Category Archives: Technology

Big Data: Are You Still Doing it Wrong?

The only buzzword on par with big data is cloud. According to the converted, or should I say the diverted, better decision are made with better data, and the more data the merrier. This sounds good in theory, but most algorithms that predict demand, acquisition cost, projected sales prices, etc. are based on trends. But these days the average market life of a CPG product, especially in electronics or fashion, is six months or less, and the reality is that there just isn’t enough data to predict meaningful trends on. Moreover, in categories where the average lifespan is longer, you only need the data since the last supply/demand imbalance, global disruption, or global spike in demand as the data you need for the current trend before that is irrelevant … unless you are trying to predict a trend shift, in which case you need the data that falls an interval on each slide of the trend shift for the last n trends.

And if the price only changes weekly, you don’t need data daily. And if you are always buying from the same geography, dictated by the same market, you only need that market data. And if you are using “market data” but 90% of the market is buying through 6 GPOs, then you only need their data. In other words, you only need enough relevant data for accurate prediction. Which, in many cases, will just be a few hundred dat points, even if you have access to thousands (or tens of thousands or even hundreds of thousands).

In other words, big data does not mean good data, and the reality is that you rarely need big data.

But you know that AI doesn’t work without big data? Well, their are two fallacies here.

The first fallacy is that (real) AI exists. As I hoped would have been laid bare in our recent two-week series on Applied Indirection, the best that exists in our space is assisted intelligence (which does nothing without YOUR big brain behind it, and the most advanced technology out there is barely borderline augmented intelligence.

The second fallacy is that you need big data to get results from deep neural networks or other AI statistical or probabilistic machine learning technologies. You don’t … as long as you have selected the appropriate technology appropriately configured with a statistically relevant sample pool.

But here’s the kicker. You have to select the right technology, configure it right and give it the right training set … encoded the right way. Otherwise, it won’t learn anything and won’t do anything when applied. This requires a good understanding of what you’re dealing with, what you’re looking for, and how to process the data to extract, or at least bubble up, the most relevant features for the algorithms to work on. But if you don’t know how to do that, then, yes, you might need hundreds of thousands or millions of data elements and an oversized neural network or statistical classifier to identify all the potentially relevant features, analyze them in different ways, find the similarities that lead to the tightest, most differentiable clusters and adjust all the weights and settings to output that.

But then, as MIT recently published (E.g. MIT, Tech Review), and some of us have known for a long time, many of the nodes in that neural networks, calculations in the SVM, etc. are going to be of minimal, near zero, impact and up to 90% of the calculations are going to be pretty much unnecessary. [E.g. the doctor saw this when he was experimenting with neural networks in grad school over 20 years ago; but due to the lack of processing power (as well as before and after data sets to work on) then versus now it was a bit of trail and error to reduce network size]. In fact, as the MIT researchers found, you can remove most of these nodes, make minor adjustments to the other nodes and network, retrain the network, and get more or less equivalent results with a fraction of the calculations.

And if you can figure out precisely what those nodes are measuring and extract those features from the data before hand and create appropriately differentiated metadata fingerprints and feed those instead to a properly designed neural network or other multi-level classifier, not only can you get fantastic results with less calculation, but less data as well.

Great results come from great data that is smartly gathered, processed, and analyzed — not big data thrown into dumb algorithms where you hope for the best. So if you’re still pushing for bigger and bigger data to throw into bigger and bigger networks, you’re doing it wrong. That’s the wrong way to do it. And the only way you can call it AI is if you re-label AI to mean Anti-Intelligence.

AI: Applied Indirection in Contract (Lifecycle) Management

Continuing our expose of why you should think “Applied Indirection” and not “Any form of Intelligence” when you hear AI, because most solutions claiming to be AI are really just dumb systems with RPA (robotic process automation) and classic statistical models from the 90’s, we move onto Contract (Lifecycle) Management which, like analytics, is almost universally touted to have AI, even when there isn’t even a shred of anything that comes close.

This doesn’t meant that there aren’t vendors with true AI, especially when you classify it as Assisted Intelligence (and sometimes even Augmented Intelligence), in the space, just that, as the buzz-acronym reaches new heights, there will be many more vendors claiming AI than those that actually have AI and you will need to do your homework to find out which is which.

