AI agents have a problem full stop - Rethinking The Hype Cycle #8
Going "AI-first" isn’t always such a hoot🦉
Hello👋
We're back with Rethinking the Hype Cycle, your practical guide to AI and what's next in tech for curious leaders.
These past two weeks have been loaded with heaps of hype, before unexpected U-turns on 'AI-first' hardline strategies, as those pesky customer needs proved harder to handle than initially thought.
Your regular reminder: If you're not working on the bleeding edge, you don't need to bleed.
On to the trends. 👉
🔮AI and frontier tech trends
And the winner in the AI tool race is...
Terribly colour-coded but useful charts show 2 in 5 US businesses now pay for AI tools. OpenAI leads adoption by miles. Anthropic races behind at a gap. X, Google and others' market share is almost non-existent.
As ChatGPT becomes synonymous with AI, much like Zoom was to online meetings during the pandemic, will this stride in advantage lock in its dominance for both business and consumer use? Bear in mind that first-mover advantage doesn't often play out long-term. Zoom's lead withered when incumbent IT took over. We 'Teams' now more than we 'Zoom'. We 'hoover' with brands other than the original from William Henry Hoover. (Henry remains the best hoover brand of all time, though. The rest suck).
Microsoft didn't let 7,000 people go because of AI
"No large-scale empirical studies confirm that Generative AI is actively replacing workers across core business functions."
AI thinker Stephen Klein crunches historical data on Microsoft's recent L-AI-offs, surmising they were reshaping anyway. We're seeing a 'blame it on AI' trend. Companies claim they're scaling back or not hiring as their magical AI machine makes them more productive. In reality, the economy bites in our geopolitical wild times, and many industries are shrinking.
For developers, the AI threat is real. Some organisations opt for AI-led coding with a smaller pool of developers fixing the post-automated mess. This report of a 6-figure salaried New York developer downsizing to a trailer: "AI isn't taking over the tech industry – it's turning it into a sweatshop."
Marketers still aren't getting AI training
If you're worried you're falling behind in the AI 'race', if you're doing anything structured at all, you're probably doing alright. The Marketing AI Institute survey, which skews to content marketers, shows that less than half had any formal training. It’s depressing how little this has shifted since last year.
Getting AI tools without training is like buying a car without a driver’s licence. You could get somewhere, but you may unleash considerable risk to you and the world. I do an AI literacy workshop if you want help shaping your plan.
What I did this quarter by Google
Got 3.5 hours to watch the Google keynote? Slacker! Try this quick 2-minute spin of Google's best bits so far in 2025. Google proves they're a player in the AI space. Let's hope they use AI to start delivering more useful search results.
AI predicts cancer outcomes from selfies
A scientific study shows that this deep learning face analysis tool could predict your biological age to improve cancer prognosis.
SEO is dead? Let's not dig its grave just yet
Interesting LinkedIn chat on how to get ahead with AI search (or try to) with some folks, including Deborah Carver aka The Content Technologist, who really know their onions 🌰
Good content design remains the most likely strategy to win in the shift to more information being served by AI chat than Googling. I've written before about getting ahead of AI bots for thought leadership.
Are we in a post-follower era?
In similar online trends, social media's purpose has shifted. In the old days, we chased followers. Then engagement. Now, most good stuff happens in private chat, what was once called 'dark social'. It’s had a big impact on all our lives, thanks to the ‘Partygate’ UK government scandal with deleted WhatsApp chats and, more recently, Signal groups stacked with bro emojis to discuss sensitive US military operations. If you're growing a brand, you most likely need a mix of these three approaches – some from the brand, some by you and your talent.
AI agents have a problem full stop
RAG (Retrieval augmented-generation) is seen as a solution to rogue AI results by looking up information from your knowledge sources. In theory, this control makes it more reliable, but it needs more crafting to integrate with agentic AI.
