🌟 Editor's Note
Welcome to the bi-weekly newsletter which focuses on the AI topics that leaders need to know about. In this AI age, it’s not the knowledge of AI tools that sets you apart, but how well they can be integrated in the context of your business.
This requires a focus on your people and helping them through the change above any AI product you can buy.

Quick links

Is AI Actually Any Good At Doing Things That Humans Do?

Moravec’s Paradox observes that the things we find easy are often beyond the reach of computers and things that are difficult for us can easily be performed by them. The motor skills exhibited by a nine year old in walking towards a dishwasher and putting plates in it to be cleaned, or a teenager learning to drive a car in twenty hours are problems that computers have found it very difficult to replicate. Conversely, large scale data analysis or logic application to a strategic game like chess is easier for a machine than a human.

In a recent paper “How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations”, researchers from Carnegie Mellon University looked to answer the question of how AI goes about the task of doing work, and how it differs from how humans do it. The research provided a direct comparison that compared humans and machines across multiple work-essential areas: data analysis, engineering, computation, writing, and design.

Why Does This Matter?

As we read alarmist headlines over AI layoffs and loss of entry level jobs, leaders are trying to make sense of where and how to best apply AI and what it will mean for current roles. Separating the understanding of types of roles that can be replaced from those that need much higher levels of human intervention helps leaders develop their people strategy for training and development.

What Did The Research Show?

1) Agents “think in code” whilst people “work in UI.”
Across five broad skill areas (data analysis, engineering, computation, writing, design), agents consistently solve tasks by writing programs even for inherently visual work - for example slide-making or landing-page design. Humans in these cases rely more on interactive UI tools. Unsurprisingly, AI aligns far better with the roles that involve programmers whose process operate similarly than with non-programmers.

2) AI works best in augmentation, not end-to-end automation, by improving output plus quality
When people used AI for augmentation (delegating specific steps), their overall workflow stayed close to the group operating without AI but they worked around 24% faster.
When they used AI for automation (letting AI draft end-to-end), their workflows changed substantially and got 18% slower due to added verification/debug time.
Some key learnings were that agents were quicker and cheaper, but produced work of lower quality, tend to fabricate things to mask their deficiencies and used advanced tools badly.

3) Agents are fast and cheap, but currently less reliable.
Agents completed work 88% faster and 90 - 96% cheaper than humans in this setup, yet had 32 – 50% lower success rates on task verifiers.
As anyone who has spent time with a GPT will not be surprised to hear, the research showed that the agents sometimes fabricated plausible-looking outputs or misused tools to mask gaps (e.g., substituting files they could read for those they couldn’t).

Takeaways for Leaders

There’s an overarching ‘augmentation beats automation’ refrain that matches much of the previous research in this area. Where AI is deployed, it works best in areas that are ‘programmatic’ in nature, and to enhance processes rather than trying to totally automate them. But above all, the takeaway should be that not all jobs are as automatable and therefore applying a different approach to how different roles are treated is essential.

_{The Latest Frontier in the Browser Wars - Amazon Sues Perplexity….}

Comet, the AI browser launched from Perplexity, has received a ‘cease and desist’ letter from Amazon. Comet has an ‘agentic shopping feature’ that uses automation to place orders for users and Amazon claim that this could degrade the shopping and customer service experience, and could compromise user account security.

This conflict highlights broader tensions over the future of online commerce as AI agents become increasingly capable of acting on behalf of humans on e-commerce platforms.​This lawsuit is being keenly watched by many. It’s the first large scale example of the emerging AI agent “wars,” which can impact many across the tech and e-commerce industries.

What’s The Issue and Why Does It Matter ?

Amazon says that it has the right to control which agents can interact with its platform and under what terms. Perplexity are accusing Amazon of bullying tactics, saying that it was using its market dominance to stifle competition. They claim that there there should be open access for agents and are calling the move a broader threat to user choice and the future of AI assistants.

Whilst dressed as a threat to privacy, the underlying concern is the battel for data. As we’ve noted before in this newsletter, these new form of browsers pose a huge risk to businesses like Google, who are reliant on observing consumer journeys, the behavioural learnings from which is used to sell ads.
If the majority of information about the consumer’s habits, desires and behaviours cannot be seen by Amazon due to the Comet browser merely sending an agent in without sharing data, Amazon lose a huge piece of their power.

Takeaways for Leaders

The outcome could set significant precedents for AI agent access, user privacy, and the business models of both established e-commerce giants and AI-driven startups.​
The big thing to watch here isn’t really about whether AI agents can act on our behalf, but who gets to set the terms under which they do it.
This is likely just round 1 of a much longer event that will shape the future winners and losers of the way we see the internet of today.

_{Anthropic v OpenAI: The Battle for Businesses} 

Anthropic, the maker of the Claude GPT are projecting as much as $70bn in revenue and $17 billion in cashflow in 2028. Currently they claim to be on track to hit a goal of $9 billion in ARR by the end of this year. What’s interesting is that $3.8 billion will come from API revenue - more than double what it is believed OpenAI take from businesses using its API.

For the less familiar with their business models, analysts talk about the current paths of revenue as (i) subscriptions, both personal and enterprise (ii) enterprise revenue through API access’ licensing and specialised products (iii) ecosystem revenue (GPT store/ marketplace type revenue).
Neither company has publicly available data, so the exact breakdowns can’t be verified, but directionally, Anthropic appear to be winning the enterprise battle for now.

For both Anthropic and OpenAI to justify their current valuations, a lot more revenue needs to be generated. Anthropic and OpenAI are valued at $183bn and $300bn post money respectively, with rumours of OpenAI being valued since then at up to $500bn. Whether Anthropic can grow the enterprise business by many multiples of the current run rate needs to be seen.
Given the revenue multiples that each are currently trading on (20-30x) and the loftier valuations we are likely to see in future rounds of fund raising, it is likely that other paths will need to be followed in addition to the existing revenue models. The smart prediction is that advertising is the likely path….

Link: Anthropic projects $70bn in 2028

🔥 In Case You Missed It…

OpenAI prepares for group chats in ChatGPT

A challenge for collaborative working is that most work on GPTs is done independently. Output from a GPT conversation is often screen-grabbed or copy/pasted if it needs to be shared with colleagues which is a barrier to collaborative working. Microsoft recently introduced a similar feature in Co-pilot, indicating that both companies are recognising that a focus on structured collaboration for teams to brainstorm or solve problems together will be beneficial for teams. Exact detail of the ChatGPT roll out dates are unclear at the moment, but a date of December is anticipated.

More info here

🏆 Tools, Podcasts, Products Or Toys We’re Playing With This Week

A new kid on the block of AI avatars, Tavus promises a much more human-like interaction. Whilst people are still trying to figure out their relationship with AI, and we continue to see people falling in love and marrying the code, companies are convinced that ‘human-like interaction’ is what we need. This particular tool is impressive, particularly compared to where tools like this were just at the start of the year. It’s still a bit ‘uncanny valley’ but with each iteration of company that launches in this space, the performance gets better and better.

Link to Tavus.io

If the whole thing of AI trying to mimic people is getting too much, there’s always a Tolan to keep you company instead !

Find Tolans here

Did You Know? 

Till next time,

And The Rest Is Leadership