Example #1 of Applied Indirection in C(L)M: Auto-Renewal Detection & Management

Yes, evergreen contracts can be a big problem in Procurement when they have outsourced their usefulness, but detecting and managing these is not hard, and certainly doesn’t require any AI whatsoever. All you have to do is specify the contract as “evergreen” or “auto-renew” by checking a box and enter a notice-by date (to prevent an evergreen renewal” as well as the start date and end date and most contract management platforms can alert you in sufficient time to take action, escalate to your supervisor if you don’t, and kick-off a termination process at the push of a button.

For anything close to AI, you require a system that can detect when a contract is evergreen or auto-renewing when there isn’t a spelled out and easily identified auto-renewal clause that can be found with a simple reg-ex search. For example, when a crafty supplier buries an auto-renewal requirement in the liability section under the notices subsection titled “methods for delivering official notices” which starts off “Official notices shall be sent by X, Y, or Z, to A or B and only treated as an official notice upon proof of receipt. This includes a notice of non-renewal, as the contract will automatically renew 30 days prior to expiry otherwise.” Even a good lawyer might miss that in a fifty page contract when it’s snuck in on the third revision.

Example #2 of Applied Indirection in C(L)M: Off-Contract Purchasing

Maverick purchasing is a big problem. But it’s not one that you need AI to detect. If you encode all the products, services, and / or categories that should be bought on contract, it’s pretty easy to identify when a purchase for that product, service, or category is not bought from that supplier. And if the contract only applies to a region, it’s pretty easy to encode that too and it’s just a simple check.

And even if you have two or three suppliers in a multi-supplier contract for risk mitigation purposes, then it’s just a matter of making sure at least one of the supplier got the purchase, and if each supplier had a geographic area, that the right one for the area. Again, simple rule checks. No AI needed.

The key is to detect when something is off-contract when it is not specifically coded to a contract, either because it’s a new product, missing a category designation, required to hit a volume commitment, and so on. And while this can often be accomplished by identifying the closest product or service (using a document likeness statistic or weighted field match), sometimes advanced NLP may be employed for better results (and this would constitute weak AI).

Example #3 of Applied Indirection in C(L)M: Clause Suggestion

On the surface, this sounds pretty smart … point out clauses that should be in my contract to protect me. Under the hood, in most CLM systems that include authoring, it’s basically a set of templates that are used to specify what to look for in a contract type, with additions or subtractions for well defined industries that the provider serves. It’s basically a check list. And it’s about as dumb as it gets.

Can it be smarter? Of course, but the smarts are more around proper contract identification than clause selection. Because the clauses that should be included generally depend first and foremost on the type of contract, secondly on the product or service, and thirdly on the regulations that affect the products and services in the origin country, the destination countries, and any points in between. Then, identifying which regulations come into play and which types of clauses will be needed. This requires good NLP, probabilistic selection, and, preferably, adaptive learning that learns over time when Legal or Procurement chooses an alternate clause over a standard clause. A system should have assisted intelligence here to be useful, and augmented intelligence to be truly useful. But few do.

Note that SI is not saying that systems with the non-AI abilities discussed above are not valuable, as any system that automates tactical processes and minimizes non-strategic busy work is valuable. We are just saying you shouldn’t pay for what you’re not getting, or overpay for what you are. Buy what you need, and pay accordingly.

AI: Applied Indirection in Sourcing

As we said yesterday, hopefully we have made it clear by now that most of the time you hear AI you should think “Applied Indirection” and not “Any form of Intelligence” because most solutions claiming to be AI are really just dumb systems with RPA (robotic process automation) and classic statistical models from the 90’s (which were available in SAS in the 90’s as well, you just didn’t have enough memory on your PC to run all the data you wanted to run).

But since we want to make it abundantly clear that most of the “AI”, even in our space, is not “AI” at all, we are going to continue to take the major areas of SPT (Strategic Procurement Technology) and highlight some areas where AI is commonly claimed, but rarely found, continuing with Sourcing.