TL;DR: most RAG systems push information then stop when they’ve served it up, whereas agents need to prod for questions and actions to move the discussion on to either solve the problem or advance the discussion.
Gen AI makes people more productive but less motivated
HBR reports on a study showing those doing admin and marketing-type tasks with GenAI are more efficient but more bored when doing tasks that need critical thinking unaided. This will be quite some transition.
🚰 Watercooler: The barmy and bluster in big AI hype
Are you Googling ChatGPT like a dad?
Altman claims ChatGPT users have different use cases depending on age: younger folks use it for life advice, older folks as a search engine. I'm feeling seen. And old.
There were alarmist reactions that the data shows young people use the app as a confidante, feeding it intimate information about their thoughts and lives. This has a potentially huge impact on data privacy if this data is sold to brands. We've been here, and still are, with social media, so perhaps nothing new. If you're thinking about your AI content strategy, it's worth considering what your demographic uses AI chat for.
And some serious knives out shade for Altman in this dissection of his kitchen and cooking choices in the FT (this feature usually happens in a restaurant so the home cooking is perhaps a power move from the OpenAI Founder). What a load of drizzle. 🔪
Let's replace those pesky AI agents with humans
Big news here for the AI whisperers and cynics alike: Fintech firm Klarna, the poster child of 'AI first' with its OpenAI collaboration to replace 600 customer support agents with AI, has now realised AI can't really replace people without hard dropping quality. They even bravely said they'd put too much emphasis on cost saving. They will hire more customer service folks, but in a gig worker platform scenario.
Let's not listen too hard to the naysayers who claim this is proof that AI doesn't work. It probably will still play a big role in Klarna and other mass-market online support services. But we could end up with a more balanced reset: AI agents collaborate between online chat and human folks to deliver a good blend of support. And let's be real, it's also possible it didn't save them much or created too many bad reviews from unhappy customers.
Going "AI-first" isn’t such a hoot🦉
Similarly, Duolingo faces customer backlash over its AI-first messaging. Its CEO's tone-deaf response amounted to: "they don't get what we're doing." If you're going hard on the AI message, most people don't like or understand AI, so it's your job to explain it with empathy.
And when it comes to your employees, show extra doses of empathy. Liz G's LinkedIn post resonated with me: in times of trade wars, geopolitical tensions and fear of AI job losses, as a leader, it's your job to provide certainty and steer the ship.
A company reminder for everyone to talk nicely about the giant plagiarism machine
"Critical thinking takes 30% of company time. Some of you aren't focused on the profit potential of outsourcing our thinking to a machine remixing thoughts that have come before."😱
(for absolute clarity: this is a parody. But far too real)
OpenAI goes global
OpenAI for Countries launches as a "clear alternative to authoritarian versions of AI that would deploy it to consolidate power." Given it's a rollout of the US 'Stargate' that seems more like throwing shade to (arguably not that authoritarian) Deepseek.
More tickers here showing we're moving to a geo-separated AI landscape where choice of AI tools will be based on national alliances rather than the best-fit tech for the user.
⚖️ Tech regulation, data security and brand safety
Consumer trust is drooping with GenAI
Survey shows over-65s' trust in media since GenAI adoption has plummeted, yet for people under 35 it's increased. We need to upweight transparency in public channels and adapt how we communicate AI use. And, again, consider how our audience demographics may impact how we talk about using AI. I'm quoted in this study by agency TFD.
I like big tech and I cannot lie
UK business Minister Peter Kyle cosies up to big tech and AI with a plethora of meetings. And, of course, Tony Blair widely grins at the back 🙄 We're seeing a concerning shift here with investment in AI moving towards big tech, and limited growth towards a startup economy for AI businesses.
Republicans want to deregulate AI
So much is on fire in the US in the AI and safety field. MAGA want to ban all AI regulations for 10 years. That also includes states from adopting their own AI regulations, and would make null-and-void some in the pipeline.
