Generative Text is the Procurement Junior Staffer You Didn’t Know You Needed
Procurement is the first mile of the supply chain. Already, before the Pandemic, procurement teams were spread too thin. When Covid hit with its full force, these employees faced tremendous additional stressors, challenges that persist in many ways. It’s no wonder that there is burnout. 89% of survey respondents in this Keelvar survey say that they are “banking on automation to reduce time spent on manual tasks.”
The arrival of “generative AI” is timely with its suggestion of the possibility of a liberating quantum leap forward after the incremental progress of robotic process automation.
Generative AI has much to offer procurement, but it’s important to understand what it can and cannot do. If we use it properly, cognizant of its strengths and weaknesses, artificial intelligence tools can improve our performance. While they’re at it, generative text applications might be able to improve the quality of our jobs, too.
The best way to think about AI tools such as ChatGPT in a procurement context is as an inexpensive, young associate: naïve, inexperienced, passably good at writing, hardworking, and willing to learn. Use them to make the most of the valuable time (and mental wellbeing) of your senior staff. These experienced hands can focus on developing supplier relationships, evaluating performance, training people, and working strategically with the business. Good AI augments staff capacity; it does not replace staff capacity.
This article dives into what the new technologies do and the conditions under which they can be most useful for procurement staff. We’ll end with a detailed example of how we might use these solutions (in combination with other technology) to write a first draft of a Food Services RFP.
What Is “Generative Text?”
Generative text refers to the technology underpinning artificial intelligence tools that can write blocks of text. Initially, researchers focused on the problem of predicting the next word in a sentence or a fragment. Imagine the word suggestion feature in a messaging app or Gmail. You can see here that the email application predicts I want to type in the words “next week” once I type the letter “n” after having written already “Let’s figure out a time to meet …”
Developers teach these models how to predict the next word most accurately by collecting massive amounts of data, preprocessing them so that the computer can read the text numerically, and training different combinations of AI building blocks to learn relationships between words. These deep learning models are evolving to become increasingly sophisticated by combining the building blocks in different sequences (and adding more of them), even as they train on bigger and bigger data sets. Models can have hundreds of millions or billions of parameters. It can be extraordinarily difficult to explain how they reach their conclusions.
Having created these behemoth machines to work with text generally, developers have introduced new so-called natural language tasks.
Generating text that sounds plausible is one of the most exciting jobs we have built models to execute. The progress we have made is akin to predicting large blocks of text, one next word at a time, iteratively, with an increasing sense of nuance for not only the relationship of words to one another generally, but to one another in a specific sentence.
There are thousands of different models out there. ChatGPT is a portal into only one of them. It is a ready user interface for the underlying GPT-3+ model from OpenAI. All the large technology companies have these kinds of models. We’re on the verge of seeing an explosion of ChatGPT-like applications that are easy for laypeople to use. Up until this point, the models have been usable but accessible only with subsequent coding to extract and manipulate the results.
The promise of ChatGPT and other tools like it is the democratization of this powerful technology.
We are moving from the equivalent of bespoke early automobiles, hand-crafted by specialists and afficionados in their garages, to the rollout of mass-model cars such as the Ford Model-T built for widespread use. Like the transition from the horse-and-buggy, there will be bumps along the way, but the evolution is inevitable.
One of the key advances in recent years has been the evolution of models called “Transformers.” They have progressed to be good at understanding the context of words within a sentence. This improved their performance in next word generation substantially. Yet the quality of their presentation obscures their limitations. Here is the MIT Technology Review:
“They are excellent at predicting the next word in a sentence, but they have no knowledge of what the sentence actually means.”
The world is complicated and messy. Yann LeCun, one of the top AI researchers in the world, says of large language models like GPT3 (the model underlying ChatGPT):
“Why do LLMs appear much better at generating code than generating general text? Because, unlike the real world, the universe that a program manipulates (the state of the variables) is limited, discrete, deterministic, and fully observable. The real world is none of that.”
The output of these models comes across as sensible, but that doesn’t mean what we obtain is reasonable.
“Gary Marcus, author of Rebooting AI, explained on Ezra Klein’s podcast: ‘Everything it produces sounds plausible because it’s all derived from things that humans have said. But it doesn’t always know the connections between the things that it’s putting together.’”
These models may not generate text per se; they may plagiarize text from their training data sets.
“Past literature has illustrated that language models do not fully understand the context and sensitivity of text and can sometimes memorize phrases or sentences present in their training sets.”
Another way to think about generative text is that it is like a modern photocopier. Contemporary digital copiers encode an image into a compressed, digital file and then decode the digital file to print it. There is some slippage ‘twixt the cup and the lip if it is not a perfect compression. A lossy compression might make sense if you wanted to make it easier or less expensive.
“Think of ChatGPT as a blurry JPEG of all the text on the Web. It retains much of the information on the Web, in the same way that a JPEG retains much of the information of a higher-resolution image, but if you’re looking for an exact sequence of bits, you won’t find it; all you will ever get is an approximation. But, because the approximation is in the form of grammatical text, which ChatGPT excels at creating, it’s usually acceptable. You’re still looking at a blurry JPEG, but the blurriness occurs in a way that doesn’t make the picture as a whole look less sharp.”
