find by username' option and type hfd.32
Email (not secure)
If you don't need a high level of security or confidentiality you can email harry.davies@theguardian.com
SecureDrop and other secure methods
If you can safely use the tor network without being observed or monitored you can send messages and documents to the Guardian via our SecureDrop platform.
Finally, our guide at theguardian.com/tips lists several ways to contact us securely, and discusses the pros and cons of each.
According to three Unit 8200 sources, the cloud-based storage platform has facilitated the preparation of deadly airstrikes and has shaped military operations in Gaza and the West Bank.
Thanks to the control it exerts over Palestinian telecommunications infrastructure, Israel has long intercepted phone calls in the occupied territories. But the indiscriminate new system allows intelligence officers to play back the content of cellular calls made by Palestinians, capturing the conversations of a much larger pool of ordinary civilians.
Intelligence sources with knowledge of the project said Unit 8200's leadership turned to Microsoft after concluding it did not have sufficient storage space or computing power on the military's servers to bear the weight of an entire population's phone calls.
Several intelligence officers from the unit, which is comparable to the US National Security Agency (NSA) in its surveillance capabilities, said that a mantra emerged internally that captured the project's scale and ambition: 'A million calls an hour'.
The system was built to sit on Microsoft's servers behind enhanced layers of security developed by the company's engineers with Unit 8200's instructions. The leaked Microsoft files suggest that a large proportion of the unit's sensitive data may now be sitting in the company's datacentres in the Netherlands and Ireland.
Disclosures about the role of Microsoft's Azure platform in the surveillance project come as the US tech giant faces pressure from employees and investors over its ties to Israel's military and the role its technology has played in the 22-month offensive in Gaza.
In May, an employee disrupted a keynote speech by Nadella in an act of protest, at one point yelling: 'How about you show how Israeli war crimes are powered by Azure?'
After the Guardian and others revealed in January Israel's reliance on Microsoft technology during the war in Gaza, the company commissioned an external review of the relationship. The review, Microsoft said, had 'found no evidence to date' that Azure or its AI products were 'used to target or harm people' in the territory.
A senior Microsoft source said the company had held conversations with Israeli defence officials and stipulated how its technology should be used in Gaza, insisting Microsoft systems must not be employed for the identification of targets for lethal strikes.
However, Unit 8200 sources said intelligence drawn from the enormous repositories of phone calls held in Azure had been used to research and identify bombing targets in Gaza. One of the sources said that when planning an airstrike on an individual located within densely populated areas where high numbers of civilians are present, officers would use the cloud-based system to examine calls made by people in the immediate vicinity.
The sources also said use of the system had increased during the campaign in Gaza, which has killed more than 60,000 people in the territory, the majority of whom are civilians, including over 18,000 children.
According to health officials in Gaza, at least 60,000 people have been killed during Israel's current military campaign, launched after the Hamas-led attack on 7 October 2023 which killed nearly 1,200 people.
The actual death toll is likely to be significantly higher, as the figure only includes Palestinians killed by bombs or bullets whose bodies have been recovered, leaving out thousands trapped under the rubble or killed by starvation and other indirect victims of the campaign.
According to the data – which includes the deaths of militants – women, children, and elderly people account for approximately 55% of the recorded deaths.
But the initial focus of the system was the West Bank, where an estimated 3 million Palestinians live under Israeli military occupation. Unit 8200 sources said the information stored in Azure amounted to a rich repository of intelligence about its population that some in the unit claimed had been used to blackmail people, place them in detention, or even justify their killing after the fact.
'When they need to arrest someone and there isn't a good enough reason to do so, that's where they find the excuse,' one said, referring to the information stored in the cloud.
A Microsoft spokesperson said it had 'no information' about the kind of data stored by Unit 8200 in its cloud. They said the company's 'engagement with Unit 8200 has been based on strengthening cybersecurity and protecting Israel from nation state and terrorist cyber-attacks'.
'At no time during this engagement,' they added, 'has Microsoft been aware of the surveillance of civilians or collection of their cellphone conversations using Microsoft's services, including through the external review it commissioned.'
The driving force behind the cloud project – described by one source as a 'revolution' within the unit – was Sariel, commander of Unit 8200 between early 2021 and late 2024. A career intelligence officer, Sariel was a strong advocate for projects of this scale.
Following a 2015 wave of deadly so-called 'lone wolf' attacks by young Palestinians, many of whom were teenagers unknown to the security services, Sariel had overseen a significant expansion of the volume of Palestinian communications that Unit 8200 intercepted and stored.
His answer was to begin 'tracking everyone, all the time', said an officer who worked for Sariel at the time. Instead of traditional surveillance of specific targets, Sariel's project relied on mass surveillance of Palestinians in the West Bank and used novel AI methods to extract insights.
'Suddenly the entire public was our enemy,' said another source who worked on the project, which sought to predict whether someone represented a threat to Israeli security.
One system developed in this period, sources said, scanned all text messages between Palestinians in the West Bank and assigned each message a risk rating based on an automated analysis of whether it included words deemed to be suspicious. Still in use, the system – known as 'noisy message' – can identify text messages in which people talk about weapons or discuss wanting to die.
When Sariel became Unit 8200 commander in early 2021, he prioritised forging a partnership with Microsoft that would give the unit the ability to go further and capture and analyse the content of millions of phone calls each day.
At his meeting with Nadella later that year, Sariel does not appear to have explicitly stated his plan to store Palestinian phone calls in the cloud, referring instead to 'sensitive workloads' of secret data, according to internal records of the meeting.
But documents suggest that Microsoft engineers understood the data stored in Azure would include raw intelligence, including audio files, while some Israel-based Microsoft staff, including alumni of Unit 8200, appear to have known about what the unit hoped the joint project would achieve.
'You don't have to be a genius to figure it out,' one source said. 'You tell [Microsoft] we don't have any more space on the servers, that it's audio files. It's pretty clear what it is.'
Microsoft's spokesperson said: 'We are not aware of Azure being used for the storage of such data.' They said Unit 8200 was simply a customer of its cloud services and Microsoft 'did not build or consult with Unit 8200' on a cloud-based surveillance system.
However, in early 2022, Microsoft and Unit 8200 engineers worked quickly and closely together to design and implement advanced security measures within Azure to meet the unit's standards. 'The rhythm of interaction with [the unit] is daily, top down and bottom up,' one document noted.
Among Microsoft staff, the project was shrouded in considerable secrecy and engineers were told not to mention Unit 8200 by name. Under the plan, vast troves of raw intelligence material would sit in Microsoft's datacentres overseas.
Files suggest that by July this year, 11,500 terabytes of Israeli military data – equivalent to approximately 200m hours of audio – was held in Microsoft's Azure servers in the Netherlands, while a smaller proportion was stored in Ireland. It's unclear if all of this data belongs to Unit 8200; some may belong to other Israeli military units.
According to the files, Unit 8200 informed Microsoft that it planned to move over time as much as 70% of its data, including secret and top secret data, into Azure and was willing to 'push the envelope' with the kind of sensitive and classified information that intelligence agencies normally held on their own servers. 'They're always trying to challenge the status quo,' one executive noted.
Asked about Sariel's meeting with Nadella, Microsoft's spokesperson said it 'is not accurate' to say the CEO provided his personal support for the project with Unit 8200. They said Nadella 'attended for 10 minutes at the end of the meeting' and there was 'no discussion' of the content of the data the unit planned to move into Azure.
However, according to internal Microsoft records of the meeting seen by the Guardian, Nadella offered support for Sariel's aspiration to move so much of the elite surveillance unit's data into the cloud, described earlier in the meeting as including sensitive intelligence material.
'Satya suggested that we identify certain workloads to begin with and then gradually move towards the 70% mark,' one record states. It adds that Nadella said 'building the partnership is so critical' and 'Microsoft is committed to providing resources to support.'
