AI Research – A Cautionary Tale
The other month, a friend referred a business owner to me for some consulting. This business owner was trying to solve for high transaction costs among his merchant clients. He had done good research and thought he might have some possible solutions.
We met twice and I gave him some pro bono feedback on his ideas. The conversations were very constructive. I outlined some hurdles his solutions would need to overcome.
About a week later, I received a detailed email from him with another solution idea. My response to his email began with the following line:
I am afraid we are backing into a consulting engagement without the appropriate contract agreement in place. . .
Why would I say that when he had provided me a detailed solution description?
In our previous conversations, he had mentioned that he uses an AI engine for his research. Knowing that and having a better feel for his knowledge base from our first two conversations, I could see that this solution description carried a fair bit of AI output.
And that is how this conversation backed into a consulting engagement.
I have a several concerns with AI as a research tool:
Lack of source citation –To give quality feedback to this client, I would need to “reverse research” the solution document to find the unnamed sources of its information. What percentage of this information came from opinion documents and what percentage from technical documents? Without any citation footnotes, it could take hours to figure that out.
Each payment network is different - Payment networks have significant differences between them, even when they use the same ISO specifications. I have worked on enough cross-network gateway projects to know the complexities those differences cause. Any AI generated document that potentially mixes data from different networks runs a very real risk of containing fatal errors. As there were no source citations from the AI output, I would need to comb through the solution document looking for clues for the source network for each statement.
I don’t know everything – “Why didn’t you just read the document and comment back?” I learned many years ago to never respond to technical questions without researching first. Even if I have answered a question many times – and ESPECIALLY with those types of questions – I need to research first. Technology and procedures change. I need to check if either has changed since the last time I answered this question. As well, two questions can look the same on the surface but be different enough to require very different answers. To modify what I said in a previous blog, “An expert is someone who knows when to stop”, drop, and research.
So, am I saying, ‘don’t use AI’?
I am and have always been an ardent “data first” person. I have spent much of my career querying databases to gather data into useful information. However, for information to be useful, it needs to be linked to its source and have context. Without those anchors, information becomes meaningless, not even useful for trivia quizzes. Information source and context are necessary to build an accurate and meaningful story or solution from data.
AI engines have computing power, and their intelligent programmers create detailed queries. However:
Are those programmers subject matter experts on payments?
How can they assess the relative value of various information sources on payments?
Do their queries segregate between process and technology documentation from Mastercard, Visa, Discovery, CUP, Interac, etc. and NEVER cross compile between them?
Without citations on data sources, I must assume “no” as the answer to each of these questions.
I do appreciate the research done by this business owner. The potential solution showed creativity and out-of-box thinking.
His next responsible step would be to spend money (probably $30K or so) to have me research and vet the solution.
Alternately, he could go straight to implementing his solution unvetted and maybe find out, about $3 million down the road, that ‘A’ isn’t always ‘I’.