AI is Not AI … And AI-Related Tech is Not Created Equal

Joël Collin-Demers recently made a post that correctly stated that real “AI solutions” tell you exactly what they do, in which sequence, and [help you] understand how it solves your exact problem. the doctor totally agrees. Otherwise, they are using buzzwords and trying to cash in on the hype to sell you old-school automation at best, or third party (Gen-AI LLM) wrappers at worst, and don’t have any real AI.

He covered ten different types of technology and attempted to capture the positives, the negatives, and the best uses therefore in source-to-pay. For a relative non-techie (compared to the doctor with a PhD, degrees in CS and Mathematics, and actual experience implementing everything but BS LLMs from scratch), he got a lot right. But he got a few things wrong. As a result, there was a need to correct him (in this post) and ensure the corrections have as much permanence as the original post.

We do recommend you read his original post, but for each tech addressed, integrate the correction below.

RPA – does not “break” when processes change; it breaks when you feed it bad data; when you change processes, it simply loses its usefulness until you change it to match the process — with a good RPA system, that shouldn’t be hard

Machine Learning – requiring clean data is NOT a bad thing; it ensures the algorithm “learns” the patterns you need it to learn to use it effectively

Natural Language Processing – doesn’t struggle with jargon, just context — you define the dictionaries, the grammar, the language — it’s accuracy boils down to that; if you are feeding in documents that use the same words/phrases in multiple contexts, it will always struggle with that to a point

Predictive Analytics – in lay terms, classical predictive analytics is essentially just multi-dimensional curve fitting based on the data available — when something has not been modelled, there is nothing the algorithm can learn to fit against — but to be fair, nothing you’ve listed will succeed with unprecedented events/data

Anomaly Detection – this is based on outlier detection, and well trained outlier detection does NOT have high false positive rates (which are no higher than false negatives), and any “wrong” classifications from a business perspective simply means that the definition of a valid transactions needs to be amended to reduce the “outliers”

Computer Vision – lighting is not as much of a problem as you think as most good algorithmic interpretations will always mathematically adjust the brightness and contrast to a consistent range in pre-processing, and sometimes even greyscale; angles are a problem, because if they can’t be determined, the right transform can’t be applied to appropriately orient the image to maximize identification likelihood

Optimization Algorithms – “requires precise problem definition” is not a bad thing, it’s a good thing — if you don’t have a precise problem definition, you cannot get a precise answer with any technique; one of the best uses is product mix, not just supplier portfolio

LLMs – not “can” hallucinate false info; “will” hallucinate false into — every single time, just a question of the degree; it’s “generative” AI which literally means it makes stuff up, and how accurate what it makes up is with respect to your problem depends on how it was trained, what was asked of it, and how you ask it … very unreliable all around

Pattern Based Recommendations – the whole point is to “filter” to what you would normally buy so that you don’t have to sort through everything, that’s not a problem

Reinforcement Learning – it does not require extensive time, it requires extensive data — computers process mathematical calculations billions of time faster than we do — it’s never time anymore!

AI Doesn’t Help You Get The Most From Your Spend Analysis System

Twenty years ago we wrote a post on how to get the most from your spend analysis system, and the reality is that not much has changed. In fact, AI has only cemented the need for the foundations.

Twenty years ago, we essentially told you the keys were to:

  • streamline the tactical
  • focus on strategic analysis that always delivers

And those haven’t changed. Because:

1. With AI overloading you, you need to verify, or dismiss as fast as possible, which means tactical should be quick, easy, and effective.
2. With AI vendors claiming they can replace you, it’s critical to start with analyses that deliver results to prove the power of manual analysis.

We’ll review the tactical capabilities that are still super relevant and the strategic capabilities you should always keep in mind.

Your tactical skills with respect to the following should be best-in-class with leading peers:

  • (sub) cube creation off of the primary cube (using inheritance)
  • re-classification, derived dimension creation, and view modification
  • baseline and compound view-based report creation
  • direct integration with the contract repository, supplier management system, and external data feeds
  • automated maverick spend and override approvals

Your strategic skills with respect to the following should be second nature:

  • overspend calculation on “best price” agreements against market prices offered by the vendor through other channels
  • purchase order vs invoice vs payment analysis for overpayment recapture
  • outlier analysis tuned to (potential) fraud identification (both internal and external)
  • loss prevention (repair vs. replace analysis, keep vs. return analysis, price reduction vs. charity donation for write off analysis, etc.)
  • compliance analysis (right products/services being used, payment terms being adhered to, etc.)

