Supio's agent chases a medical-records request for weeks before escalating to the legal team

In the workflow Supio described to SiliconAngle, the fax goes out and then somebody has to remember to check whether anything came back, for days or weeks ^[4]. That somebody is the part of legal work most AI products have left alone. They summarize medical records, prepare a demand letter, search discovery or answer questions about one case ^[17], and staff still have to notice that an action is needed, find the right data, work the channels, follow up and document the outcome ^[8]. SiliconAngle wrote that this is why most legal AI has improved task-level efficiency and had minimal effect on firms' bottom line ^[7].

Eight steps: validate the provider's contact details, work out its request process, complete forms and HIPAA-related paperwork, fax the request, follow up by phone or email, monitor for a response over days or weeks, ingest the records when they arrive, and alert the legal team if the process stalls ^[4]. Seven of those the agent does alone. The eighth is a message to a human, and it fires on a condition ^[12].

The other number Supio gave is roughly two-thirds of the work in a case being some form of communication with an external party other than the client ^[5]. Take that at face value and about a third of the work covers everything else, client contact included ^[13]. It is the company's figure, from a briefing, and Supio did not disclose customer names, pricing or measured outcomes ^[14]. "Firm OS" and "system of action" are Supio's words for the platform, which pulls case data, firm knowledge, work status and Thomson Reuters case law into one place ^[1][6].

So the interesting question for anyone buying this is what the system does on day 14 of silence from a provider's records department. The numbers that answer that are cycle time from request sent to records in the file, the share of open requests older than your current median, and how many requests a paralegal touched at all. None of that shows up in seat counts or session counts.

Here is what I would establish before a pilot, on any agent pitched this way. Count the steps and count the handbacks, then check that each handback lands on a named person. Ask where the bounds on the agent's own decisions are written down, since a long-horizon agent by SiliconAngle's definition makes bounded decisions and escalates the exceptions ^[9]. And test the silence case, because the failure mode of an agent that owns a multi-week task is nothing happening while everyone assumes something is. SiliconAngle's own test of the category is whether the system can determine when to send an email, whether a response has arrived, when escalation or approval is appropriate, and where the result belongs in the system of record ^[11].