This doesn’t meant that there aren’t vendors with true AI, especially when you classify it as Assisted Intelligence (and sometimes even Augmented Intelligence), in the space, just that, as the buzz-acronym reaches new heights, there will be many more vendors claiming AI than those that actually have AI and you will need to do your homework to find out which is which.

Example #1 of Applied Indirection in Sourcing: Automated Auctions

Some of the shiny new sourcing platforms are really slick and can run an automated auction and do everything from the time you do a product/service selection all the way to final award recommendation. Now, I’m sure you’re thinking such a platform must be at least on the order of augmented intelligence, approaching cognitive, to do all this, but the reality is that you can do all this with simple RPA and a rules-driven workflow. Supplier selection? Just select the past suppliers and pre-approved suppliers from the last sourcing event. RFIs? Use the template with the standard terms and conditions. Ceiling prices? Use current price or current market price if the price has been relatively flat for the past year or falling. Floor prices? Pull in the should-cost model, marked up by a fair margin, if it is available, or use the lowest of the lowest paid historical price and lowest market advertised minus the typical savings percentage in the category. Minimum Decrement? 1%, rounded to three significant digits. Pre-populated bids? Current prices or last bids or advertised market price. Duration? Standard auction duration for the category. And so on. It’s literally just a set of rules, with tolerances, and RPA.

There’s only AI if the platform can run a sophisticated market and category analysis on internal, external, and automatically identified market data, identify a prime set of products in a category for an automated auction, determine the appropriate contract length (and projected demand) to result, and basically take a bunch of the analysis off of your plate for non-strategic / low-value, regular, purchases.

Example #2 of Applied Indirection in Sourcing: RFX Auto-Fill

One of the most time-consuming parts of the Sourcing process is waiting for the suppliers to fill out the RFIs with not just the bids but all of the other information you require. However, there’s no reason that a lot of this information can’t come from their profile, their catalog, and previous RFI responses (where you asked the same questions).

It doesn’t take AI to encode meta-data mappings between profile fields and standard RFI fields, catalog fields and standard RFI fields, and re-use of the same question across RFIs (and previous responses) to pre-fill the majority of an RFI and simply require the supplier to confirm and enter new, additional, or changed data (and, of course, check the boxes that say they accept the T’s & C’s, verify the data, and confirm their bid).

Unless the platform can seek out additional data on the supplier’s web-site, third party directories, third party audit sites, and process semi-structured and un-structured data using natural language processing and identify and extract new data and new information relevant to the RFI (and the sourcing event), it likely doesn’t have anything close to AI (even in AI’s weakest form).

Example #3 of Applied Indirection in Sourcing: Outlier Identification

Here’s another capability that doesn’t require anything close to AI. Statistical algorithms to identify outliers have existed for decades and decades. Many even pre-date modern computers. Run a simple mean/modal comparison, do a simple clustering maybe even use a regression. Easy-Peasy. Even Kitty can do it!

The trick is to identify outliers that aren’t easily identifiable mathematical outliers. A bid in range that is not sustainable for a supplier because the bids were supposed to be landed cost bids (and included shipping as the buyer wasn’t taking possession until the goods hit the warehouse) and one supplier’s bids clearly didn’t, when you think about it, is the type of outlier we want the system to detect for us.

If four of the suppliers are near-shore, and the cost of shipping for them on a unit basis is typically 10% of unit cost, but a fifth supplier is offshore, and the cost of shipping for that supplier by unit is typically 30% of unit cost, even if the bid looks okay, it might not be. For example, if the bids for all the nearshore suppliers are between $100 and $120, their actual unit prices are about $90 to $110. If the off-shore supplier, with a shipping cost around $30, comes in at $100, a mathematical outlier algorithm won’t detect that it’s underlying unit bid is roughly $20 less than the lowest bid, and if the driving costs are expensive raw materials, and the should cost model for that supplier indicates a production cost of $80 (even though their production and overhead costs are significantly less), then a bid of $100 is unsustainable and an outlier (or will not be honoured as the supplier misunderstood and expected to bill shipping separately). If the platform truly has assisted intelligence, it should detect this, even when a human doesn’t. (While you always want the lowest cost, you want the lowest sustainable cost — it doesn’t do you any good if the supplier goes bankrupt halfway through the contract and you have to scramble to find another.)

Note that SI is not saying that systems with the non-AI abilities discussed above are not valuable, as any system that automates tactical processes and minimizes non-strategic busy work is valuable. We are just saying you shouldn’t pay for what you’re not getting, or overpay for what you are. Buy what you need, and pay accordingly.

AI: Applied Indirection in Supplier Discovery & Management

Hopefully we have made it clear by now that most of the time you hear AI you should think “Applied Indirection” and not “Any form of Intelligence” because most solutions claiming to be AI are really just dumb systems with RPA (robotic process automation) and classic statistical models from the 90’s (which were available in SAS in the 90’s as well, you just didn’t have enough memory on your PC to run all the data you wanted to run).

But we want to make it abundantly clear that most of the “AI”, even in our space, is not “AI” at all. So, to do this, we’re going to take the major areas of SPT (Strategic Procurement Technology) and highlight some areas where AI is commonly claimed, but rarely found, starting with supplier discovery and management.

This doesn’t meant that there aren’t vendors with true AI, especially when you classify it as Assisted Intelligence (and sometimes even Augmented Intelligence), in the space, just that, as the buzz-acronym reaches new heights, there will be many more vendors claiming AI than those that actually have AI and you will need to do your homework to find out which is which.

Example #1 of Applied Indirection: New Supplier Identification

A true assisted intelligence system will scour a database, network, etc. and identify potential suppliers based on common product categories, like production or service capabilities, and community profiles and use some fuzzy logic* and adaptive modelling to make recommendations that you might not even thought about.

In contrast, many systems that claim to be AI will simply use SKU, key word, or strict sub-category meta-data matching to suggest the same suppliers over and over again, most of which you’ll already know as these will be the ones coded in the database or network to meet a particular demand. That’s not AI, that’s just multi-faceted search.

Example #2 of Applied Indirection: Auto-Profile Completion

Many systems that claim to be AI will simply use meta-data to map profiles from one system to another where a mapping between the field names and types exist (in a canned profile) and the data types are compatible. That is just ETL that has existed for over two decades, with good RPA behind it to identify the right mapping file and deal with exceptions appropriately.

In contrast, a true assisted intelligence system will be able to automatically construct mapping profiles from a new supplier record in a new system to the current system based on meta-meta data, automatically identify missing data, and automatically identify that data in semi-structured / un-structured text in a supplier description or overview from the supplier’s page on a directory or their “about-us” site using state-of-the-art NLP (natural language processing) technology, only asking a human to intervene and approve this automatically identified data if the probability of accuracy is not sufficiently high or to manually enter data (or contact the supplier for such data) when such data is not easily found.

Example #3 of Applied Indirection: Auto-Issue Identification

Many modern supplier management systems can automatically identify supplier-related issues, notify you with an alert, and even kick off a corrective action management process with a single click. But in most systems, this is not AI by any stretch of the imagination, not even close. It’s just RPA and classical statistical trend analysis in the best case, and simple rules and workflow in the average case. After all, you can detect an issue if a defect rate in the latest shipment is above a tolerance, if an invoice is for five times the number of units, or the satisfaction survey is less than 80% with a simple arithmetical rule. And in a slightly more advanced system, if the OTD rate is on a downward trend that will drop below a minimally acceptable level within three shipments, and so on, a simple trend analysis will suffice. And kicking off a corrective action management process is just automatically starting a workflow. No AI, by any stretch of the imagination, is needed.

In comparison, if there was true assisted intelligence, the system would go beyond simple rules, trend analysis, and notice early deviations in typical performance by looking across standard metrics and surveys to spot outliers that might indicate a trend, augment this with sentiment analysis on recent buyer feedback, and see if there is any external data that could indicate a potential downward trend is coming (such as a lot of recent negative sentiment directed to the supplier’s twitter feed or an article indicating a natural disaster in the immediate vicinity of the supplier’s plant). A true assisted intelligence system will give you early warning of a potential issue so that an account manager can investigate, and if there is a potential issue, take action before it materializes, or at least mitigate the issue (such as a supply disruption) if it can’t be prevented.

