Coming from a consulting background, I resonate with your point about the apprenticeship gap. It can be overlooked (not in a bad way) due to generous leaders who prefer to give juniors an opportunity to shine. With AI crushing the junior layer, it would be great to hear the thoughts from your interview and podcast guests about apprenticeship in the world of redefined jobs and services.
Have been seeing the writing on the wall for some time now in B2B marketing services (in the UK). Agencies and consultants have been used to bundling lots of time-based ancillary work in their invoicing (producing reports, audits, SEO content, PPTs etc), as it's very difficult to cost creativity and judgement in ways that clients will accept. (I remember working in an ad agency where an invoice to P&G was sent back with the creative work cost crossed out, with the words "We do not pay for creativity.")
Business units were happy to pay for that sort of busywork as it was stuff they could pass on up to show how hard they were working. Now that LLMs are "good enough" to take almost all of the friction out of that sort of work, agencies and consultants must find another pricing model.
What makes this an even more difficult environment is that LLMs are poised to produce online ads personalised with social media understanding of individual engagement triggers (Facebook is promising that already), which can be produced at speed and scale (with customised imagery and text content fitting the brand voice perfectly). And A/B tested almost infinitely to get the optimum response - something an agency would previously have charged a lot for (and much slower).
Brands that are very serious about a high-level image will need more than just "good enough" - but for an awful lot, cheap, updatable quantity may be much easier to keep in-house than pay for externally...
Hi Grace, great perspective on MOAT constraints from the Agentic AI and professional services perspectives. Agent orchestration heavily depends on JUDGMENT rather than Speed-to-Market within the Vibe coding environment. Despite the known friction along the agentic AI transformation, retention of current workflow expertise will continue until there is a sustained plateau of decreasing return on token productization. The latter will take at least another 5 years. Meanwhile, the legacy professional experts are indispensable to inject consistency and credibility in the context of AI Applications performance with limited success other than cheaper first rollout software yet without solid relationship maintenance.
The repricing of knowledge work is already happening at the tool-buying level too. Mid-market companies used to hire consultants to evaluate and implement software - now the same analysis is something ops and IT leads can do with the right frameworks and AI tools. The consultant gatekeeping is eroding from both ends.
I think lawyers, consultants, and similar are actually fairly rent seeking in society. As economic activity increases, they’ll see growth. It’s less established entities that look similar that’ll see depression, but ofc I’m not an economist.
I’d expect more disruption to the back office rather than front. People managing systems, inbound, monitoring, auditing, and junior people roaming around.
I think the initial disruption will be in the operational roles. But as this article argues it aligns with what many are saying now, you can outsource your thinking but not your understanding. Hence the knowledge workers that used to do a lot of the initial grunt work of “thinking” can be then assisted or replaced in some capacity but in the longterm whether it’s in lawsuits or m&a deals or whatnot, lawyers will be paid and required for the “understanding” and even to further step which is ability to back the understanding.
I feel like sharing some of my recent experiences in biopharma/biocomputation again after reading this. I am building out some pipelines, which is a fancy way of saying I am using open source orchestration software made for stringing together open source biocomputational software to transform certain data inputs into certain data output. It is great for AI to perform all steps of, but debugging is a key step. Claude Code is great for working on this either with a lot of permissions and autonomy, but I use it more with me in the loop to save on tokens--trading my time for token use. In an ideal world where I was short on time and long on budget, I could see myself setting up CC on several HPCs running the same workload to debug or build. Doing everything by hand would still take a few hundred hours if including time to download repos, wheels, and build containers, so being able to set and forget would free me up for other tasks like.. bioinformatics.
Gathering information is stupidity fast and easy, and you can get amazing coverage and accuracy by how you prompt to add methods of verification. These tools are amazing for scraping without fussing with antibot mechanisms or managing different APIs. Managing my Claude web chats, skills, and projects has let me build really useful databases of tools, cell markers, patents, research papers, and more... I have been able to use these resources to design cutting edge pipelines I want to implement or greatly speed up annotation, or speed up data processing (using handy scripts, but having AI run the data through the scripts). And I can create workflows using a local orchestration model to utilize a lot of the data... And I can use Claude to help me with specific code to achieve specific goals using R packages I haven't touched in years. My end goal at the place I work is to cut out my job--I need a local interface that the people in the lab can just prompt (because they may be comfortable using a webserver hosting an ancient tool, but they can't handle seeing code or terminals) and get their pipeline or data retrieval or data processing done instead of piling work on my desk and dragging me off to hours long meetings.