What Does Generative Text Mean for Procurement?
Put this all together and it means that AI doesn’t eliminate work; it changes work.
The machine lays out the building blocks for a project and the human, with the knowledge of context and complexity, ensures that the building blocks fit together as intended. The AI does the brute set-up work; the individual brings the meaning. This is where the value lies.
“What makes businesses highly profitable is using humans to do these higher order things, these creative strategic thinking things that are still a long way from anything I’ve seen the computers able to do.”
AI tools such as ChatGPT (laughingly referred to as “mansplaining as a service”) deliver their output with what appears to be complete certainty wrapped in a reassuring grammatical correctness that together make them seem more credible. But as familiarity with these instruments becomes more widespread, we’re beginning to realize that we can’t trust the answers that they give us blindly. We need to verify, too.
“… it has a tendency to confidently deliver incorrect information. This means that step one in making this technology mainstream is building it, and step two is minimizing the variety and number of mistakes it inevitably makes …
“… Dr. Dai says that one analogy for the future of trust in AI systems could be one of the least algorithmically generated sites on the internet: Wikipedia. While the entirely human-written and human-edited encyclopedia isn’t as trustworthy as primary source material, its users generally know that and find it useful anyway. Wikipedia shows that ‘social solutions’ to problems like trust in the output of an algorithm — or trust in the output of human Wikipedia editors — are possible.”
Over time, armies of laborers may end up tweaking these products to avoid the myriad errors they generate today. For now, we can use generative text to get us started on projects, doing the first cut of the tedious work, freeing us up to work on the meaningful work. Who wouldn’t want that?
AI can make our jobs better by bearing the brunt of those things we don’t like to do, so we can focus on making a real impact.
One other key potential benefit is that tools such as these have the potential to turn anyone into a procurement officer. Instead of being an arcane process that other parts of the organization tolerate, it can give basic purchasing knowledge to people on the front line. This will help everyone understand the important benefits procurement brings.
We can obtain value from generative text tools if we remember that they are not foolproof, despite an appealing presentation. Once we realize their limitations, then we can use them responsibly. They will be the plough horse in the fields of digital transformation, emancipating manual labor to focus on adding greater intellectual and commercial value. It just may be the avenue for the long-delayed, organization-wide recognition of procurement as a strategically vital business process.
It’s already being tested. The US Department of Defense has something they call “Acqbot.” It doesn’t make any contracting decisions, but it does generate text. Ideally, this would make it easier for suppliers to respond, as well.
It makes sense that the largest organizations would move first. They have the largest base over which to spread the costs.
But there is nothing to stop even the smallest organizations from using these tools.
Consider an Example: a RFP for Food Services
Perhaps the best way to understand what is happening here is to look at an example. Consider a company that is looking for a supplier to provide food services.
Here’s what we get when we “prompt” ChatGPT with the following question: “show me a detailed food services statement of work for an rfp.”
Sure, here is an example of a detailed statement of work for a food services RFP:
Overview: The purpose of this RFP is to solicit proposals for food services for a new corporate office building. The food services should provide a high-quality dining experience to employees, guests, and clients.
Scope of Work: The selected vendor will be responsible for providing food services for the following areas of the building:
Cafeteria: The cafeteria should offer a variety of healthy, fresh, and locally-sourced breakfast, lunch, and snack options to employees and visitors. The vendor will be responsible for menu planning, food preparation, service, and cleaning.
Coffee Bar: The coffee bar should offer a variety of hot and cold beverages, including coffee, tea, and specialty drinks, as well as baked goods and light snacks. The vendor will be responsible for menu planning, food preparation, service, and cleaning.
Catering: The vendor will be responsible for providing catering services for meetings, events, and conferences hosted in the building. The catering services should include a variety of menu options, professional service, and on-time delivery.
Vendor Requirements: Vendors interested in submitting a proposal for food services must meet the following requirements:
Experience: The vendor must have at least 5 years of experience in providing food services to large corporate offices.
Quality: The vendor must have a proven track record of providing high-quality food services, including menu planning, food preparation, service, and cleaning.
Sustainability: The vendor must demonstrate a commitment to sustainability, including the use of locally-sourced and organic ingredients, waste reduction, and recycling.
Staffing: The vendor must have a sufficient number of qualified staff, including chefs, servers, and catering personnel, to provide timely and professional service.
Compliance: The vendor must comply with all applicable health and safety regulations, including food handling, storage, and preparation.
Proposal Requirements: Proposals must include the following information:
Executive Summary: A brief summary of the vendor’s experience, capabilities, and approach to food services.
Scope of Work: A detailed description of the vendor’s proposed services, including menu options, staffing, and hours of operation.
Pricing: A complete pricing schedule for all food services, including cafeteria, coffee bar, and catering.
References: A list of at least three references from current or previous clients.
Certifications: Proof of compliance with all applicable health and safety regulations.
Sustainability: A description of the vendor’s commitment to sustainability, including the use of locally-sourced and organic ingredients, waste reduction, and recycling.
Timeline: The following timeline applies to this RFP:
RFP Issued: February 23, 2023
Proposal Deadline: March 23, 2023
Vendor Presentations: April 2–6, 2023
Vendor Selection: April 13, 2023
Contract Start Date: May 1, 2023
Evaluation Criteria: Proposals will be evaluated based on the following criteria:
Submission Instructions: Proposals must be submitted by email to the contact person listed below:
Contact Person: John Smith Title: Director of Facilities Management Email: firstname.lastname@example.org
Please include “Food Services RFP” in the subject line of the email.
Thank you for considering this RFP. We look forward to receiving your proposal.
We can see several things straightaway.
One, at less than two pages, it is much shorter than we would have anticipated.
Two, this is because it lacks specificity. There are very few questions related to the buyer’s requirements. There is no description of the buyer’s situation. In fact, it assumes that this is for an office building.
Really, it is an outline of what a general bid solicitation might look like. It asks the supplier to fill in the blanks. The best projects tell the supplier’s as much as possible about the buyer’s constraints. This is so open-ended as to discourage suppliers from investing in a response.
Of course, we can prompt ChatGPT for more detail. We did that by asking, “Give me more detailed questions.”
Sure, here are some more detailed questions that you can include in an RFP for food services:
Menu Planning: a. How will you develop and revise the menu to ensure it meets the nutritional needs and preferences of our employees, guests, and clients? b. How do you source your ingredients, and what steps do you take to ensure the quality and safety of your ingredients?
Food Preparation: a. What measures do you have in place to ensure food safety during preparation and service, including temperature control, hygiene, and cross-contamination prevention? b. How do you accommodate special dietary requirements, such as allergies, intolerances, and cultural preferences?
Service: a. How do you ensure timely and efficient service, especially during peak hours? b. What steps do you take to ensure that the dining area and service stations are kept clean and tidy throughout the day?
Catering: a. What types of events do you specialize in, and what is your capacity for catering large events? b. How do you handle last-minute changes to catering orders, and what is your policy for cancellations and refunds?
Sustainability: a. How do you source and dispose of food waste, and what steps do you take to minimize food waste? b. How do you incorporate sustainability practices, such as using eco-friendly packaging and reducing energy and water usage, into your operations?
Staffing: a. How do you ensure that your staff are properly trained and qualified to perform their duties, including food preparation, service, and cleaning? b. How do you manage staffing levels to ensure that the dining area is adequately staffed at all times?
Technology: a. What technology solutions do you use to streamline your operations, such as online ordering, inventory management, or kitchen automation systems? b. How do you use data and analytics to optimize your menu, staffing, and operations?
Pricing: a. How do you structure your pricing, and what factors influence the cost of your services? b. What is your policy for pricing adjustments and how frequently do you review your pricing structure?
Contract Terms: a. What is the minimum term for your contract, and what is your policy for early termination? b. What is your policy for price adjustments, and how frequently do you review your pricing structure?
These questions should give you a more comprehensive view of a vendor’s capabilities, experience, and approach to food services.
This is a page in length. Much of it consists of general questions related to contracting with no reference to the specifics of a food services project.
ChatGPT does not have sufficient category-specific knowledge to write a complete first draft.
The ideal application would be to give ChatGPT to someone who knew the category well but didn’t know how to write clearly.
What if we needed category-specific questions, too?
Ideally, we would like to have specific information about multiple sub-topics within the category, such as the following:
Maintain control receive and store food supplies properly ensuring storage and first in first out basis with minimum wastage
Meat must be from federally inspected source and Canadian Food Inspection Agency CFIA approved or equivalent
Only products that meet the standards of the food industry and of the Canadian General Standards Board may be used
Foodstuffs must not be spoiled dirty or infested at the time of delivery
Frozen food must show no signs of defrosting Perishable food must be delivered in refrigerated vehicles in which the temperature does not exceed 4C
To do something with this level of detail, repeated across various sub-topics, requires access to relevant data.
Either the buyer would have a way of cataloguing and sharing structured data about prior work in the category or they would have access to services and consulting firms that had this kind of intelligence themselves.
Once a buyer had the data, it should be possible to fine tune the same sorts of tools underpinning the ChatGPT interfaces of the world to extract, curate, and present the detailed, myriad questions a buyer would need to ask.
Being more specific actually makes it easier to attract quality supplier proposals. It demonstrates that the buyer understands the problem they seek to solve. The detail limits the problem space to something familiar and straightforward to address.
This second piece is what we have built at EdgeworthBox. You can think of it as a limited procurement-as-a-service functionality. We have hundreds of thousands of statements of work, spanning multiple categories, in our structured data, made accessible by the standardization of its formatting.
Combine this service with the writing of a ChatGPT interface and the procurement cycle becomes much shorter and buyers attract more (and better) supplier responses.
We’d love to talk to you about it. Reach out anytime. EdgeworthBox is a set of tools, data, community, and market intelligence that help B2B buyers purchase the right solution, from the right supplier, at the right price.