Over the last few months probably millions of words have been written about the impact of ChatGPT and other AIGC applications are going to have on the way we work. Every conceivable issue has been considered and argued over except one……Read on.
When Google started up operations in September 1998 Carl Shapiro and Hal Varian, both at the University of California, Berkeley, were waiting for their book Information Rules: A Strategic Guide to the Network Economy to be published by Harvard University Press in October. This was not only the first book to consider the dynamics of the information economy but also the last. There has been no book since with a similar scope despite the impact that information products and services have had on the global economy. You would be excused for saying that this book was pre-Google and pre-WWW but the insights they offer are still relevant today. Sadly, the book is no longer in print. At the core of the book was a discussion about the interrelationships between costs, prices and perceived (by the customer) values.
Jump forward to 2023, and not a word about AIGC costs, prices and perceived values!
As Yejin Choi recently pointed out in her superb TED lecture the scale of computation for LLMs is such that currently only a few tech companies can make the resources available. Water consumption alone is very significant at a time when the conservation of natural resources is a major concern. Sunyan has published a background paper on the costs of LLM creation and management which does suggest that costs will be coming down but will prices follow? OpenAI publishes a tariffs data sheet but translating that into enterprise budget format is very challenging. Things get much more complicated when you start to consider the use of AIGC applications inside the enterprise, perhaps embedded into existing applications.
Enterprise search pricing models are complex enough without adding in AIGC. The license fees are usually only a relatively small element of the cost compared with the ‘per query’ charges. It is unclear at present what the basis of the transaction charges will be – per prompt, per text download, per application (e.g. summarization versus machine translation versus information discovery) and then there is the periodic cost of updating of the LLM database. If the AIGC application is integrated under license into the ‘search desktop’ along with one or more search applications the transaction charges may not align with the enterprise search license structure. .
However, the technical costs may well be relatively light compared to the disruption costs and training costs you will have to accommodate. Getting the value out of AIGC is going to require a substantial rethink on process management (have you ever tried to reconfigure a BPM/PM application?), certainly a hit on productivity and a considerable amount of cross-the-enterprise training. Reading a recent analysis from the US Federal Trade Commission will give you a hint of the work that lies ahead.
The bottom line is how you are going to present the ‘value’ element into the investment plan for the business. This is where we come to three ‘productivity’ challenges.
- How does your organisation measure ‘productivity’ in a knowledge worker environment where the knowledge in a single email may be of more value to the business than a 200pp report?
- How far back do you have records of these measurements at an appropriate level of granularity to be able to show demonstrable improvements through the implementation of AIGC applications?
- Will your business case be that with the improvements in productivity (assuming you are able to definitively measure them) your organisation can increase its output and revenue or that you will be able to reduce your headcount and thereby your reduce your costs?
At present all the proposed benefits of AIGC take no account of the bottom line of the supplier or the user, a somewhat artificial situation which is going to have to change. There is no free lunch!
Martin White
2 May 2023