AI two-step

11 minutes estimated reading time

The phrase AI two-step is something I first heard from my friend Antony Mayfield. He used it to talk about how companies were adopting the latest developments in AI for business processes. And then reduce headcount to reflect the newly AI derived tasks instead.

The AI two-step isn’t necessarily a new concept, companies like Pegasystems were using rules-based systems to take away the drudgery of back office work in banking and fund management for decades.

Further back, companies like Experian, through their access to CCS’ CardPac software provided a service for credit card issuers in the UK using rules-based credit scoring and applications approval. This ran on time-shared mainframe computing resources, which also provided Experian with a good source of ongoing credit worthiness data. All of which reduced the back office work and employees needed by the credit card company. MBNA used to make a virtue out of having every decision reviewed by real live credit analyst, who could overwrite a scoring decision if they saw a compelling reason to do so. (CCS became part of First Data and eventually part of Fiserv).

As these services were being rolled out, there was a corresponding cut in jobs.

Examples

Here are just a few examples of businesses adopting AI, some of which are prime examples of the AI two-step.

IBM

While IBM may no longer trumpeting its Watson AI service as loudly as it used to, AI methods are dispensing with the need to replace staff who leave the technology company.

Pfizer’s Charlie

One might think in the UK that Pfizer should have thought a bit more carefully about the name Charlie, but the aspiration behind the platform is interesting. Charlie was noted to be helping with content creation, fact checking and legal reviews. Research by Bain & Company have found that it isn’t just Pfizer in the pharmaceutical and biotechnology sector that are taking this approach. Some 40 percent of executives who were surveyed said that uses of generative AI were factored into their 2024 budgets.

Bain indicated uses across a wide range of business functions within pharma:

IT programming code review
Competitive intelligence
Research and biomedical literature review
Marketing copy
Augmenting the selling process as a sales co-pilot and contact centre automation

Publicis

French listed marketing combine Publicis made a high profile adoption of machine learning and AI-based services back in 2017 under the moniker Marcel. Back then Marcel was being used for workflow type tasks and organisation of data. This year Publicis rebranded its approach to the less playful CoreAI, so far it has cut the use of freelance staff – which are usually essential for project delivery in ad agencies, rather than the usual AI two-step of lay-offs.

UPS

UPS adoption of AI techniques in everything from workflow to customer service allowed the logistics company to make the largest lay-offs in its 116-year history.

Clear analogues to the AI two-step?

Various commentators compare the AI two-step happening to the dot com boom of the mid-1990s to the early 2000s. The comparison with the dot com boom is easy at first. You have businesses that have phenomenal share price growth, widespread interest and experimentation. Business sectors from advertising to Hollywood are concerned about massive disruption.

The examples I would think about would be factory automation and business process re-engineering. In factor automation, over decades companies used machines to negate the need for unskilled and semi-skilled workers. A friend of mine worked in Huddersfield in a textile mill. He was one of just a couple of people who worked a shift. None of them were weavers, they were engineers and an IT admin who maintained the lines of machines turning out high-end suiting fabric that was mostly sold to Japanese clothing manufacturers. This came very close to being a ‘lights out production line‘ where the product is handmade by robots as they used to say in the old Fiat car advertisements.

Weavers and machine operators were replaced by a lot fewer, but more expensive roles.

Business process re-engineering was driven by enterprises implementing enterprise software to drive efficiencies and automate workflows. This was a lucrative time for consultancies who were brought in to shape a company’s workforce and processes to fit a software company’s pre-defined template for that industry. This was usually based on average industry standards. Software giant SAP have been building and refining these templates for the best part of 50 years, each industry template draws on individual units that might cover a business function like HR, finance or asset management.

A bit of software customisation was needed to fit a given business, and it might have to interface with third party products to handle market complexities such as different tax regimes.

The consultancy teams also laid-off employees that didn’t fit the framework. That’s what business process re-engineering actually meant.

Automation was responsible for putting up to 47% of American jobs at risk. However other research indicates that new forms of skilled or professional jobs are being created. One of the big problems with this data is that they are speculative models. More positive takes from businesses fuelling automation like McKinsey and Company versus more critical predictions from government think tanks and academics.

Factory automation and business process engineering are both similar to the use of AI in business, in that they are primarily helping mature businesses maintain their position and drive efficiency. The dot com boom on the other hand was much more disruptive and spawning more upstart businesses – some of which were very successful and leaving mature businesses struggling to cope. From financial services to media – pre-internet businesses are still struggling to cope with the innovation and disruption that begat the dot com boom.

Optimists versus pessimists

The optimists highlight a number of nuances that they think mediates the impact of automation and machine learning over time.

Tasks over jobs.

