AI Digest - May 24
What matters to organisational leaders each month
As I hit send on this newsletter speculation is swirling about a General Election announcement in the UK - another domain where shifts can happen suddenly and with unpredictable consequences. I’ll keep the political analysis to the pub though - all the uncertainty & strategic responses discussed below are AI related!
I hope you’re enjoying the gradual arrival of summer (if you too are London based)
James
P.S. Thanks to the half dozen or so people who have shared their experiences navigating procurement processes to buy access to AI tools and services. I am still keen to hear from anyone else with thoughts in this space.
If you only read one thing
What does ChatGPT currently say about your brand? GenAI is changing elements of customer relationships and the most forward thinking companies are recognising this. In some cases, control of the customer journey – a key element of winning repeat business – risks being taken out of their hands as new technology provides new layers of intermediation. If you don’t control this layer you risk losing lots of the key touch points that shape how a customer feels about you. Do you want your product to be recommended by an AI agent that sometimes hallucinates? Do you want customers to use ChatGPT rather than calling customer services? Maybe not.
This challenge is prompting a new wave of thinking and effort within companies to ensure they continue to actively shape these interactions, rather than letting others do so: see Apple fighting for control over the whole user interface in cars that include their software.
Powerful consumer brands (and the best B2B service providers) recognise that they don’t just provide a solution to a need, but a range of other ‘elements of value’ alongside this (see image below) and this is embodied in how customers interact with you, not simply what they get. The Big 4 sell confidence, not strategies; Apple sells an ecosystem, not a smartphone; Coca-cola sells… something, but it’s not just fizzy soda. You need to think about how your customer journey and relationships will change as a result of their changing use of technology. Or others will change them for you and you’ll end up like newspapers who relied on social media for distribution - until they couldn’t. The Elements of Value. Disruptive Interfaces. Apple Car Control.
Contents
If You Only Read One Thing
What are you selling? Who owns your customer relationships?
What Is GenAI Good For?
Case studies, use cases and examples you can lift and shift
Simulations and Surveys
Advertising takes another leap forward
How To Successfully Integrate GenAI With Existing Organisations
Removing constraints, raising the floor not the ceiling
Finding valuable data that you already produce
Prepare for failure - and estimate the cost
Our Recent Work
AI in Job Applications
Zooming Out
Emotional understanding
Private Equity targets
Global Talent Flows
Learning More
Public Service Announcement: Slack is training on your data
Second Order Effects - Worked example for the future of Marketing
Job-specific benchmarks
The Lighter Side
What Is GenAI Good For?
Month on month it feels like there are more examples to point to! Here are some of the best ones to help you brainstorm what you could do, from the mundane to the ridiculous.
101 real world GenAI use cases from leading organisations. Some of these are very high level and don’t actually give you any understanding of what is being done. Others are very clear. Either way, the list is useful for giving a sense of the scale of change, underlining the extent to which companies want to put their names to this, and as a prompt for brainstorming. 101 Use Cases.
Uses of LLMs for educators: simulations, mentoring, coaching and co-creation. With broader implications for ways that LLMs expand capabilities. Innovation Through Prompting.
AI products people would pay for that could exist today but don’t exist yet. Some more useful than others. Nat Friedman.
Simulations and surveys. We said in March that one of the best uses of GenAI in product development was using GPT personas to have a digital product manager or user voice in the room for every decision,. Examples are also now emerging of GenAI being used to run larger simulations of environments, such as a hospital, where agents learn over time how to effectively treat LLM-generated patients – with potential implications for product development and marketing. Some are also using GenAI to bring real people into the room more cheaply, conducting surveys and focus groups run by AI, with positive results. Agent Hospital. Focal Data. What Conservative Voters Think. Survey Data. AI Interviews.
Image and text generation for advertising. Meta have launched enhanced GenAI features for advertising, including automated variations on marketing images with text overlay, and on the advert heading and primary text. While still early days, this is big news for big businesses doing lots of targeting variations as well as for small businesses with no creative team. Meta. Advertising Driving Differences in Firm Performance.
How To Successfully Integrate GenAI With Existing Organisations
Use AI for what you are weakest at. This applies at both the individual and organisational level. There’s always the temptation to see if ChatGPT can do what you do best. But its power really lies in enabling individuals or teams to overcome constraints that derive from weaknesses in their own capabilities: bringing data scientists up-to-speed on the field they’re working in so they can communicate better with subject-matter experts, supporting nurses to do more of the work currently done by doctors to mitigate bottlenecks, helping companies with limited marketing skills catch up to their competitors. With GPT4 now free, and GPT-4o integrating into any work that someone might be doing on a Mac (with Windows presumably coming soon), there are endless opportunities to bring GenAI into the mix to plug skills gaps. Worker Augmentation. GPT-4o.