More than other recent news, if this comes to pass, it's the most concerning for US citizens. An unregulated AI state would be wildly dangerous, as much of this tech works its way into citizen services and daily lives. If you're in the US, keep an eye on this and take action with your local representatives.
Case in point: putting all your sensitive citizen data in one pot as Musk's DOGE hit-squad wants to do, makes it easier to exploit people and a delight for hackers where X, literally, marks the spot.
And in a Kristallnacht move, the US copyright leader is ousted after a report saying AI training isn't totally great for copyright holders.
"Words are our earliest human technology, like water they appear insubstantial, but like water they can generate tremendous power"
Sage queen of near-time dystopia Margaret Atwood says she cannot remember another time “when words themselves have felt under such threat”. Meanwhile, Academics who study fascism are fleeing US for Canada. We are now entering the prologue bit of her novel, The Handmaid's Tale.😨
Your DNA data sold to big pharma
The unintended consequences of sharing DNA data. As DNA test provider 23AndMe bankruptcy unfolds, now your genetic makeup, and those of your extended family, is in the hands of big pharma. They claim they’ll try to honour the privacy users originally signed up for, but want the big data set to develop new drugs.
Are deep research tools digging deep enough?
Deep research' LLMs are now freely available, with premium subscriptions of ChatGPT and DeepSeek, for example. Research is a painstaking process of reviewing often hundreds of densely worded papers to convey new insight and analysis. Some might say a prime use case for GenAI. But can AI work for the painstaking academic research?
This materials research was pulled due to a high level of fabrication. Looked too good to be true, and it was. Peer review, to an extent, protects rogue research from getting too far, but with increased pressure on academics, research could become less trustworthy.
Take Grok off the block
Is Grok on your list of LLMs you're using or testing? You may want to reconsider. Any AI that's programmed to share wicked fake news about a false genocide should not be passing any of the tick boxes on your software, marketing or governance risk register.
Not united by music
Turns out a small media spend and block voting is almost enough to swing an election (though they didn't count on the power of Austrian opera). Israel government tried to persuade the world that their Eurovision entry was the best, in the face of vocal opposition to their participation in the otherwise delightful, outlandish annual song contest.
Ousted 2024 contestant Joost and 2025 should-have-been runner-up Tommy Cash's take (written before this) is pretty spot on.
🧪 Weird and wonderful new tech
Unexpected AI in the bagging area🍷
Does AI have a sweet bouquet or a velvety aftertaste? Great example of why if you're using AI in your product development, you need to explain what it does and the benefits, or avoid looking like a total plonk(er).
Robotic centipedes are coming for invasive weeds
Robotic centipede gets in the weeds to help farmers root out pesky unwanted fauna in tough terrain. This device looks spooky but lovely.
Robot centipede = good 🤖
Human centipede = bad 🐛
A very touchy robot
Robots lack the human touch – or rather the sensitivity needed to correctly handle things. Vulcan is a more touchy-feely robot co-working in Amazon warehouses. In the PR they're achingly keen to emphasise the exciting new jobs these cobots who will do 75% of warehouse fulfilment will create.
"We've done away with human beings in this department."
Brilliant satirist Michael Spicer takes on the uncanny valley of a glitchy chatbot that wants to create images of the Pope instead of dealing with a fraudulent transaction. Well, wouldn’t we all. Neatly illustrates why unstructured AI really doesn't work for complex customer services.
💼 AI business use cases
American schools were deeply unprepared for ChatGPT
Excellent investigation by 404 Media, one of the trailblazers of independent journalism, on how unprepared most US states were for arrival of AI in classroom. Big AI funded trainers in pro AI propaganda to persuade educators to "ditch the textbook". 😲
I wrote before about education and AI. It feels like we need a gentle slope here to empower schools with new tools.
Of course, some teachers are early adopters. They turn to AI to mark kids' work. Have a guess how accurate it is (hint: 🐒 + 📌).
Do we still need to learn prompt engineering?
