A company runs automated eSign workflows for their case-management platform. They send out envelopes containing guardianship agreements, available for signing, and once everything’s been signed and executed, the final document can be downloaded as a PDF. Standard procedure. Happens every day. Except one time – the PDF renderer didn’t capture the document content on that final step — and that was fundamentally the difference between a child’s placement being settled or it falling apart in court.
Looking at numbers such as rejection rates published by courts, which is not only high but also unstable in its nature (in LA County, the 2022/2023 rate swung from 8% to 80% of filings across 13 months), I can’t help but think these were the ones the courts found and rejected. What about the ones that made that cut and went all the way to trial?
And now, with agentic AI, we are making that process simultaneously more complicated and efficient in reaching specific outcomes. It really is “the year of the technical debt” as pointed by Salesforce Ben, and the reckoning for solutions dealing with human workflows is here.
Running SaaS for key human functionality like signing documents and identity verification now requires understanding of the 90-day reckoning, the increased cost in shipping quickly, the AI token bills increase not as a sign of profitability, but as a switch in the organizational culture with no visible harness. And, sometimes, agents actively causing irreparable leaks or loss of customer data. Many stakeholders in such workflow-based SaaS business roles (outside of engineering-adjacent roles more specifically) are inevitably being jumpstarted into having to understand how their business creates and maintains API governance and data security. With no visible harness either.
The obvious counter here is that AI models are going to evolve and the failure rate will continue to decrease. It has already turned functional requirements like uptime and HTTPS into table stakes. Very soon it probably won’t make sense to use reliability and compliance as differentiators to position a business. There is also an immense network of highly efficient vendors increasing their know-how and scrutinizing security in this space. Humans will get even better at creating utilities to solve small problems, go-to-market with security solutions and fight for a share of the market.
There is a strong point here for making this not about the trust in the solution or the artifact used to achieve it, but the accountability as the thing that creates trust. It’s the network of humans and systems around this workflow arena that establish what is accountability, not uptime or architecturally-sound code. It’s about the institutions and the people ensuring that the end-to-end workflow has the appropriate legal liability.
eIDAS is a great example: it provides assessment of conformity through auditors, certificates and qualified trust service providers who have to prove and keep their trust seal through mandatory legal obligation. Public organizations and courts in the EU/EEA must legally recognize an electronic signature to the same legal effect as a handwritten one with the appropriate caveats of each institution or legislation. It’s not about the certificate, but the trust built around it.
When a problem like the case management one happens, everyone who manages, buys and uses software thinks 1. How could this have happened? and 2. What are the ways in which it could be undone? So, before getting to that point, ask yourself
about how far the long tail of compliance of your company goes, and how your vendor network supports that tail, make it as reversible, verifiable and secure as possible for your business.
Is there a level in which APIs, MCPs and AI can create efficiency paired with control and security to avoid such cases that can’t be undone? A good answer is by using context and natural language in agentic software to make these irreversible errors visible and easy to detect and monitor. When an agent is invoked, the MCP can act on that direct boundary immediately when signing or redacting a document for the user.
There are many examples of reversible errors and artifacts where agentic software can grow with less strict rules and protocols. Agentic software allows developers to build front-end interfaces and document-driven apps incredibly faster, make user experience more tailored and creative, explore parts that were too far from the MVP before. Those are easy to revert. (Will that eventually change? Probably yes. Will it change dramatically and quickly? I don’t think so.)
On the other hand, a signature, a redaction, a fulfilled order, an evergreen signed agreement, a record permanently deleted are not. Those can’t be reverted without human interaction and accountability. And what is worse, is when these errors happen silently due to lack of awareness or acting on confirmation bias. There are many examples of errors like this, but the poorly redacted Paul Manafort Court Filing of 2019 is the one that comes to mind. In that case, while visually it looked like the document had been redacted (black rectangles in the page), the sensitive content had been silently available to the public due to lack of harnessing of the PDF tool and doing a simple action to test it. This is a prime example of “one that made the cut” that I mentioned earlier.
Better AI models will reduce the risk margin, but the impact of the irreversible errors, the potential cost, the lawsuit, it doesn’t need to be significantly large in quantity to create major problems. These situations tend to point to the opposite direction – small quantity of issues with high level of detail and potential to cause bigger problems. Speed is free where mistakes can be undone cheaply; and ruinous where they can’t.
A key buyer of an application is not the developer who’s running the demo, or the admin. It’s the person who must explain to a judge how this could have happened and what are the ways in which it could be undone. They don’t care about agentic software, they want evidence and accountability: what happened, where is the
record, who’s accountable at that boundary. Aim to be the buyer who knows exactly how you want a provider and their product to handle irreversible errors like the one discussed here.
I’ve been building tools and utilities using APIs and MCPs as a vendor while also building the underlying infrastructure for use cases where errors can’t be undone. I find this to be one of the most challenging moments to be operating in a SaaS company in the regulated workflow space. The stakes are as high as they’ve ever been. Irreversible errors happen, every day, and will continue to, and the key thing any vendor and operator in the industry can provide to their user is carrying the liability when errors happen. That’s what we strive to compete on at Foxit – carrying the liability with the appropriate safeguards. Because while the tools improve, someone will still need to be on the hook.
Related Categories