The copilot framing is right for now. The harder structural read: every previous "augmentation" technology became a replacement technology once the unit economics tipped. Spreadsheets augmented accountants until they didn't. ATMs augmented tellers until they didn't.
The 15% B2B end-to-end resolution rate is a starting point, not a ceiling. Once account intelligence improves and context windows expand, the math flips. Staffing hybrid teams costs more than letting AI handle 80% of tickets at 90% quality. The CFO makes the call before the product team does.
The augmentation phase is what every replacement looks like in year three. The interesting question isn't whether AI replaces support. It's how long the augmentation window lasts before unit economics force the next move.
1) Coding jobs are becoming less available or at least there are mixed signals on that in the US. Maybe there are more builders, but, lets be honest, not everyone has talents and interest to be an entrepreneur.
And well, logically, if I can use AI to automate some aspects of developer work, I will need less developers to ship the same number of features. There is limited value in shipping more features. Platform complexity, GTM bottlenecks and well slow/ no growth in demand put a natural limit on your production.
2) Two things are true of customer service:
-- In a mature org that is well designed enough to implement AI effectively, customer support is likely already pretty lean through automation and outsourcing. It is a cost-centre after all, however much gloss you put on it. So there isn't much to cut with AI. Smart routing is a decades-old technology, AI probably made it better, but marginally.
-- People running customer services orgs have a strong incentive to keep at least some people under management, because, well, why would you need them if there is no contact centre to manage. Also, having human at least as a back-up for hard cases creates an impression of a luxury service for customers
The labour problem will come, if AI will automate petty bureaucrats and middle-managers. The more common middle-class/lower middle-class jobs nowadays. When industrial revolution hit and agriculture was intensivised, there was already a large demand for unskilled/semi-skilled peasant workers in the factories. When NAFTA hit the industrial jobs, there was nowhere former industrial workers could go, but at least some of their children could take bureaucratic bullshit jobs.
If you automate accountants, clerks, junior/ middle PMs and HR -- where would these people go? Industrial output is growing, but automated factories don't create as much demand for industrial workers. Especially if the economy is not booming, which is true for the most of the world. Similary new builders would not create huge demand for medium-skilled HR managers.
Patient experience in healthcare demands a much higher outcome. Complete resolution for a majority of incoming patient queries, not just triage to the right human. The crisis is there aren’t enough humans, or enough documented knowledge to service the patient population.
The easy/hard distinction is the most important line in this piece. Easy tickets automate well but they were never the time problem. Hard tickets resist AI resolution because they require investigation first. That's where the hours actually are. The real copilot opportunity in B2B is in investigation, not resolution.
The 5.3 versus 3.9 messages-to-resolve gap is easy to misread as AI making tickets harder, when it's the reverse. The tickets it sends straight to a human are pre-selected for difficulty, so the higher message count is the triage working, not failing. The same selection effect sits under the deflection numbers: the 15 versus 35 percent B2B-to-B2C spread tracks how much is safely automatable, not model quality.
The fact that AI silently triages two-thirds of tickets is the real story here and it's exactly the work that deflection rate as a metric is structurally blind to. Feels like the category is overdue for a "context lift" or time-saved-per-human-touch metric that actually credits the copilot for the work it's doing or something like that.
Is any of this meaningful? "Support" is a signal that the service a customer is using is broken, ill-designed, or is some deliberate dissuasion mechanism. It's a pointless cost to the business: users whose money they already have, causing them more costs for no reason (the business already decided not to fix the cause of the problems requiring the customer support in the 1st place).
Making charts about this stuff is pure fantasy as far as I can work out?
The copilot framing is right for now. The harder structural read: every previous "augmentation" technology became a replacement technology once the unit economics tipped. Spreadsheets augmented accountants until they didn't. ATMs augmented tellers until they didn't.
The 15% B2B end-to-end resolution rate is a starting point, not a ceiling. Once account intelligence improves and context windows expand, the math flips. Staffing hybrid teams costs more than letting AI handle 80% of tickets at 90% quality. The CFO makes the call before the product team does.
The augmentation phase is what every replacement looks like in year three. The interesting question isn't whether AI replaces support. It's how long the augmentation window lasts before unit economics force the next move.
I think there are two things to note here.
1) Coding jobs are becoming less available or at least there are mixed signals on that in the US. Maybe there are more builders, but, lets be honest, not everyone has talents and interest to be an entrepreneur.
https://www.washingtonpost.com/business/2025/03/14/programming-jobs-lost-artificial-intelligence/
And well, logically, if I can use AI to automate some aspects of developer work, I will need less developers to ship the same number of features. There is limited value in shipping more features. Platform complexity, GTM bottlenecks and well slow/ no growth in demand put a natural limit on your production.
2) Two things are true of customer service:
-- In a mature org that is well designed enough to implement AI effectively, customer support is likely already pretty lean through automation and outsourcing. It is a cost-centre after all, however much gloss you put on it. So there isn't much to cut with AI. Smart routing is a decades-old technology, AI probably made it better, but marginally.
-- People running customer services orgs have a strong incentive to keep at least some people under management, because, well, why would you need them if there is no contact centre to manage. Also, having human at least as a back-up for hard cases creates an impression of a luxury service for customers
The labour problem will come, if AI will automate petty bureaucrats and middle-managers. The more common middle-class/lower middle-class jobs nowadays. When industrial revolution hit and agriculture was intensivised, there was already a large demand for unskilled/semi-skilled peasant workers in the factories. When NAFTA hit the industrial jobs, there was nowhere former industrial workers could go, but at least some of their children could take bureaucratic bullshit jobs.
If you automate accountants, clerks, junior/ middle PMs and HR -- where would these people go? Industrial output is growing, but automated factories don't create as much demand for industrial workers. Especially if the economy is not booming, which is true for the most of the world. Similary new builders would not create huge demand for medium-skilled HR managers.
Patient experience in healthcare demands a much higher outcome. Complete resolution for a majority of incoming patient queries, not just triage to the right human. The crisis is there aren’t enough humans, or enough documented knowledge to service the patient population.
The easy/hard distinction is the most important line in this piece. Easy tickets automate well but they were never the time problem. Hard tickets resist AI resolution because they require investigation first. That's where the hours actually are. The real copilot opportunity in B2B is in investigation, not resolution.
The 5.3 versus 3.9 messages-to-resolve gap is easy to misread as AI making tickets harder, when it's the reverse. The tickets it sends straight to a human are pre-selected for difficulty, so the higher message count is the triage working, not failing. The same selection effect sits under the deflection numbers: the 15 versus 35 percent B2B-to-B2C spread tracks how much is safely automatable, not model quality.
The fact that AI silently triages two-thirds of tickets is the real story here and it's exactly the work that deflection rate as a metric is structurally blind to. Feels like the category is overdue for a "context lift" or time-saved-per-human-touch metric that actually credits the copilot for the work it's doing or something like that.
Is any of this meaningful? "Support" is a signal that the service a customer is using is broken, ill-designed, or is some deliberate dissuasion mechanism. It's a pointless cost to the business: users whose money they already have, causing them more costs for no reason (the business already decided not to fix the cause of the problems requiring the customer support in the 1st place).
Making charts about this stuff is pure fantasy as far as I can work out?