Recent stock movements away from Software as a Service stocks has been interpreted as a result of the fear that pureplay generative AI companies like Anthropic and OpenAI would disrupt companies like SalesForce, SAP and ServiceNow which form the bedrock of companies’ CRM, Finance and support functions. This echoes a frequent perception by Investors, purportedly experts in technology, that AI is going to take over everything and most of the jobs too. That remains to be seen, but in this particular case the conception that “generative AI software companies will disrupt and replace ERP systems” seems over-optimistic bordering on ridiculous.
Let’s investigate the story that started it all. At the start of February, Anthropic released a plug-in for legal departments, which “could automate legal work such as contract reviewing, non-disclosure agreement triage, compliance workflows, legal briefings and templated responses”(The Guardian). When you go to investigate this plug-in, you find it is indeed built to “speed up contract review, NDA triage, and compliance workflows for in-house legal teams” all of which can probably be stitched together in a standard prompt by any half-studied youtube proficient AI specialist in an afternoon. These are functions that most legal companies have already built themselves. Anyhow, these features have existed in legal software for years already.
What the story conveniently doesn’t tell you is that Claude just needs access to the data, and oh, you just have to build an MCP infrastructure around your services for this “free” plug-in to work. It also does not handle any of the regulatory requirements that are license-to-operate-level requirements for legal counsel. So, basically Claude offers a plug in that you still have to configure and integrate and you still need to handle all of the document management and regulatory controls somehow on the side. What’s more, it doesn’t even integrate with Word which is the standard document management system for legal counsel, like it or not. All in all, in the real world it appears to be relatively light in its value add and ill aligned to how humans actually work.
If we generalise this to SaaS in general, as seems to have been the case, are we going to trust the LLM with keeping the records of book keeping, statutory reporting, and compliance reporting, and if we are, how are we going to make regulators trust it? Systems of record like ERP and CRM have one important characteristic that LLMs do not have and which is indeed orthogonal to their mode of function: they are stupid, predictable and transparent. When you put a number into an ERP system it saves it, and it is unchanged until you look at it again, it doesn’t make evaluations or judgements. When it calculates it does so in a dumb transparent way that one can audit.
Using LLM plug-ins is potentially a great idea, it just doesn’t solve the fundamental problems that Software as a Service systems solve, which is providing a system of record for authoritative information that can be accessed and audited in a transparent manner to satisfy internal and external requirements and regulation. To use these plug-ins, you need to build an integration infrastructure, which is just as costly and complex as any other system integration. The value add will be much lower than expected and the investment in integration and data access much higher.
Companies are therefore moving to embedded AI instead. Generative AI is already being commoditized, so it is much less risky and costly to use existing embedded AI functions or even wait for the vendors, with greater industry understanding than Agentic AI vendors, to build tailored functions for them.
Photo by Vitaly Gariev on Unsplash