But, most importantly, your perception skills need to be top notch. When the AI spits out an opportunity, you need to:

  • identify precisely not only what the AI thinks the opportunity is, but how you would verify it, and if the opportunity is real, capture it; but if not, disprove it quickly
  • when the opportunity turns out to be false, identify the direction the AI was heading, what the AI got wrong, how to identify such errors instantaneously in the future, and what direction to give to the AI handlers to stop wasting your time with obvious garbage
  • how to identify the gaps in what the AI analyzes to find opportunities even faster than the cognitive atrophied team driving the AI

While the AI can kick off the tactical, and support some of the strategic, it should be clear that it falls short in all three categories, and doesn’t even scratch the surface where real perception is required.

AI Doesn’t Change the Fundamentals of Spend Analysis

Twenty years ago, we wrote a post that said, so you want to do spend analysis?

And in that post, we noted that you want to get on with it, but you’re not sure where, or how, to start. So we provided you with a step-by-step process you can use to get your spend analysis effort under way and keep more of those corporate dollars in the corporate coffers, where they belong. And, after 20 years, the steps remain, more or less, the same. Because the song hasn’t changed. Only the noise by those trying to sell you fancy tools (that don’t work) has gotten louder in their attempts to drown the song out.

1. Locate Your Data

You need to identify where all of the source data is. Not the data warehouse or data lake you’re pointed at or someone’s random, monthly data dump, but the source data. That’s because it’s not likely that all of the data you need will be in the warehouse, lake, or dump because that likely consists just of the ERP or AP system, and we know that no single system contains all the data you need. Spend will be split between the ERP, AP, and T&E. Supplier information in the ERP, SXM, and Risk Management Systems. Contracts, and key data, in the CLM and Legal Systems. And so on.

Make sure you know where all the source data you need is, what warehouses and lakes it is pushed to, on what schedule, and which of these you can use. If you can work with system owners to get the missing data pushed to central locations, or if you will need to retrieve it directly.

2. Create a starting taxonomy

You can’t integrate the data otherwise. As long as the taxonomy integrates and federates all of the data, it will be sufficient. Remember that a modern spend analysis system supports multiple cubes, with inheritance, that each cube can have its own hierarchy, and that cubes and sub-cubes can be created on the fly.

3. Centralized the Data

Centralize the data into a single (virtual) starting cube that everything can work off of. Do this by defining the master transaction records, absorb all of the related information, and build the initial (virtual) cube(s) that every (derived) cube will work off of. (By virtual, we mean that the system can pull data on demand as needed with predefined definitions, it doesn’t have to be stored in a single cube in a single warehouse or lake.)

4. Family the Data

In just about any ERP or enterprise system, there are multiple entries for every supplier, product, or other entity. Make sure these are familied into one instance in the master data set.

5. Map the Data

As the data gets pulled in, map to the starting taxonomy so that the data is ready for use.

6. Pick the Low Hanging Fruit

Always start with the easy opportunities — today those might be the highest ranked opportunities spit out by the AI, the ones from the “start here” guide (supplier rationalization, bypass spend identification, growing, off-contract, categories, etc.). A few quick wins will build up support for a real, manual, spend analysis effort — which is where the real results will come from.

7. Make sure you get the right tool!

Remember, the tool, which should have all of the features defined in our previous entry in this series, must be one that gives you the agility, speed, and power you need to apply your intelligence to the problem at hand.

8. Get a good consulting mentor

You want someone to teach you not only how to do spend analysis, but how to approach and even think about spend analysis so that you can translate your instincts into analysis with quick, pinpoint, accuracy and answer your queries faster and better than any AI ever could. That’s not your run of the mill consultant who just executes the standard slate of analysis and applies the AI tools in their stables, that’s a real analyst who understands what analysis truly is and is willing to teach you. They are few and far between, and may not come cheap due to high demand, but once you learn, you won’t need them anymore. And what you spend up front will be insignificant with respect to the opportunities real analysis can uncover.

AI Doesn’t Re-Define What True Analysis Requires Part II

Real analysis requires more than just classifying data into a simple hierarchy, rolling up some numbers, and generating a report or two. It requires more than traditional BI tools which, when all is said and done, don’t do much more than that.

Going back to the beginning when we first defined the New Horizons (Part 1 and Part 2), we find that modern AI still doesn’t always support even the foundations that the Spend Master Eric Strovink defined two decades ago.

But that was just the beginning. Today, we also need:

Cube Reverse Engineering and Easy Rule Construction

If the analyst has an existing cube, then the spend analysis system should be capable of extracting all of the rules flawlessly (as well as detecting the classification errors) and presenting them to the user for easy review. If not, then the system should be able to extract all of the rules used for data organization in the underlying systems.

Moreover, for data that still remains unclassified, it should be easy to use multiple AI techniques to generate suggested rules that can be easily accepted, modified, or rejected.