Several months before meeting Microsoft CEO Nadella in 2021, Sariel had published a book about artificial intelligence under a pen name – revealed by the Guardian to be the spy chief's – in which he urged militaries and intelligence agencies to 'migrate to the cloud'.
Known within Israeli intelligence as a tech evangelist, Sariel valued what he characterised to colleagues as a friendly relationship with Nadella, according to a senior intelligence source. 'Yossi bragged a lot, even to me, about his connection with Satya,' they said. (Microsoft denied that Nadella and Sariel had a close relationship.)
'He sold [the partnership] internally and got a huge budget,' another former intelligence colleague said. 'He claimed it was the solution to our problems in the Palestinian arena.'
Sariel declined to comment and referred the Guardian's questions about the project to the Israel Defence Forces. An IDF spokesperson said its work with companies such as Microsoft was based on 'legally supervised agreements'.
They added: 'The IDF operates in accordance with international law, with the aim of countering terrorism and ensuring the security of the state and its citizens.'
After publication, the IDF issued a new statement. 'We appreciate Microsoft's support to protect our cybersecurity. We confirm that Microsoft is not and has not been working with the IDF on the storage or processing of data.'
For its part, Microsoft viewed the multiyear partnership as a lucrative commercial opportunity. Executives anticipated hundreds in millions of dollars in revenue and 'an incredibly powerful brand moment' for Azure, according to the files.
'[Unit 8200's] leadership hopes to expand the mission-critical work tenfold in the coming years,' one executive noted.
As Unit 8200 began to make use of Azure's storage capabilities in 2022, intelligence officers rapidly grasped the new powers at their disposal. 'The cloud is infinite storage,' one source familiar with the system said.
Calls – which include calls made by Palestinians to international and Israeli numbers – are typically retained in the cloud for about one month, though storage can be scaled up, allowing the unit to keep hold of calls for longer periods of time when needed, several intelligence sources explained.
This allows the unit to go back in time and retrieve the phone conversations of people who become of interest, they said. Previously, surveillance targets would need to be pre-selected for their conversations to be intercepted and stored.
Several of the sources insisted the cloud-based system had prevented deadly attacks against Israelis. One said 'saving lives' of Israelis was the principal motivation behind Sariel's vision for the system. But it notably failed to prevent the Hamas-led attacks of 7 October 2023 in which nearly 1,200 people were killed in southern Israel and 240 people were kidnapped.
In the wake of the attacks, Sariel faced criticism for his apparent prioritisation of 'addictive and exciting' technology over old-fashioned intelligence methods, which some critics said contributed to the disaster. Sariel resigned last year, accepting responsibility for '8200's part in the intelligence and operational failure'.
In the ensuing war in Gaza, the cloud-based system pioneered by Sariel has been put to frequent use alongside a series of AI-driven target recommendation tools also developed on his watch and debuted by the military in a campaign that has devastated civilian life and created a profound humanitarian crisis.
Israel's destruction of Gaza's telecoms infrastructure has reduced the volume of phone calls in the territory but sources said the information held in the cloud remained useful. According to one, enthusiasm for the system had grown among intelligence officers working on Gaza as the war progressed and they saw the military is 'heading towards long-term control there'.

Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles


The Herald Scotland
3 hours ago
- The Herald Scotland
University of Edinburgh 'pushing staff to cut exams'
In the latest guidance on minimum standards for assessment in the College of Arts, Humanities and Social Sciences, the section on exams states: "Reducing the number of exams limits the considerable additional academic, administrative and estates costs associated with exams, and minimises the need for alternative assessments for student with adjustments". It's understood that behind the scenes considerably more overt pressure is being put on departments in the school, which represents more than half of the University, to move away from exams. Read More: However, staff have questioned the expenditure the university has on external venues for exams, and raised concerns about removing an aspect of courses which cannot be influenced by Artificial Intelligence (AI). A concerned staff member, who has been kept anonymous to protect their position, told The Herald: "There is regular encouragement to get rid of exams. "A lot of the subjects within our college have gotten rid of exams in the past few years, they got suspended during Covid and everything was online. 'Then, partly under encouragement from management, a lot of them just never came back and now so many have got rid of exams a lot of those still using them are being pressured to get rid of them because the university spends a lot of money renting out the EICC (Edinburgh International Conference Centre)." The conference centre is an external venue that is used to host exams for the University of Edinburgh. It's understood the University of Edinburgh receives a discounted rate to hire rooms at the EICC. The Edinburgh International Conference Centre However, the cost of doing so is likely to run well into six figures. The university hires space at the conference centre for four to five weeks in May and approximately two weeks in December to hold exams. It's understood the rate for the university to hire the Cromdale Room in 2025 stands at £9500 + VAT per day for 2025, rising to £9700 + VAT for 2026. For the Lennox, Lammermuir and Moffat rooms the combined costs are £13,000 + VAT per day, rising to £13,650 for 2026. A staff member said: "We used to have all our exams on site, because we have a lot of large spaces like gymnasiums and stuff like that. 'Before the pandemic the regular exam diet had two exams, through May and December we'd have exams usually at 10am and 2pm. 'During the pandemic people were doing exams online and they were in different time zones all over world, so we did all of our exams at 1pm UK time which wouldn't be too insane for most other people. 'When they started doing exams in person again they just kept that, they only did one exam per day so we didn't have enough space and we started renting it at the Conference Centre - which seems mad to me. 'The whole enterprise is incredibly expensive, we're renting out at least three of their large rooms and possibly more than that for four or five weeks in May, two or three weeks in December. 'So because they're so expensive they're pressuring us to use fewer exams, but they also seem very reluctant to revert to the old system of just using the gymnasiums which we have – we do use them, but not very intensely – and some of the spaces we used to use locally aren't being used at all, we're just renting the Conference Centre and spending huge amount of money on it in the middle of a budget crisis. "If you're making cuts that seems like the first thing you'd get rid of, exams themselves should be about 10,000th in your priority list." The University said the EICC "has been used in specific circumstances as part of contingency planning to minimise any possible disruption to students". While there are some subjects where exams are less useful than in others, and ones where students will naturally sit fewer of them, a move away from exams is raising concerns over the degrees being handed out. The rise of Large Language Models (LLMs) such as ChatGPT has made assessment by essay less reliable. In January 2023 a survey of 1,000 college students found that nearly 90% of them had used ChatGPT to help with homework assignments. Read More: A staff member said: "I sit in meetings with the same people who don't like exams and we're talking about AI and they'll say, 'it's crazy, it's obvious how much our students are using them, all the worst essays have disappeared'. 'You'll always have students who can barely write, can barely get words on a page, so you used to have this left hand tail of extremely low quality for all assignments and that has almost disappeared. 'The ability to rank students and assess them on their writing is so far below what it was five years ago before LLMs. "Even if we don't think exams are a great way to assess they're becoming, relatively speaking, better and better over time because it's where we know the students aren't using LLMs to completely undermine the whole process. "As universities have got bigger there's more and more bureaucracy, you have people making decisions who are not involved in teaching at all, and the people who are teaching have no way to get their voice into these decisions that are being made. 