Preparing for a second industrial revolution

It's easy to get excited about the AI revolution. We’ve all heard the one about the future where white-collar workers are obsolete and advanced AI systems pull information from the sum total of a corporation's knowledge whenever employees or customers need it.

And that will happen. But what very few people are talking about is that, before the AI revolution can happen, companies are going to have to take a careful look at their knowledge infrastructure first.

What we've learned at Harriet, from building a real product in "non-hackathon" mode, is that, without orderly company data, you can’t take advantage of the massive efficiency gains AI offers. Fortunately, we're showing organizations with Harriet that AI can facilitate both the preparation and the application of enterprise knowledge.

Fundamental shift

The fundamental shift that will occur with the introduction of AI into knowledge work is in how companies think aboutknowledge preparation and application. Until now, most organizations have had a fluid boundary between these two domains.

Most organizations don't think strategically about how they create, refine, and share institutional knowledge.

The naive approach is to simply write information down when it needs to be scaled or shared. However, written documents are often just a substitute or supplement for some type of communication — a training session, a phone call, or a handover.

This lack of strategy leads to a situation all of us are familiar with: massive shared drives filled with files that have slightly different filenames to indicate different versions, authors, regions, or processes. There's no real overarching strategy, just a set of ad-hoc preferences.

Some professions are better than others at avoiding this trap. Engineers for instance use sophisticated version-control systems to organize code.

Assumptions waste everyone’s time

Although organizations don’t think strategically about knowledge management, they do invest an enormous amount of time and effort into organizing knowledge for operational purposes. Any document authored by a human represents a hypothesis about the consumer of that information and is created with a view to being consumed.

What this leads to is a tremendous amount of waste. Just think of all the onboarding, policy and training documents that employees are forced to consume. Think of all the time spent in training sessions or clicking through e-learning. A huge amount of time and effort goes into packaging information into delivery parcels. The trouble is that employees almost invariably ignore information if it is communicated at a time that it is not actually relevant to them.

The consequence is that when they do need the information, they need a reliable way of finding it. That’s a problem that a generation of enterprise search companies have tried, and mostly failed, to solve. Mainly because a solution requires exactly this "grooming" issue of the strategic issue of knowledge curation to be addressed.

Sources of truth

Some areas are sufficiently critical that organisations have tried to build sources of truth that reflect the totality of institutional knowledge in a way that is relatively easy to update. CRMs are a great example of this. Most organizations accept that it’s hard to do business without a clear overview of who your customers are, what your touch points are with them and who is engaged in the relationship.

When it comes to internal operational data, the dedication to efficiency and consolidation often falls apart.

Barrier to AI

Initially we thought that this would be a problem. Internal document management is, with rare exceptions, not a burning platform in a lot of organizations. Our finding from speaking to customers was that even those who were excited about subscribing to Harriet spent a lot of time testing her, partly just to make sure our AI works as described, but partly because they were genuinely unsure what precise information had been successfully loaded into her memory.

This can be a frustrating process. We found early on that Harriet would sometimes be fed documents - inadvertently - that said inconsistent things, were out of date, or duplicated information. This is the sort of thing that also happens to humans (“I can’t find the latest version of the travel policy but I’m pretty sure this version is mostly still appropriate”). We found that these sorts of issues made buyers realize that they had a bigger job to do before they were ready to adopt AI - first they would need to clean house. A huge, deadline-less task routinely kicked down the road.

If everything has to be in a perfect state for you to adopt AI, you're never going to adopt AI. But don't worry, it turns out this is exactly the sort of problem AI itself can solve ;)

Knowledge management

A quick digression on knowledge management. Historically, knowledge was managed in big buildings with lots of books: libraries. For a really long time the card index was the height of sophistication when it came to organization.

The metaphor of the printed book has been very resilient. Many documents continue to be essentially a big string of text optimized to a greater or lesser extent for printing on dead trees. That’s how my computer shows me the window that I’m using to write this post.

Who remembers Knol?

Lots of people have had the insight that the way humans process information is not particularly similar to the way it’s stored in books. For instance, Knol was a failed Google service launched in 2008 that tried to define a “unit of knowledge” (the eponymous “knol”) that could be linked together in a network of other knols. Sort of like a more granular version of Wikipedia.

The failure of Knol was overdetermined - among other things, Wikipedia was pretty good back then and the case for an alternative unconvincing. One challenge was surely that humans struggle to create very focused, single-issue pieces of knowledge without wider context but rather like to create narratives that weave together lots of fact in a structure that is intended for a purpose. That’s because when humans consume information, we find it extremely difficult to do so in the abstract.

Your personalized knowledge graph

We have found that they key to buyer and user confidence is both up-front organization and taxonomy and demonstrated ability to answer questions.

What that means in practice is that we use AI to digest artifacts (documents) to their component parts (information) which is then:

1. Communicated in a human-digestible way (a dashboard of your organization’s information, gap analysis and checklist of inconsistencies across your knowledge estate).
2. Made available to an AI agent that can easily locate relevant information and apply to specific user questions dynamically

How we apply this at Harriet

Armed with more and more capable models such as Claude-2 and GPT-4, we realized that in fact we could both help our customers get confidence in Harriet while also giving them tools to unlock their broader knowledge management challenges.

As part of Harriet’s onboarding process, Harriet now performs an analysis of all the documents she is given, cross-referencing them with common policy and process areas to give you a bird’s eye overview of exactly where your company’s policy posture has gaps.

She’s even able to tell you where different documents contradict each other or give suggestions for how to improve policies with gaps.

Think of it as digesting your organisations information before reconstituting it as knowledge - which can then be instantly applied to real situations via Harriet’s Slack chat interface.