Reusable Cube Template Definitions

That includes derived, aggregated, and federated dimensions as well as pre-defined views that can be easily mapped to existing data sets, edited, and modified to an analyst’s liking. A user should never have to start from scratch when they did a similar analysis before. And tweaking should be as few clicks as possible.

It should also allow for pre-defined cleansing and categorization rules, easy tweaking thereof, and easy integration of Gen-AI for mapping rule suggestions when the built in capabilities are insufficient.

Reusable Filters

It’s not just cubes that need to be reusable, but filters that allow for customized drill-downs into federated hierarchies — it should be super simple to extract, replicated, modify, and save filters that customized cubes and views to exactly what the user needs to see when they need to see it.

Inheritance

By now, you know the big problem with spend analysis is one-cube with one hierarchy based on one schema. A hierarchy defined (on a schema defined) by consensus satisfies no one. A hierarchy defined by an analyst satisfies only the analyst, and only for the problems she is working on at the time.

As a result, in an average organization, whenever analysis needs to be done, a copy of the data is made, another cube is built, another customization performed, another report printed off, another decision made, and when it comes time to revisit the analysis, or rerun it with updated data, no one knows which of the many data sets is the right one to start with, or even where the most recent data from the source apps got pushed to. Because everything is disconnected, analysis makes a mess.

But if the spend analysis application supports inheritance, where cubes can be derived from existing source cubes, and where data updates are automatically propagated down, but not up, then there only needs to be one copy of the source data, and as long as that source is kept up to date, all of the sub-cubes can be kept up to date as well. They don’t need to be tracked, or even maintained when their usefulness has come to an end, as they are just temporary instances a user can create to answer specific questions, where they can manipulate the data, and hierarchy, to their hearts content without impacting any other user in the organization or corrupting the data.

Because they purport to eliminate the need for these capabilities with their advanced capabilities, most AI systems just don’t provide most of these capabilities. But these are the capabilities that true analysis requires.

AI Doesn’t Re-Define What True Analysis Requires Part I

Real analysis requires more than just classifying data into a simple hierarchy, rolling up some numbers, and generating a report or two. It requires more than traditional BI tools which, when all is said and done, don’t do much more than that.

Going back to the beginning when we first defined the New Horizons (Part 1 and Part 2), we find that modern AI still doesn’t always support even the foundations that the Spend Master Eric Strovink defined two decades ago.

Two decades ago, in the original series, we defined:

Meta Aggregation

You don’t just need roll-ups and derived dimensions, you also need “cluster dimensions”. The ability to consolidate spend / vendors in groupings in pre-set ranges (such as the number of consulting vendors under 100K, the number between 100K and 500K, and the number getting the big bucks at more than 500K a year). In other words, classic f(x) doesn’t cut it. You need g(f(x)). You need recursive functions. Which can dynamically update in real time not just when the ranges are changed, but when the hierarchy of the source dimensions is altered (or the source data altered).

Visual Crosstabs

The utility of Shneiderman diagrams (or “treemaps”) to display hierarchical information is well known; the treemap is useful because it is visually intuitive. The relative sizes of the rectangles represent the relative magnitude of spending; the colors indicate relative change in spending; and inner rectangles show the breakdown at the next level of the hierarchy.

Now, suppose that rather than the inner rectangles showing a lower level of the same hierarchy, instead they showed the breakdown of spending within another dimension entirely — i.e., a “visual crosstab.” The visual crosstab would not only show magnitudes, but trends as well.

Federation

 

One of the most serious limitations of OLAP analysis is the schema structure itself — typically a “star” schema, where a voluminous “fact” or “transaction” file is surrounded by supporting files, or “dimensions.” In the case of spend analysis, dimensions are Supplier, Cost Center, Commodity, and so on; transactions are typically AP records.

There are only certain ways that a dimension file can be linked to transaction files, and it isn’t always clear which file ought to be the transaction file and which files ought to be dimensions. For example, suppose that the transaction file consists of AP transactions, and a dimension file consists of invoice line items. The problem is that the invoice line item file is “moving faster” than the AP file; i.e. for every invoice number that appears in AP, there are multiple invoice lines that match. Which invoice line item should we link to?

There are numerous other examples. But the essential problem is that we have two separate datasets, and we’re trying to join them at the hip. There is an AP dataset, and there is an invoice line item dataset, and never the twain shall meet, except artificially. Even when there is no granularity issue at all, and when one dataset can be normalized or snowflaked such that every matching line item can be joined through from the other, the amount of effort required to set up the index->index->index relationships can be daunting.

Instead, why not create two separate datasets, efficiently and quickly; and then as a final step, federate them together on a common dimension? … Federation represents a key productivity enhancer for dataset creation, as well as a simplification to the dataset building process in general.

But that was just the beginning.