Is AI the Poly-Fil enterprise software needs?
AI has the capacity to fill in the gaps that held back the first IT revolution
The first IT revolution, roughly the arrival of desktop PCs and the tools built for that platform, was a time of heady hope for large enterprises. The clean abstractions sold by vendors and consultants promised management a degree of control and transparency over business processes that had previously only existed in their wildest dreams.
It’s hard to answer questions like “in aggregate, did the hype around enterprise IT in the 90s deliver?” I suspect the spectacular implosion of the Dotcom Bubble lives larger in the collective psyche than the real and widespread incremental gains that IT adoption did deliver.
My personal guesstimate is that to the extent the tools failed to deliver, it is at least partially because the available software lacked the malleability required to capture how business is actually done. We lacked a “Poly-Fil” to efficiently smooth the touch points between processes and tools, and the way people and computers work.
Change the software vs. change for software
Nothing exemplified the promise vendors were selling, and implicitly what they were missing, more than Larry Ellison’s pitch for the Oracle E-Business Suite. In Softwar, part biography of Ellison the man, part realtime account of Oracle’s attempts to sell enterprises on Ellison’s vision for the suite, the subject of the book laments the obstacles keeping management from the Holy Grail of business process automation:
Most business processes are not completely automated; they’re a combination of online and offline activity. For example, in the sales process called ‘opportunity to order’ each line of the sales order is always entered into the application system, but the price quote and sales contract might be stored separately in an Excel file or a document image file. If you store your price quotes in Excel files, that means your applications database is incomplete, so you can’t ask the system questions like ‘What is the total value of all my outstanding quotes?’ Similarly, if you use a document image file to store contracts, you cannot ask the system, ‘How many of my contracts have nonstandard limitation-of-liability clauses?,’ because that information is not stored in the applications database. In other words, if your process automation system is incomplete—and most are—then your applications database will be incomplete as a consequence.
Almost twenty five years later the characterization of data living in silos across data formats and productivity tools (images, Excel, etc.) feels like a fair description of how many business processes actually operate.
Ellison clearly understood there was a problem. His solution was to urge businesses to change how they operated to work with software, which would come as laughably naive were it not coming from one of the great winners of the first IT revolution, and a great technical visionary (something horror stories about Oracle databases tend to obfuscate).
It didn’t work
The Oracle E-Business Suite was not a success, especially in the form Ellison was pitching as a set of integrated tools. The market eschewed both his philosophy of integration, evidenced by the explosion of “best in breed” SaaS products, and the product itself.
Nevertheless, he certainly understood what enterprises viewed as the promise of IT. He knew what he was selling, even if the product and his preferred adoption path didn’t work out. The examples he provides are exactly the kind of realtime information that management often crave, and in turn lean on unit leaders (or analytics, or business operations, etc.) to assemble:
What is the total value of all my outstanding quotes?
How many of my contracts have nonstandard limitation-of-liability clauses?
In “internet first” businesses (many of which, not coincidentally, came after 2000), this sort of information can be quite straightforward to obtain. When all customer interactions are digitally mediated, it’s just more likely that the resulting data is amenable to querying, slicing, aggregation, and anything else you need to do to get to an “insight.”
Most business doesn’t happen online, though, and the more complex the transaction is, the less likely it is to be completely described by a blob of JSON. A lot how business operates is spread across formats and documents that don’t play nicely with available workflow tools, but don’t cause enough friction to motivate the lift required to be fully captured in an end-to-end software solution.
No easy solution
To make things concrete, let’s focus on Ellison’s example of identifying contracts with “nonstandard limitation-of-liability clauses,” but you could substitute this for any other query over the unstructured data a non-trivial business process throws off. How would organizations go about being setup to ensure they can answer questions like this?
- They don’t. If a question like this becomes urgent, then someone just has to go and read the contracts. For a lot of use-cases, and for some kinds of question, this is probably totally fine.
- It’s captured in Salesforce (or something similar). Many organizations, ranging from a furniture store I shop at to investment banks, use Salesforce to capture the various aspects of their customer interactions.
The problem with the second solution is that organizations need to configure a tool like Salesforce ahead of time to capture the data they need. This sounds like it might be a solution, but it’s often the case that the most interesting questions are not known before they are asked, and the end result of this is a tool that is configured to capture anything anyone in management ever asked about (a Google search for “Frankenstein CRM” will substantiate this observation).
I think this is a place where LLMs can have a big impact on the enterprise landscape, by acting as a kind of Poly-Fil that helps bridge the gap between between the way people actually work and how computers they use capture information in a way that useful.
LLMs as software Poly-Fil
I should say upfront that I am pretty skeptical about some of the more garish takes regarding how AI is going to help execute business processes. These tend to range from the truly hysterical AI-will-run-the-sales-process down to the more plausible claim that it might be able to identify, for example, a relationship between a salesperson’s behavior and the existence of non-standard business terms (or discounts, or anything management might want to discourage).
I am, however, optimistic about the more prosaic goal of radically changing the cost of capturing and querying information, and finally seeing the heady dreams of the early 90s fulfilled. In a world of LLMs there is no need to choose between salespeople remembering to check a box in Salesforce that says “non-standard liability” and asking a recovering consultant from BizOps to check all the contracts from the last two quarters (often the second one when it actually matters, and then the first forever more when it no longer does). LLMs can help answer questions like this, and they can do so without knowing what the questions are ahead of time, or implementing pedantic data capture strategies that burden teams in their day to day work. They have the capacity act as a translator between systems, allowing the components to operate in their native language (Word, Excel, PDF, etc.), and then integrate without making enormous data integration investments.
One might say this article is the enterprise IT edition of Text Is the Universal Interface, a fun article that appeared on Scale’s website written by an OpenAI researcher.
Conclusion
The first wave of the IT revolution sold enterprises clean abstractions with phrases like “end to end process automation,” and created a tremendous amount of (justifiable) excitement over what the future might hold. Management dreamed of realtime visibility into their business processes, and powerful querying capabilities for gathering the data to understand and improve them. Many of these dreams came true, especially for businesses that were founded with the assumption of these new tools.
The tidal wave of innovation in generative AI is incredibly exciting, but I think one under-discussed opportunity is for enterprises to reevaluate process automation, workflows, and analytics strategies that the previous wave of software lacked the capacity to digest in an economically viable way. Management should be able to grasp the themes of customer conversations in the middle of the quarter without relying on a vibes check from a sales manager, or burning the midnight oil scraping together data.
The productivity gains from digitizing business have been enormous, for businesses and consumers, and I believe there is much more to come.