I know many don't agree with me, but we’re moving away from the nerdy skill of prompt engineering. I'd call it now 'better briefing'. If you know how to manage work requests and articulate clearly, LLMs today can get you far without needing a library of prompts.
For now, the need is still real. I've looked at a few guides relevant to marketers and content creators. This one's worth a bookmark with some easy to follow templates. It's from Anthropic/Claude but guidance will likely work for any AI tool.
What would good AI look like?
There's plenty of worthy AI examples, but how about we apply some effort to what good AI can look like for everyone? Anil Dash has a wish list. With better collaboration and less hype we could do this.
Are you hiring a human or a machine?
AI video job interviews may be sold to employers as more efficient and less biased, but are they? The disembodied robot interview may not hold water for those with neuro-divergent folks (or those who would prefer talking to a human to work out if they are interested in working for you or not).
The 'talent problem' sees candidates firing out thousands of AI-optimised CVs to target AI-systems that filter most people out. One guy used a tool to blast out 5,000 applications and while it wasn't that successful by response rate, he did land a fair amount of interviews for limited effort.
Think about how you can keep the human touch in hiring and applying fairness. Otherwise, you could shortlist those who are good at AI optimising more than job doing.
Mind the gap
Thinking ahead: We need to accept that AI will be able to do, or shrink, lots of information-processing tasks that are a part of many knowledge-based roles. Thinking ahead about your career, what skills can you develop in the grey space where AI breaks?
➡️ What to do next to get ahead
Plan for human-AI collaboration
AI works best when enhancing those pesky humans rather than replacing them entirely. Design workflows where AI handles routine tasks so your folks can spend more time dealing with the knotty problems your customers/clients/stakeholders have.
Invest in AI literacy, not just AI tools
With less than half of marketers receiving formal AI training, there's a competitive advantage in proper education. Create structured learning programmes that teach your team how to prompt AI effectively and recognise its limitations while keeping critical thinking skills.
Prepare for the geopolitical AI divide
With OpenAI's global expansion and proposed US deregulation, consider how AI tool choices might become politically influenced. Diversify your AI toolkit now and evaluate tools based on functionality and data sovereignty rather than just today’s model performance.
Plan how to switch off
AI can make us more productive but less motivated in critical thinking. How can you do some non-AI switch up to keep your cognitive skills in check? Going offline – Post-it notes, hand-written sketches, a walk or quiet time without music and chat – can help you connect with your thoughts.
🌍 I'm out and about
This month, it's all about AI and careers.
Grow your personal brand with AI
Fri 23 May, 5pm GMT (online)
I'm talking with Amy Au on how women in STEM can use AI to grow their leadership brand. It will be fun (promise). Bring vino (or if you're lucky enough to still have the Friday drinks trolley, mine’s a G&T).🍷
AI's impact on investing and financial careers
Thurs June 5, 9:30pm GMT (online)
I'm joining a panel for the Women in Fintech Initiative to talk about how fintech folks can plan for the AI impact on their career growth.
I've also updated my speaker topics for 2025 with talks on inclusive AI, frontier tech trends and my favourite, Move slow and fix things, about rethinking AI hype. If you need a speaker who can bring some levity and real-world insight to help leaders move past tech hype, you know who to call (me. Not Ghostbusters).
🔁 ICYMI in Rethinking the Hype Cycle
Will AI bring us joy, division or tear us apart?
Social tech erodes how we connect. Here's how we can bring it back.
To think big about AI's potential, start small
Could small language models plus slow and steady actions move you faster?
Back next week with more thoughtful (or perhaps scrappy) insights, and in two weeks with another trends round-up. Sign up to get it first.
Until then, keep it curious 🤔
Susi O'Neill
EVA trust in tech www.evadigitaltrust.com
People-centred tech and AI adoption 🤝 Transparent communications 🎙️ Tech talks and inspiration. Need this? Get in touch