Build tools with data that is primarily yours. The corollary to last month’s advice to build only what is unique to you and let the market solve general problems is this: you need to identify what data is proprietary. Perfect timing for this handy guide which identifies 6 types of information you (likely) already produce and how you might incorporate it into LLMs. By which we mean give it to ChatGPT as context at the start of a conversation or use it to set up a GPT).
Training & Instruction manuals: Automate manual tasks where the process is already written down. GPTs will do a good job of following.
Frameworks, Templates & Outlines: Teach models your preferred methodologies so they respond with answers that fit with how you already communicate.
Example Outputs: Like new employees, AI models learn from gold standards and are good at copying “what good looks like”
FAQs: Interaction is a better way of retrieving information than reading a wall of text. Remember the guardrails if you release it to customers though.
Transcripts: Contain lots of information that is never read again - although these might not be information dense enough to use often.
Context: Instead you can extract important background information from transcripts (step 1) and ensure that all your GPTs have this as background so they know company priorities, organisational changes etc
Using your data guide. Moderna-OpenAI GPT Partnership.
First you fail, then you fail to realise it. Automated Customer Service presents big measurement problems. How do you know if you have misunderstood a customer's request? Most Phone Tree systems have no way for a user to say “my option isn’t listed” and the trope of someone yelling “I want to speak to a real person” is well worn for a reason. Customer service AI-agents (which are taking off) open up this problem to an even wider audience. Keep in mind that there are two risks to automating customer service: firstly you fail, secondly you fail to realise that you’ve failed. What happens next depends on your relationship to your customers. Perhaps they show up to your business in other more costly ways (termed Failure Demand), or perhaps they stop showing up at all (see If You Only Read One Thing). Either way, measuring what you miss is an important metric and one that requires thought. Be sure to measure all the costs when assessing cost savings. Failure Demand. Gartner: Contact Centre as a Service. Challenges uncovering what customers think.
Our Recent Work
I spoke at the launch of Hays’ What Workers Want report, looking at the role of AI in job applications (both writing and screening). 49% of applicants said they were seeing more success when using AI tools to help them write. Report.
Zooming Out
Emotional understanding is the next theme of development for LLMs. Previous generations turned all speech into highly compressed representations of text, losing all emotional content along the way. New ways of training models (which give rise to multimodal models, like OpenAIs new GPT-4o and Google’s Gemini) allow them to comprehend more than just the linguistic content, they can see video and process audio (not just words) in near-real-time. This opens up possibilities to, for instance, interpret vocal modulation and intonation, or hand gestures and facial expressions on film. How many corporate trainings have you sat through which tell you that 75% of communication is non-verbal? The number doesn’t matter, but the principle does. Models which understand emotional state are going to be far more useful for roles such as supporting customer buying journeys. Detecting Human Emotion. GPT-4o launch. Was GPT-4o flirting? Hume.AI.
Related: Police using AI to write reports from Body Cams. Exercise to the reader: What else will these Body Cams automatically be scanning for? Announcement.
Private Equity buyouts target companies with high automation potential. There is nothing new under the sun. Where PE firms have previously championed outsourcing and “cost control”, this takes a new form in LLM-powered automated workflows (or, quite often, simply Robotic Process Automation masquerading as AI). New owners have power (whether we like it or not) to break social promises, such as laying off workers, which previous owners couldn’t. Andrew Ziperski spells out the thesis for all to see. Blog.
Follow the talent. Excellent visualisations of the flows of global AI workers. A reminder that the UK both punches above its weight, and is a very small part of global research. Marco Polo.
Learning More
Public Service Announcement: Slack (which really means Salesforce) will train on your corporate messages unless you invoke their opt-out. Lots of discussion about keeping proprietary data safe and many organisations now prevent use of ChatGPT outside of an enterprise plan. For organisations that use Slack, however, the message content will be a much richer source than anything being pasted into ChatGPT. And as ever, a reminder that all the big tech players are training models, AI doesn’t equal ChatGPT. Slack Discussion - Tweet Thread.
First order effects are relatively easy, forecasting how systems shift is harder. As the saying goes, if you understood how cars differed from horses you could have predicted the petrol station, but you likely wouldn’t have predicted supermarkets or Milton Keynes. Andrew Chen looks through this lens to ask how marketing will change. Start with the building blocks: (almost) infinite labour, content and internationalisation and build models from there. Well worth applying this line of thinking to individual business areas (or hitting reply to this email and asking us to help). Marketing: What Happens Next.
Independent benchmarks for job-specific tasks. So far covering: Corporate Finance, Contract Law (and some US specific tasks such as Income Tax prep). Vals.AI.
The Lighter Side
Do you want to (byte)Dance with me? What’s more fun than a party… a party that comes with a pre-brief on who to meet derived from a 2 day virtual party simulation using AI agents. My RSVP is in the post. Link.
Nobody reading this far needs to see this, but you all know someone who does.
