In case you missed it, we've already written an article defining what artificial intelligence is! This article will go a bit deeper.
Classify All The Things! Harnessing AI for Records Management
- What is AI?
- Artificial Intelligence at its foundation is algorithms designed to teach machines to learn and make decisions.
- It's all about machines learning from the data we give it, and be able to make basic decisions which means automation for us.
- The goal is just to make some decisions automatically.
"What is now"
The foundation of all of this is machine learning. Without machine learning, the machine cannot learn to make decisions. It needs to understand how to learn first, then it can make decisions.
What we focus on with machine learning is:
1. Generating Data Models:
- Directly coded (via statistics) (in the world of insurance or finance, based on stats)
- Trained from data sets (this is where machine learning really shines: finds patterns in seemingly unrelated sets of data)
Once we've got that, it's about executing algorithms.
2. Execute Algorithms
1. Run input through model(s)
2. Extract new data
3. We then use the output data to either enrich or predict.
That's the foundation, we haven't made it to AI yet.
Three Types of Machine Learning
(twenty types, but these are relevant to those working with data)
1. Supervised Learning: basic yet powerful. providing the system with fixed data and training the system to understand what it's looking at. Can then input new data and have it action on it for us. Example: use invoices and train the system to understand "what is an invoice?" Looks at structure, by content., etc. Builds out mathematical model of common elements between these documents (vector space). Goal here is to train on invoice, so machine will identify the content as an invoice and extract the data is needs to, and put the invoice through processing.
2. Unsupervised Learning: slightly different approach: we give it data but no information. So it will look for commonality, and it will cluster (ex: may cluster invoices via products, by vendor, etc.) without it being trained.
3. Reinforcement Learning: concept is prediction. Take data (ex: corpus of invoices), but the objective is predicting what future insights we can extract from our data. Ex: invoices and ask what vendors are willing to provide additional discounts (show right and wrong, so it can become highly predictive)
Supervised data (Collabspace example): machine learning allows a basic "trained model can start to identify things that are meaningful in your data sets" (entities in Collabware Press Release). Also identifies the Sentiment- uses vector space- to calculate the positive/neutral/negative.
Can be done with images (supervised learning- engine has been trained to identify people, table, etc.) so it can relate what it's seeing. In unsupervised sense, can identify colors (how many?) and commonalities between different images.
Can be done with audio, not only words spoken but who is speaking. "voice print differences" unsupervised is grouping common voice together, and using machine learning to extract and understand what is being said (supervised). How unsupervised and supervised go hand in hand to work together.
What is AI?
AI is when you take that learn, generate new knowledge based off of that, and start making decisions and taking action based on that.
This looks like:
Generate Knowledge:
- Integrate Machine Learning models and algorithms (not just one, need multiple)
- Capture input and make decisions
- Reinforce with decision outcomes
Then...
Making Decision and Taking Actions
- Requires enough data/models to ensure confidence
- Allow for adaptation to varying inputs
- Account for foundational bias
The diagram below (from UBC CAIDA) is a fantastic representation of how it all ties together.
Machine Learning covers our data, and must add more and more as we go.
What is the business case for applying AI to RM?
RM: Automation and Transparency
- Classification and Categorization
- Reduce manual effort and bias/mistake
- Leverage rules-based or Supervised Learning (or both!)
Data Enrichment and Discovery
-Identify and extract valuable data
- Tag and enrich with new information from Machine Learning
- Efficiently identify specific data or patterns in large data sets
IT: Security and Supervision
Threat Identification and Mitigation
- Predictively identify suspicious or dangerous activities by pattern
- Take action automatically based on requirements
- Examples: Identify Phishing emails and quarantine
Reduce Administration
- Automated categorization/classification reduces user support
Executive Management: Insights and Predictions
- Connecting Dots
- Mine protected data for new insights and patterns (based on new knowledge)
- Surface related content in context
- Example: Surface similar proposal responses to RFP to support Sales
Predicting Outcomes
- Apply predictive models to operational data
- Example: Predict potential machine breakdowns from maintenance records
Future State: Where is AI going?
Data Enrichment
- Develop layers of information from data
- Tag data with that information to encapsulate knowledge and context
- Discover enriched data more easily and in context (eDiscovery/FOIA)
Virtual Assistant
- A tool to quickly query all data and knowledge
- Allow execution of basic actions
- Answer questions
- Book meetings, locate relevant data, and capture attendee output
Practical steps to apply AI to RM today
Build a Foundation
- AI Program/Policies
- Align with the mission
- Identify high-profile/high-value targets
- Address challenging issues like biases and ethics
Data Lake
- Tear down system-specific silos -> stream in data lake
- Removes source system technology barriers
- Protects data so it can be used to build models
- Enable data to be computed at scale
- Flatten structure and plan for enrichment
More practical steps in presentation, available to listen to in the presentation or in the session slides below.
You can also check out our whitepaper on this topic:
