The excitement around periods of accelerated innovation like we are in today with AI partially stems from the fact no one has perfected a solution.
Incumbents and upstarts, practitioners and investors, specialists and generalists alike are piecing together the mosaic of what the end-state(s) for AI might look like, which feeds the intense online discussion and media coverage that rivals the current equity bull market…or drives it depending on your view.
We are all learning, iterating, and comparing views, while searching for analogues in history and from the very human behavior we are trying so hard to mimic.
Jaya Gupta’s and Ashu Garg’s post expanding on Jamin’s Ball’s view on the future of system of records, argue the missing link for AI is context. Whether captured as a graph or in other forms of memory, context or judgement traces, as David Park refines the argument further, is essential for effective AI to turn ambiguity into action and faithfully understand how human knowledge work gets done.
Rather than re-inventing a solution that nature had perfected, we ask how do humans learn context? Specifically, how does a new hire learn the context and develop the judgement appropriate to do the work they are tasked at their new job?
At Metis we deliver the infrastructure for firms to infuse their AI with the context and judgement that fuels their processes and transparently learns how to bring together their resident intellectual property to yield the unique output their clients value.
We are often asked by our clients and prospects – who does Metis compete with?
The answer is a new hire assistant for every person at the firm that inquires, listens, and learns how to help them get more work done and perform better.
A new hire for a company comes equipped with an education, skills, and set of experiences that have been vetted by the company and their direct manager.
They have zero context on why their colleague formats an Excel sheet a certain way or prefers one data source over another apparently identical source or that their manager expects internal reports completed one way but external reports on the same topic a different way.
A new hire has no context on the history of how processes have evolved and why and what that portends about how they could be improved in the future.
Every system of record and the systems of engagement that existing employees use to access those records is customized to the user or their team’s preferences and learnings accumulated over years of experience.
A new hire, regardless of pedigree, does not know how work gets done within a new firm until they take their seat and start the most important part of being a productive and accepted part of an organization– learning by asking questions and observing.
If we think back to the best people we’ve hired, we describe them as “self-starters”, “inquisitive”, “resourceful”, “quick study”, and “proactive”. Star performers share many of these accolades, which they earn by asking astute questions and intently observing the work being done around them.
What is the file structure on the shared drive for my group, and is it different from other teams? Why do we do this specific action for this client and not this other one? Where is that data stored and how can I access it more efficiently? Who is the most effective at presenting the Monday morning meetings and how can I learn from them?
Eventually star performers anticipate your needs and exceed expectations for the work they take on and quality of output.
This process takes time, in some cases months of managers and peers investing effort to answer the new hire’s questions, impart the context and cultivate the judgement for effectively working within the standards and culture of the organization.
For an investment firm, context and judgement is taught by learning the unique process and philosophy of the firm that define its people and ultimately drive long-term performance for LPs.
Importantly, employees are not identical machines that can be finely calibrated like a welding robot on the factory floor, the process by which they produce ostensibly the same output might vary markedly and could permanently change any given day. While on paper knowledge work might be processed mapped to look like an assembly line, the actual way work gets done is not nearly as structured and is tied together by context and human judgement.
Volodymyr Pavlyshyn in his piece “Beyond Context Graphs”, adds further insights, by arguing AI infrastructure needs to integrate multi-layer/ multi-state memory, explainability, and deterministic causality. The interplay of all three is the prerequisite for creating trustworthy enterprise automation of knowledge work.
Metis delivers AI infrastructure for our clients that integrates all three vectors. Our platform engages with users and builds trust with two key tenants that are inspired from how the best new human hires operate:
1) Show your work! Since grade school humans are told by their teachers to write out how they solved a problem. Students inevitably make mistakes and only by showing their work can a teacher spot the error in the process to help them learn how not to do it again. The workplace is no different.
Managers ask for the Excel back-up or the sources of data when a new hire first completes a project so they can audit the output, particularly if it falls short, and teach the new hire how to improve. We place significant value on a new hire that assimilates feedback and doesn’t need to be corrected again.
Metis shows its work so the user, as the expert in a process, can provide feedback in plain natural language, as to how to improve. Importantly, Metis has the memory to retain the learning and even extrapolate the user’s feedback to other similar jobs, while always visually showing the user its work.
User trust in Metis is built on transparency and speaking the same language.
2) Ask questions! The old saying, when you assume, you make an a$$ out of you and me, is what users feel when traditional AI hallucinates on them.
New hires that don’t jump to conclusions and ask for guidance and feedback when they don’t know or are unsure are highly valued by an organization as it gives managers and peers an opportunity to teach and avoid compounding errors.
Metis is programmed out of the box to be inquisitive and because it breaks down every task and problem into the most basic sub-components, it surfaces the steps that are low confidence or where the required information doesn’t exist and asks the user for clarification and answers.
By asking questions, Metis further builds trust with the user and enables the user to build automation with high fidelity and confidence.
Metis carries the same burden as a new hire, over time earning the trust of its human colleagues and managers, by being transparent and inquisitive, and despite its extraordinary processing power and capabilities, staying humble that the user has a lot to teach!