You know, but I like this example.
Transformation Vs. Optimization
I have a real life cautionary tale from the last two weeks.
I wrote and have been sustaining a Trello integration across a couple of ERPâs for several years. Service interruptions increased dramatically starting a few months ago. Hmm.
The morning of the 17th, Trello really broke. The API was down until the afternoon of the 26th. Totally dead for 9 and a half days. Active userâs work was also corrupted during the recovery period reported as âoperationalâ.
We were optimizing one of our largest databases, specifically the one storing comments and card activity. At the end of this process, a bug caused some database partitions to go live without their indexes.
Thatâs quite the oops. Donât throw out the guardrails with the essential institutional assets when putting the alien rookie in control of the revenue machine, I guess is the moral of the story.
That is a really good cautionary example.
I would be careful not to automatically assume AI caused that specific failure, but I do think the broader lesson absolutely applies.
When companies aggressively shift toward AI, automation, and cost reduction, they need to be extremely careful not to throw experienced people out with the bathwater.
I do not think AI is mature enough to simply replace knowledge workers. At least not in the way a lot of companies seem to be imagining. The truth is, there is not some fixed amount of work where AI does 10% and you can just remove 10% of the people.
AI is better understood as a force multiplier. It allows mature knowledge workers to increase their effectiveness, move faster, analyze more, document better, and build more. But that only works when the people using it actually understand the domain.
If a company cuts deeply into its experienced staff and then leans heavily into AI to fill the gap, it may be saving payroll in the short term while losing the institutional knowledge that made the systems practical and reliable in the first place.
In ERP especially, the boring parts are often the most important parts: audit trails, data integrity, deterministic logic, exception handling, rollback plans, support knowledge, and people who understand why the system behaves the way it does.
That is why I get nervous when AI is talked about like a replacement for expertise instead of an amplifier of expertise.
The Trello example is a good reminder that business systems are not just interfaces and features. They are also history, context, operational memory, and trust.
AI can absolutely be useful, but it needs to be layered in carefully: assistive first, governed always, and never at the expense of the people and controls that keep critical systems reliable.
There is a LOT about what I do that an AI agent could be trained to do. As an Operations Consultant that mostly does implementations, all the âtypicalâ paths are pretty well-documented.
But one of the most telling things I ever heard about being a consultant was from my first consulting manager (some of you may recognize the name Terry Ellis), who told me, âYou have to know the software. But if youâre not 80% psychologist youâre going to fail.â
Reminds me of that famous old IBM quote from a 1979 training manual: âA computer can never be held accountable, therefore a computer must never make a management decision.â
Yeah, I remember sitting down with a customer who was looking to spend well over $100,000 on an advanced OCR solution for AP Invoice processing and asking the question to find out what their distribution of Suppliers were.
It turned out that over 80% of AP Invoices, and >90% of AP Invoice Lines, came from their top 10 suppliers and that most of them were able to provide CSV or XML format AP Invoice files already. And all of the remaining suppliers were able to provide CSV or XML formats within a few days. It turned out that after a week or so of effort, they were able to setup some fairly simple workflow pipelines to import the CSV and XML files via DMT and programmatically attach the PDF copies from the supplier at the same time. While for the remaining AP Invoices, since the volume for manual entry was now much lower, it became far cheaper to simply have their staff continue to enter them - especially since the remaining AP Invoices were the uncommon/exception ones that OCR would have struggled with.
So, while I certainly see a huge potential for AI and see document processing as a major strength/opportunity to use it, I think that it is also worth taking a step back a lot of the time and considering other/business/human solutions too. And for documents, when there are key customer/supplier relationships, a lot of the time Iâve seen picking up a phone or sending an email to request/discuss sending and receiving structured easily digestible CSV/XML/EDI files as being a very successful/valuable approach.
Funny I asked that/exact question yesterday, instead of going down the bespoke third party SO automation route (Not even using ECM) btw). The response, âWe donât want the customer to have to changeâ⌠
The new solution has AI, so weâll be fineâŚ
Yes, I have experienced and seen that too many times too. I think we all fall guilty of stubbornness and ignorance a lot more than we like to admit (particularly when we are under pressure), but there are certainly many levels, and too many extreme cases, of both.
Fair point
I remember when Epicor had just purchased DocStar, or it was a year or two before brain gets fuzzy that far back, and they were showing the new âishâ AI based Invoice matching and going though the motions of how you train it for different documents for different customers etc. Really depends on what is meant by AI.
Trying to get customers to change is a uphill battle though. Your 1 company and you need to delicately ask people that are buying your goods to change which puts a cost on them. Itâs a chore trying to get come customer to put a partnumber on the order form let alone your partnum.
Yeah, absolutely. Circumstances definitely differ. My main point was itâs often easy for us to immediately pick up the hammer (or AI in this case) and using/applying it to everything.
AI capabilities are absolutely fantastic for lots of uses but with all the hype, I suspect that there will be plenty of wrong tool/approach for the job scenarios which could have been prevented by taking a step back and considering other options too.
Agreed, the way Iâve implemented it was to solve a problem that was a PITA before GenAI but I see to much of âlets use GenAI for everythingâ. If you talk to GenAI it literally tells you to not use it for quantitative problem solving. But you know I think GenAI is a solution looking for all the problems.
People might not pass that standard either. I think what we want is transparency in the logic that arrived at an important answer.
But we do not care about say the wording choices that end up having the same meaning but with a slightly different style. Peace.
There are people who understand Quantum Stochastics, and those that have yet to crack that book. Yes I am familiar with the Central Limit Theorem, but this gets to the point about understanding the level of abstraction you are talking about.