🎣 The Catchup

Dear readers,

Everything aches tonight…

In June, I had promised a gym buddy I would team up with him for a “friendly” CrossFit competition in the fall. “Sure, why not? Could be fun…” I was reminded two weeks ago that this commitment was coming up…

“Oh no… I have not done the work this summer…Yikes…”

It was still great fun. I didn’t do anything silly to hurt myself. However, everything except my fingers hurts 24 hours later… Wish me luck to get out of bed tomorrow… 😅

…

Tonight, I wanted to share 2 issues that will prevent Generative AI from reaching its full potential in the enterprise space… I’m sure solutions will quickly come to the fore but, so far, I haven’t found satisfactory answers/solutions to these questions.

I don’t consider myself an AI expert… Far from it. However, as a “Procurement processes and systems guy”, I wanted to share these with you so that:

1. As a key stakeholder in Generative AI technology Procurement in your company, you can screen technologies for these issues and avoid taking on the associated risks if vendor answers are unsatisfactory.

2. If you have the answer, you’ll tell me I’m wrong and I can forward it onto the rest of the Pure Procurement community 😅

Have a great week ahead.

Best,

Joël

P.S. In case you missed it, here was my best LinkedIn post of the week (it goes out to all the Procurement and Lord of the Rings fans…):

Standing up a new centralized Procurement department is a tough gig...

🌙 Sunday Night Note

Since Chat GPT launched November 30th, 2022, we haven’t stopped hearing about all the wonders Generative AI and Large Language Models (LLM) are unleashing into the world…

For me, the first 3-6 months was all about understanding the ‘art of the possible’ with Generative AI and LLMs. The question that always came up in conversations with clients was: “What are the use cases for Procurement?”.

Daniel Barnes wrote a great post this week on this exact topic, laying out many of the Procurement use cases Generative AI can help optimize. He touched on 3 main categories:

1. Accelerated document creation (drafting)

2. Rules-based workflows/processes (vs step-by-step process design)

3. AI agents/assistants (Q&A and knowledge retrieval)

The outputs in each of these categories becomes much more valuable and potent if it can be adapted to your specific company’s context:

• Drafting. A generic draft for an RFP is great but a context-specific RFP draft that considers the nuances of how you run RFPs in your company is better.

• Generating Business Rules/Processes. Generating a generic set of “best practice” rules for invoice management to guide users is great but generating a set of rules that considers the specific configurations of your AP system is better.

• Querying an AI Assistant. Having an AI agent return the general “best practices” around the use of framework orders is great but getting answers based on your internal Procurement policy rules is better.

An LLM/Gen AI tool + your contextual data/documentation = Much more potent outcomes. This was made clear to me when I saw the recent announcement on OpenAI Dev Day where they ran a demo of their “no code” GPT builder. If you feed your data to the model, you get better outcomes. This is required to take Gen AI tools to the next level of usefulness.

Now for the issues…

No. 1 - No Clear Way to Protect Intellectual Property (IP) from the Outside

The standard ‘marching order’ in Enterprise right now is: “DO NOT input any personal, sensitive or confidential information into an LLM/Gen AI tool.” However, to get contextual information back from an LLM/GenAI tool, you necessarily need to feed it privileged data and information about your company.

For structured data (anything you maintain in a dedicated field in a database, like a vendor legal name), it seems that the answer may lie in Sensitive/Confidential Data Vaults. Essentially, you can randomize information fed to the model, get contextual but randomized information back and “translate it” live for the end user so they see the real value. Sort of how password or credit card information storage works (I’m sure this is wrong on a technical level but it helps illustrate what I mean…). So that’s promising…

However, for unstructured data (documents like process models, standard operating procedures, policies, procedures, etc.), I don’t see how you can share this information with a model without “releasing it into the world” by making it part of the overall public model everyone else is using.

It’s like saying you’re going to add water to a pool but that you want to be able to keep it separate from the other water in the pool but also part of the water in the pool…

You could buy your own pool but that will be cost prohibitive for many…

Until there’s a good answer to the following question, you have to be weary of trying to achieve contextualization with an LLM/Gen AI tool in a company:

No. 2 - No Clear Way to Manage Permissions to Data

The second issue still relates to data. However, the focus is on figuring out how you can partition access to data according to the job roles and responsibilities at your company.

Let’s use an example to illustrate.

You want to build an LLM/Gen AI chatbot for suppliers. You want them to chat with the model instead of calling your contact center for all their different queries on invoice statuses, payment dates, how to update their data, how to interact with your systems, etc. Wouldn’t that be wonderful!

Well, if this is the same tool/model as you are using internally with your employees, you don’t vendors to be able to receive answers that contains information outside the scope of what they should know. For example, you don’t want vendors to know the amount under which invoice price changes are automatically accepted by your system for fear of fraud.

Similarly, you could want to partition information this way for different employee groups within the company as well.

To go back to my pool example, how can I get my model to ensure that swimmers can only swim in certain portions of the pool with pre-designated water? Tough ask…

Here’s the second question that requires a good answer before I advise trying to achieve contextualization with an LLM/Gen AI tool in Enterprise:

Other Requirements for Good Contextualization Outcomes with Gen AI

As we await the answers to these questions from the market, you can still work on your pre-requisites…

• High quality master data

• High quality documented policies, processes and procedures (business rules)

• High quality role-based security roles (and documentation)

If you have:

• Crappy data

• Business rules that aren’t coherent with each other

• Unclear job roles and associated system security

…then it doesn’t matter if you have the answers to the above questions… You won’t be able to get very far with AI contextualization anyways…

Work on your fundamentals in the meantime. It will pay dividends.

💭 Quote of the Week

🎁 Offer of the Week

If you are looking to jump start your Digital Procurement literacy journey, my introductory eCourse is 25% off for readers in the next week.

In this 2-hour video and text self-paced eCourse, you’ll learn:

• How to logically divide the end-to-end Procurement process when thinking about implementing systems

• The different types of Procurement systems on the market; their strengths and weaknesses

• How to approach Procurement system selection and implementation to ensure you get a system that works for you, and not the other way around.

Use code NOV25 at checkout to get 25% off. OR, even better: subscribe to the premium version of the newsletter and get the course for free (coupon code provided in your welcome email) along with all the other perks.

You have an offer for Pure Procurement readers? Reply to this email. Let’s see if there’s a fit.

💬 Help Spread the Word

If you like this newsletter, please share it with another Procurement professional in your network. The email writes itself:

“Just read this. Thought of you. Think you’d enjoy it.”