'As a result I think some people just don't realise how quickly the LLMs have taken over. "You can see it, even in the style of writing, it's so clear but part of the problem is that it's very clear, but it's also very hard to say it absolutely, definitively and fail a student for it. "The typical way of plagiarising was copying a student on the same course, which is just idiotic and you can easily recognise it because you have two assignments that are the same. The other way would be copying blocks of text from another source. 'So you could bring a student in, put them through a whole disciplinary process and give them zero. 'But when they use AI, you don't really know. A lot of students are only using it to basically clean up their essays, a lot of them don't speak English as a first language so they write the essay either in bad English or their native language and run it through an LLM to clean up the grammar. 'So that's bad - but it's not too bad, right? We would like them to develop writing skills but at least they're learning the subject and researching the core material. 'But some students are just doing it wholesale, and the trouble is that at the end it's really hard to tell the difference. They're both written in the style of an LLM, so I don't know if you wrote the thing in your native language and cleaned it up with AI, or if the AI just did the whole thing for you. 'Even if I suspect it's one of those things it's hard to punish them anyway, because you can't absolutely prove it. You can say 'you've used the word 'delve' about 12 times as much as a normal person', which is a big ChatGPT giveaway, but I can't say 'you used the word 'delve' a lot so I'm going to fail you'. You just can't prove it. 'When they copy blocks of text you can prove it, you can just compare the blocks of text, but you can't do that with LLMs so students basically just get away with it. 'You have a lot more students who have no idea what they're doing at all and they could pass – it's not like they're going to get a First, but they'll pass and do OK. 'The point is you can pass and do OK with almost no knowledge at all, it's a major concern for a lot of people in universities.' A University of Edinburgh spokesperson said: "Marking and assessment procedures vary across the University, reflecting the diverse range of disciplines we offer. "Any decisions around exams and ongoing assessment take into account feedback from students and staff, the specific requirements of each programme and support needed for those who require learning adjustments. "Colleagues can share their insights or concerns through our College and School-level committees, and the University is actively engaging with students and staff on the appropriate use of generative AI. Our published guidance reflects a commitment to upholding academic integrity, while recognising acceptable and innovative uses of new technologies.'


Times
15 hours ago
- Times
They said I was late to AI, but I've picked a winner in Microsoft
Which would you rather back with your money: artificial intelligence or natural stupidity? Do you stand with Demis Hassabis — a Nobel prizewinner and chief executive of the AI researcher Google DeepMind, who last week expressed cautious optimism that AI 'will be ten times bigger than the industrial revolution' — or the losers venting cheap cynicism online? I can't pretend to be neutral about this, having suffered my fair share of abuse after reporting how AI prompted me to invest a little more than 2 per cent of my life savings in the software giant Microsoft (stock market ticker: MSFT) at $233 and $241 in January 2023. Several pessimists said I was too late and predicted doom. Some of them might have felt justified, briefly, when the arrival of DeepSeek, China's AI champion, wiped nearly $1 trillion off American tech giants' stock market value last January. It's early days yet, but those Microsoft shares were trading at $525 on Wednesday and are now the ninth most valuable holding in my 50-stock forever fund, so I really mustn't grumble. Less happily for society as a whole, it remains unclear whether the commercialisation of AI will lead to the 'radical abundance' predicted by Hassabis or mass unemployment. Here and now, Alphabet (GOOGL) is extending its AI search facility to Britain this month, after launching in America and India. Unlike conventional Google, which has proved so successful that its brand has entered the language as a verb, Google AI can answer complex questions at length and in plain English, instead of providing a list of links. Never mind, for now, that this business is essentially disrupting itself, with some advertisers grumbling that fewer folk are linking through to them than they did before. Google had to go higher up the AI ladder to avoid being rendered obsolete by ChatGPT, one of the most successful app launches ever. There's no need to take my word for this, Google AI reports that its rival ChatGPT 'achieved a remarkable feat by reaching 100 million users in just two months after its November 2022 launch'. • Google has signalled the death of googling. What comes next? Unfortunately for investors, ChatGPT