How to Plan with Generative AI
Ever since the Mayor of my little town told me he wanted GenAI to tell him what he was missing, I have been on a mission to figure out how organizations can best use GenAI in planning, including answering that question. Custom GPTs enabled a very productive way to accomplish those goals, which I have developed with lots of experimentation and testing and will describe here. The facts that OpenAI has delayed roll-out plans for the ecosystem of GPTs since the Coup and that last week we entered the Gemini Era only makes the telling of this tale more important; developers at Google (and Anthropic and elsewhere) need to see how useful the custom GPTs are for the high-level and strategic uses described below, and even OpenAI’s COO appears to need a pep talk on big-picture uses.
Plausible, Unforeseen Futures, Not THE Future, and No, You Cannot Just “Talk to Your Data”
Some foolishly want AI to predict the future, so let’s stipulate first that no AI will ever predict the future as well as Daniel Solove. GenAI doesn’t even try to predict the future of anything except the sentence or image it is completing, but by drawing on lots of data and disciplines with which you may not be familiar, it may raise possibilities of multiple plausible futures of which you were previously not aware, which may inspire some reexamination about where you think you’re going. And that is just the beginning of the conversation.
GenAIs are not just bad at predicting the future; they are not even reliable spokespersons for your data today. Many organizations believe consultants who promise to train enterprise data to talk with GenAI, but as Ethan Mollick recently pointed out, that too is generally hallucination.
The following process avoids such golden calves.
1. Start with Your Favorite GenAI-Driven Search
In the pre-Gemini era, I have generally started planning exercises by casting a wide net with a specialized GPT focusing on extensive, slow, GenAI-controlled investigative search in the domain in question. When OpenAI’s version first came out, I knew it would change the search world with its (1) less-word-dependent, natural language processing (NLP) semantic investigations and (2) direct, real-time connections to websites rather than through indices, notwithstanding its early operational glitches. Since then, it was taken away for a while, but upon returning it has been much more reliable, showing its work as it browses and including links in its responses. Bing’s GenAI-first entry point in “more creative” mode is a good, faster alternative with better footnoting if you don’t need the fine-tuning of a specialized GPT, and Bard with Gemini Pro looks based on my initial test drives like it can blow both Bing Chat and OpenAI native web-browsing chat away, in finally matching an LLM in the same league as GPT-4 with Google’s superior search. My best advice to Google and you is that it’s not all about blinding speed when the search really matters; (1) I hope it will be possible to fine-tune Gemini as one can a GPT before the query, and (2) I hope Gemini-Google will take the time to hunt through the web in real time using its multi-modal (not just NLP) semantic superpowers rather than fetching material from a pre-existing index.
A cautionary note on confidentiality: I have carefully avoided entering confidential or sensitive information into GPTs, and will continue to do so even though GPT Builder now offers you the option of turning off contributions to training data; OpenAI to my knowledge offers no opt-out from exfiltration and harm/abuse monitoring for 30 days as is offered by Microsoft Azure. Bard with Gemini Pro just comes right out and tells you that it will suck your data right up.
2. Scenario Planning with Your Favorite Ecumenical Futures Studies Methodology
I provide the learnings from the search results into one of the GPTs built to start the planning conversations with the alternative plausible scenarios that are fortunately plentiful in futures studies. My favorite given its simplicity and inclusiveness is still Jim Dator’s Four Futures approach I learned from David Giguere, GISP, which I wrote about here and built into one of my first GPTs, Manoa Ideator. Another appropriately non-fundamentalist approach, particularly well-named given the workings of GenAI as plausibility engines, is the “cone of plausibility” analysis taught to me by I. David Daniels, PhD, CSD, VPS. Whatever framework you prefer, what matters most is that it opens up a number of possibilities and opens them up as conversation starters rather than as conclusions. Now on to the conversation.
3. In-Depth, Multi-Factorial Discussion
Specialized GPTs make their greatest contributions to strategizing and planning here. It is as if you can assemble exactly the diverse team of experts that you want, each of which has the professional background and approach to research, inference, analysis and communication that you tell it to have, and each of which can draw on up to 20 (if that is still the limit) knowledge documents plus websites specific to it as well as the NLP-driven web search powers discussed above. Decades ago when I was a Big Four consulting practice leader I used to love what a great team of diverse professionals could do; now it is as if any of us can create and deploy such a team to serve an organization of any size or an individual in just a few hours, and for $20 a month.
Let me offer a quick example. My little town has preserved its woods and fields through then-unusual zoning approaches to which everyone managed to agree two decades ago, including Transferable Development Rights (TDRs). TDRs enable property owners in areas designated for conservation (sending areas) to sell development rights to developers in areas slated for growth (receiving areas), after which the sending area is conserved. The natural land has been protected since then through other aspects of the zoning regulations, but the market for TDRs never heated up, in part because the development pressures have been more limited until now. In a contentious City Council election, a slate of candidates was recruited who generally took the position, apparently without doing much research, that TDRs will not work. Attending candidate forums, I was struck by the extent to which neither the opponents nor the supporters of TDRs understood the mechanism well enough; the need for education and deeper thinking seems apparent lest we throw the baby out with the bathwater or risk its safety in the tub.
Now that the town’s zoning has survived the election, and to take advantage of what I hope is residual interest around the issue, I just created for my community a GPT stocked full of data and studies on the hundreds of successes and failures of both TDRs and the more targeted Transferable Development Credits (TDCs) around the world, Global TDR/TDC Planner. It has a lot of focused guidance to offer regarding all of the assertions I heard about flaws in the program, and offers a much richer array of possible solutions than we have contemplated. Again, I am not saying that it has THE answers, but if we engage with it we will know more than we can learn from a few consultants.
Engaging with it, I quickly came to see TDRs and particularly TDCs as mechanisms that can support a multitude of different policy goals, and could examine the track records of many such attempts. For example, the TDC program of Boulder, CO addresses the important issue of limiting or at least moderating house sizes. Diving in, I created and stocked a GPT focused on incentives for reducing housing sizes, Eco Housing Planner. The hope is that each GPT will help our community and others with critical perennial or recurring issues, and they may be brought together, including as conflicting voices, to address the more complex issues.
Once we can get past the confidentiality issues, I believe that such ongoing sources of information and preliminary guidance will be preferred by our clients in many cases to the static guidance now generally offered by advisory consultants and lawyers, which is one reason I piloted and announced AI Governance Ideator.
4. Rapid Generation of Draft Policy Proposals, Bills and Regulations
When planners are satisfied with their choices supported by their dialogues with their specialized “expert” GPTs and one another, GPTs specialized at producing draft policy proposals or draft bills, regulations or other authorities can quickly turn their ideas into draft language. This is a particularly valuable capability for those of you in government, because as you know if you don’t grab the pen first, a lobbyist will. I use BOTH GPTs and Anthropic ‘s Claude 2 for drafting statutes and regulations. GPTs involve greater customization for specialized drafting and Claude, perhaps given its training as a constitutional lawyer, is generally the better statutory drafter. My normal current process has been to start with a specialized drafting GPT and let Claude critique the work before I make my own changes.