It’s tasks rather than whole jobs are being lost. Yet if you look at the data that Scott Galloway shared in his newsletter and the speedy ‘these job losses aren’t down to AI denials’ this optimistic assumption is pure fiction. The jobs being lost are the second part of the AI two-step.

Creative destruction.

Jobs are being created too and it’s often about ‘skill shifts’ rather than ‘job shifts’. While there are redundancies being made, there is a requirement (at the moment) for people skilled in writing ‘prompts’ to get the most out of the AI models created.

Overconfidence.

Overconfidence in technology and what it can do. An extension of this is a belief in the perfectibility of technology. A classic example of this is Air Canada’s recently aborted use of an AI-powered customer service chatbot. The airline quietly pulled its chatbot offline after being found legally liable for bad advice given by the customer service bot to a customer.

Moffatt booked airfares and retrospectively submitted an application for a refund to the reduced bereavement fare after travelling. Air Canada denied the request. Moffatt challenged that decision, saying he was owed the refund because he had relied on the information provided to him by the chatbot on Air Canada’s website. Air Canada admitted that the information provided by the chatbot was “misleading”, but it contested Moffatt’s right to a refund, highlighting that he had been provided with the correct information via the link the chatbot shared in its message.

The Civil Resolution Tribunal considered whether Air Canda was liable for negligent misrepresentation, which arises under Canadian law when a seller does not exercise reasonable care to ensure its representations are accurate and not misleading. Moffatt was required to show that Air Canada owed him a duty of care, that its representation was untrue, inaccurate or misleading, that Air Canada made the representation negligently, that he reasonably relied on it, and that that reliance resulted in damages. The court held that Moffatt met those requirements.

The Civil Resolution Tribunal noted that Air Canada had argued that it could not be held liable for information provided by one of its agents, servants or representatives, including a chatbot, but had not explained the basis for that suggestion. The Civil Resolution Tribunal rejected as a “remarkable submission” Air Canada’s suggestion that the chatbot was a separate legal entity that was responsible for its own actions.
Air Canada chatbot case highlights AI liability risks by Meghan Higgins, Pinsent Masons

Demographic change.

Demographics – the idea that aging countries from the west to China, Japan and Korea have skills deficits due to population decline. Automation is one of the coping mechanisms alongside globalisation and migration that have been suggested solutions. The Chinese are also looking at building factories in countries like Ethiopia, who have a young and growing population. Automation makes sense where migration would adversely affect social cohesion and the cost of globalisation would be more expensive than automation technologies. Workers in the global south are dependent on being cheaper than machines, rather like the American legend of John Henry versus the steam engine.

Companies like Automata have been looking to help businesses automate repetitive low skilled work, such as sandwich making in food service factories or low volume manufacturing tasks.

The state of automation in different roles is running along at different rates of progress. While John Deere have managed to make the most of arable farmland through the use of telematics and GPS guidance of tractors, automating farming for tasks like harvesting is proving more difficult. This is exasperated in the UK at least by the challenge of getting sustained v enture capital for hardware. Technology automation in other sectors such as construction and healthcare continues to move at a slow pace.

In an area like consumer electronics we have seen benefits and declines in automation. Benefits in the way a company like Apple can manage a sophisticated global supply chain workflow via automated software. Apple has also pioneered the use of robotics in dismantling its more modern smartphones when they are brought in to be recycled.

The declining area has been one of design choice. Prior to the smartphone and broadband internet, companies like Panasonic designed circuit boards that were less dense with components. The reason for this was to facilitate automated board manufacture through the use of ‘pick and place’.

Nokia used similar techniques for its cellphones and smartphones until the business was disrupted. Apple iPhones needed much more manual assembly because of the tightly packed components in their phones. Young women were valued for their small hands and manual dexterity leading to concerns about worker conditions.

More work to be done.

Creation of new jobs seems to be a matter of faith. IT in businesses drove an increased amount of management, the move online drove a need for webmasters, web designers and online marketers. There is an assumption that over time AI will have a similar effect, beyond people who can write prompts.

Limiting factors

Technology

While GPT based models have surprised both in terms of what they can do and fail to do, there is a belief amongst experts that:

Data sets will only get you so far. There is no clear path to a new technique, or what older techniques would need to be combined with GPT-based systems. Of the data sets out there, a significant minority could be filled with ‘poison’ data like nightshade.

That there isn’t enough data to train models in a lot of cases and synthetic data is often used instead. Others believe that this will corrupt and stunt future AI models rather than help them.

Resources

AI systems like crypto mining consume a lot of energy and require a lot of water for cooling which is already straining data centres and infrastructure. All of which will impact corporates ESG profile and larger investor relations health. You could have an amazing AI model, but if you have as bad an ESG rating as Exxon you willl struggle to raise funds.