is owned by OpenAI, a company that is not listed on a stock exchange. Fortunately, news that Microsoft had invested $10 billion in its unlisted Californian technology competitor in January was enough to prompt this small DIY investor to take the plunge and buy a stake in the future via Microsoft. The company combines long-established streams of revenue — including the world's most popular desktop operating system, Microsoft Windows, plus PowerPoint and Word — with substantial exposure to capital growth in future, through its stake in OpenAI and ChatGPT. Microsoft's modest dividend yield of 0.63 per cent has increased an eye-stretching annual average of 17 per cent over the past five years, according to LSEG, formerly the London Stock Exchange Group. Dividends are not guaranteed and can be cut or cancelled without notice. However, if that rate of ascent could be sustained, it would double shareholders' income in less than four and a half years. So this investor, who hopes to fund an enjoyable retirement, sees it as a relatively safe each-way bet. By contrast, my biggest technology shareholding, Apple (AAPL), has largely failed so far to make the AI trend its friend. Worse still, most iPhones are made in China, and so this business is extremely exposed to the unpredictable trade war between America and China, causing Apple's share price to plunge 12.5 per cent since the start of this year. That plucked it off the top slot in my forever fund, pushing it down to third place by value. But, having originally invested in Apple at $23.75 in February 2016, as reported here at that time, allowing for a subsequent stock split, I remain sanguine about these shares, which were trading at about $212 on Wednesday. One reason is that I suspect Apple Vision Pro, an augmented reality (AR) headset, has been widely misunderstood. This company has long-established success in selling technology to people who aren't that keen on technology, so I think Apple may be first to achieve commercial success with AR — which enhances the real world by overlaying graphics and information — when the price comes down and the choice of apps goes up. This would be a good time to confess that my cerebral software dates from the 1950s. So I can't claim to understand the more technical aspects of these trends, which was why I began my exposure to this sector with an investment trust more than a decade ago. Polar Capital Technology (PCT) shares were trading at 43p each, allowing for subsequent stock split, when I transferred them from a paper-based broker in September 2013. They were trading at £3.97 on Wednesday and may have further to go. Better still for bargain-hunters, shares in the £5.1 billion fund continue to be priced 10 per cent below their net asset value. Fund management charges of 0.8 per cent seem reasonable for professional stock selection in a sector where older investors may struggle to keep up with the pace of innovation. • Read more money advice and tips on investing from our experts For example, when I worked in the City office of another newspaper, three people were employed in the library just to collate clippings about companies listed on the stock exchange. Now all that information, and much more, is available on my mobile. Returning to where we began, I take contrarian comfort from the fact that many critics claim AI is all hype. As I may have pointed out before, perennial pessimism is an easy way to simulate wisdom about the stock market, but it ain't the way to make money. Space is the final frontier for new technology, and few have gone there more boldly, albeit by proxy, than the eccentric billionaire Elon Musk. Even if you wouldn't dream of getting into one of his rockets, with their star-studded passenger lists, there is good reason to consider gaining exposure to Space Exploration Technologies, or SpaceX, as an investment. Musk's antics with the equally eccentric President Trump may have distracted attention from the fact that SpaceX has lifted more than 8,000 satellites into low earth orbit. This has already extended the internet to parts of our planet that were previously offline. Most importantly, from a commercial point of view, if data capacity can be expanded sufficiently, SpaceX and its Starlink wi-fi subsidiary could eventually replace every internet service provider on Earth. Coming down from the clouds of technical speculation, small investors willing to accept high risks can gain a ground-level stake in SpaceX via a handful of investment trusts holding these unlisted shares. To be specific, just more than 14 per cent of Edinburgh Worldwide's £770 million assets is invested in two tranches of SpaceX stock. I paid £1.52 an Edinburgh Worldwide share in January 2024. It has been a bumpy ride, as you might expect, and there are no dividends, but I am happy to hang on to shares trading at £1.94 on Wednesday. Drawing inspiration from the Spitfire that flew over my head a few moments ago, I am reminded of the Royal Air Force motto, per ardua ad astra, which means through adversity to the stars. Perhaps even more appropriately, this small shareholder could say: 'Beam me up, Musky!'


Times
19 hours ago
- Times
Is it time for your firm to hire a ‘head of AI'?