Coming from a consulting background, I resonate with your point about the apprenticeship gap. It can be overlooked (not in a bad way) due to generous leaders who prefer to give juniors an opportunity to shine. With AI crushing the junior layer, it would be great to hear the thoughts from your interview and podcast guests about apprenticeship in the world of redefined jobs and services.
Have been seeing the writing on the wall for some time now in B2B marketing services (in the UK). Agencies and consultants have been used to bundling lots of time-based ancillary work in their invoicing (producing reports, audits, SEO content, PPTs etc), as it's very difficult to cost creativity and judgement in ways that clients will accept. (I remember working in an ad agency where an invoice to P&G was sent back with the creative work cost crossed out, with the words "We do not pay for creativity.")
Business units were happy to pay for that sort of busywork as it was stuff they could pass on up to show how hard they were working. Now that LLMs are "good enough" to take almost all of the friction out of that sort of work, agencies and consultants must find another pricing model.
What makes this an even more difficult environment is that LLMs are poised to produce online ads personalised with social media understanding of individual engagement triggers (Facebook is promising that already), which can be produced at speed and scale (with customised imagery and text content fitting the brand voice perfectly). And A/B tested almost infinitely to get the optimum response - something an agency would previously have charged a lot for (and much slower).
Brands that are very serious about a high-level image will need more than just "good enough" - but for an awful lot, cheap, updatable quantity may be much easier to keep in-house than pay for externally...
Hi Grace, great perspective on MOAT constraints from the Agentic AI and professional services perspectives. Agent orchestration heavily depends on JUDGMENT rather than Speed-to-Market within the Vibe coding environment. Despite the known friction along the agentic AI transformation, retention of current workflow expertise will continue until there is a sustained plateau of decreasing return on token productization. The latter will take at least another 5 years. Meanwhile, the legacy professional experts are indispensable to inject consistency and credibility in the context of AI Applications performance with limited success other than cheaper first rollout software yet without solid relationship maintenance.
The repricing of knowledge work is already happening at the tool-buying level too. Mid-market companies used to hire consultants to evaluate and implement software - now the same analysis is something ops and IT leads can do with the right frameworks and AI tools. The consultant gatekeeping is eroding from both ends.
this is one of my best 25 min spent in a long time
I think lawyers, consultants, and similar are actually fairly rent seeking in society. As economic activity increases, they’ll see growth. It’s less established entities that look similar that’ll see depression, but ofc I’m not an economist.
I’d expect more disruption to the back office rather than front. People managing systems, inbound, monitoring, auditing, and junior people roaming around.
I think the initial disruption will be in the operational roles. But as this article argues it aligns with what many are saying now, you can outsource your thinking but not your understanding. Hence the knowledge workers that used to do a lot of the initial grunt work of “thinking” can be then assisted or replaced in some capacity but in the longterm whether it’s in lawsuits or m&a deals or whatnot, lawyers will be paid and required for the “understanding” and even to further step which is ability to back the understanding.
I feel like sharing some of my recent experiences in biopharma/biocomputation again after reading this. I am building out some pipelines, which is a fancy way of saying I am using open source orchestration software made for stringing together open source biocomputational software to transform certain data inputs into certain data output. It is great for AI to perform all steps of, but debugging is a key step. Claude Code is great for working on this either with a lot of permissions and autonomy, but I use it more with me in the loop to save on tokens--trading my time for token use. In an ideal world where I was short on time and long on budget, I could see myself setting up CC on several HPCs running the same workload to debug or build. Doing everything by hand would still take a few hundred hours if including time to download repos, wheels, and build containers, so being able to set and forget would free me up for other tasks like.. bioinformatics.
Gathering information is stupidity fast and easy, and you can get amazing coverage and accuracy by how you prompt to add methods of verification. These tools are amazing for scraping without fussing with antibot mechanisms or managing different APIs. Managing my Claude web chats, skills, and projects has let me build really useful databases of tools, cell markers, patents, research papers, and more... I have been able to use these resources to design cutting edge pipelines I want to implement or greatly speed up annotation, or speed up data processing (using handy scripts, but having AI run the data through the scripts). And I can create workflows using a local orchestration model to utilize a lot of the data... And I can use Claude to help me with specific code to achieve specific goals using R packages I haven't touched in years. My end goal at the place I work is to cut out my job--I need a local interface that the people in the lab can just prompt (because they may be comfortable using a webserver hosting an ancient tool, but they can't handle seeing code or terminals) and get their pipeline or data retrieval or data processing done instead of piling work on my desk and dragging me off to hours long meetings.