Note that SI is not saying that systems with the non-AI abilities discussed above are not valuable, as any system that automates tactical processes and minimizes non-strategic busy work is valuable. We are just saying you shouldn’t pay for what you’re not getting, or overpay for what you are. Buy what you need, and pay accordingly.

*Fuzzy Logic is a recognized area or discipline of mathematical study. The name of the domain was first proposed by Lotfi Zadeh in 1965, but it is actually an extension of infinite-valued logic that has been studied since the 1920s, by Lukasiewicz and Tarski, among others. And while such systems might not use this particular technique, they will use similar techniques that can use vague, or incomplete, or only partially matching data to derive conclusions and make recommendations with reasonable statistical probability.

AI: Applied Indirection Part III.B

Again, since we are in the situation where most claims of AI are just Applied Indirection to the lack of new technology being offered by the platform which is wrapping up old tech in a new UX with a little bit of RPA and, hopefully, better canned reporting and analytics, we are diving into the different levels of analytics to help you understand where AI might be and, more importantly, where it definitely isn’t. Because you don’t want to shell out six or seven figures (or more) for a “modern” solution that is actually only “modern” in the literary sense of the word (which defines the modernist period that started around 1900 and ended around 1965). And we’re not exaggerating here … some of the core statistical algorithms that form the foundation for a few of the bigger name analytic systems on the market date back to the 60s (and even 50s). (In other words, even the old grey beards who remember working on the last of the mainframes forty years ago wouldn’t have thought these techniques new back then.)

Yesterday we covered the first two levels of analytics. The next three are:

Level 3: Predictive

This is what most of the “advanced” analytic solutions on the market offer, predictive analytics, which, when you unwrap the messaging and peel off the fancy packaging, are simply statistical trend fitting and classic trend analysis algorithms that have existed in ERP for 20+ years and MRP for 30+ years. If the price is more-or-less going up according to a slight nonlinear curve, then the price is going to be predicted against the best non-linear curve the box-of-statistical-tricks can fit the data too. And so on. Again, not even a hint of AI here.

Level 4: Prescriptive

This is where AI in its weakest form MIGHT creep into the picture. The keyword here is MIGHT. You see, a prescriptive software application takes the results of a predictive analysis and makes recommendations on what you should do to improve the situation. However, there are two categories of recommendations here. The first category, which most of the applications are based on, is canned recommendations. For example, if the organization is currently spending over market price, prices are projected to go up, but demand still exceeds supply, the canned response will be an auction that invites the suppliers used in the past and highly rated alternative suppliers on the supplier network, as identified by community peers. No real intelligence, or even computation, there. The second category is dynamically computed recommendations, which may be based on a large set of rules or may actually use machine learning and dynamic computation and fall into assisted intelligence and actually make atypical recommendations when situations outside of the norm are detected due to unusual trend patterns or externally identified data (as per our example of web scraping in Part II).

Level 5: Permissive

A permissive system is a system that automatically executes a recommendation on your behalf but, contrary to manic marketing, is not autonomously intelligent. These systems are really just slick RPA (robotic process automation) systems that use a large rule base to drive workflows based upon whether or not recommendations are above a certain confidence interval, costs are within a certain bound, timelines are within reason, and so on (as configured by the vendor and the client on system implementation). More advanced systems will use analysis designed by experts to determine whether or not a certain recommendation can be automated, and then automate it with RPA if it can, and the most advanced — and these are extremely few and far between — will use Machine Learning that will record what a user does and then learn when a user is more than likely to take a certain response (based on past behavior) and when it can just begin to automate an action based on past behavior (and, in effect, define and modify it’s own automation rules). But the vast majority of systems still have no AI here whatsoever.

So, at the end of the day, while many vendors have sold their auto-classification, visibility, and prediction systems as AI — there was actually no AI under the hood and all the AI was applied indirection in the marketing organization. So, again, before buying such a system, be sure to apply a bit of logic and a sniff test. And if all you can smell is parfum de mouffette, you can be pretty sure there’s nothing there.