Amazon, Microsoft, Alphabet and Meta are on track collectively to invest more than $300 billion in AI infrastructure and R&D over just a few years. This wall of cash sends a clear message: Big Tech believes its big bet on artificial intelligence will pay off. But for that to happen, it must convince the customers it needs the most: the boards of the Global 2000 — the world's biggest listed companies. When I last interviewed this group in 2023, AI barely made the agenda. Fast-forward two years and much has changed. Intelligent boards are getting smart about artificial intelligence. Leaders are under increasing pressure from shareholders and investors not only to articulate their AI strategy but to deploy what at first appears to be radically productivity-enhancing technologies. The shift from boards being passive observers to actively engaging with AI has been 'notable', according to Jenni Hibbert, global managing partner at executive search firm Heidrick & Struggles. 'It's very clear that many companies have evolved from a period of either observing or piloting early forms of AI adoption, to now becoming laser-focused on aligning AI tools and technologies with their core business goals.' With few board members having a background in engineering let alone machine learning, for many the first step has been to gauge the AI literacy of the board. In the past year we have educated CEOs trying to master AI jargon ahead of shareholder presentations; managers looking to learn from case studies of successful AI deployment; family offices wanting to assess their AI investments; and executives eager to get hands-on with large language models (LLMs) to understand how they are built and how data is trained. But above and beyond, boards want to understand the risks. Risk is not something Big Tech likes to talk about, but managing risk is arguably the No 1 priority of the board. And the risks when it comes to AI are not just numerous, but only just beginning to be understood. They span the potentially unsolvable inaccuracy of LLMs (hallucinations); the damaging implications of copyright and IP infringement; the negative impact on employees' wellbeing and retention, with recent research showing that individuals working alongside AI exhibit lower levels of motivation and satisfaction; the erosion of critical-thinking skills; increased loneliness of users, even those using chatbots for general purpose tasks; to AI downright lying to achieve its goals. Some AI experts even warn that we have all been enthralled by a 'stochastic parrot', a highly deceptive mimicking machine, incapable of producing original content let alone thought. It's no wonder intelligent leaders aren't putting the parrots in charge yet. One person the board is increasingly turning to, to guide and drive AI strategy, is the AI executive. Once a new-fangled role, the AI executive — such as a head of AI — has grown in maturity. Research by Heidrick in 2024 showed that 31 per cent of AI executives now report directly to the CEO, nearly doubling from 17 per cent in 2023. The head of AI's primary skillset is technical expertise, typically coming from a background in advanced analytics or machine learning. However, they also need to be able to communicate, manage cross-functional teams and decode how AI is likely to affect every facet of a company's business strategy. • Palantir raises revenue forecast to $4bn as AI demand takes off One of the people holding this rare combination of skills is Greg Ulrich, chief AI and data officer at MasterCard. Ulrich is responsible for driving AI strategy across the organisation and leads its AI Center of Excellence. Its objectives include model risk assessments, bias testing, developing secure data pipelines and engaging with global regulators. If you are wondering how effective initiatives such as this can be, consider this: Mastercard was once a physical credit card business that has transformed into a global digital payments company, taking it from a market capitalisation of $30 billion in 2010 to more than $500 billion today. It might be on to something, and it is putting education and training at the heart of its AI strategy. In some cases those education opportunities are offered to everyone in an organisation. Paul Hollands, chief data and analytics officer at Axa, pioneered the insurance industry's first generative AI apprenticeship and launched Axa's data and AI academy in 2024, providing learning opportunities to all employees in the UK. During the week it launched, 4,700 people registered. It's an investment in human capital, a belief in human potential even, which flies in the face of the zeal with which many AI leaders discuss the potential of their technologies to automate the workforce. Aside from educating boards on the risks of AI, training the workforce, and reading the tea leaves about what changes it will bring, the head of AI has another essential job: to demonstrate that AI does, indeed, have value — and how. A recent survey by IBM of more than 2,000 global CEOs revealed that only 25 per cent of AI initiatives had delivered their expected results and only 16 per cent had scaled across the business in the past three years. Until the technology becomes more mature, smart boards are right to move cautiously. Kathryn Parsons MBE is founder of Decoded and co-chair of GBx