Exploring the business applications of multi-modality
In trying to work out what processes could look like in the future, it’s interesting to see how we are anchored:
“This meeting will have this app to support…”, “the preparation will have this process and toolset…”, “the aftermath will involve automated action here and here creating template x…”.
We assume a process of preparation, event, and follow up. That is a natural process for people who need to put time into understanding (computing) the inputs and forming something professional and cohesive. But for an intelligent multi-modal AI, it may not be necessary.
Hypothetical: financial advice
A financial advisor joins a Teams meeting with a client. So does an Assistant AI that has a knowledge base of the firms’ products, services, and policies, as well as integration to their CRM.
Taking verbal and non-verbal cues from the financial advisor and the client, the assistant instantly and expertly crafts a package within parameters established by business administrators. The client may then use their own assistant, with a deeper knowledge base of the clients’ financial position, to review and meaningfully challenge the package. In the end it’s agreed, set, invoiced and paid. This all happens within a simple conversation, with relevant information appearing dynamically throughout.
No clicks. No navigation. No follow up required. Just AI agents performing processes based on systems, audio and visual data.
This is not waiting on some imagined AGI. The hypothetical presented is within the bounds of current technology.
Considerations
Accuracy and reliability: AI hallucinates and doesn’t often say when it doesn’t know something. Somewhat mitigated by direct human oversight (more to come on how to make oversight performant).
De-risking a real-time process: unlike more traditional processes, a real-time process doesn’t have takebacks. You are dealing with live ammunition. However, this is something that I liken to working with new colleagues, and another reason why oversight exists. What I will note is that with a higher role afforded to automated processes, the need for trusted systems increases.
Performance of integrations: even with a high performing AI model, the integration needs increase without humans connecting the dots. As such the impact of any associated performance bottlenecks would be felt in real-time.
Responding to new external stimuli reasonably: the chaotic nature of reality is tough for traditional automations, and whilst AI can make sense of the disordered, it needs to drive things back on track (a tall order). The question is, how can we make AI drive an agenda?
Everything really comes down to the quality of the AI model, and any fine-tuning that may have been applied.
