Scale AI documentation the “right” way

AI Governance & Assurance
Risks & Liability
scaling ai documentation - image

Avoid common mistakes when it comes to model documentation

Documentation is considered the bane of corporate existence by some. For others, a great documentation system makes their organized minds super happy. In a few instances missing documentation can be a minor inconvenience - in others, missing a digital paper trail and/or knowledge base is seriously detrimental. This is the impetus for the question many business leaders are asking today: how do we step up our AI documentation? 

A secondary question, that may or may not be top of mind - but should be: how do we scale documentation in a way that is enabling for teams across the enterprise?

As with most technology, the adoption of AI and machine learning has been much faster than the implementation of governing practices for AI and ML - including model documentation.

Many organizations today are using one of the “bandaid” solutions below for documenting ML models, controls, and other important information relating to AI. None of these solutions are sustainable for scaling enterprise AI.

Five places your ML documentation probably lives now (but shouldn’t)

  1. Documents (Word, Google Docs, etc.)
    Using a digital word document tool usually leads to disjointed files, user access, and modification problems. It’s also not very enabling, as far as setting up (and following) best practices for model governance.
  2. Spreadsheets
    “I could use another spreadsheet in my life” said (almost) no business professional ever. Spreadsheets are great for some functions, but they’re not nearly enough for creating a true lifecycle model documentation system.
  3. Email
    There are many issues with using email to discuss and document important information related to AI and ML initiatives. The top one is security, but what about the fun of finding email trails from a person that is no longer at the company?
  4. Chat tools (Slack, Teams, etc.)
    Slack and other chat tools have been game-changers for remote team collaboration over the past couple of years. But we probably all agree chat tools are not appropriate for important approvals and documentation. It’s easy to get caught up in the busy day-to-day and just use what’s in front of us, but this is not a sustainable practice.
  5. Notebooks (Jupyter, Colab, Zeppelin,  etc.)
    Model builders often place comments and key pieces of their model building process in notebooks and integrated development environments where the models are built. A fundamental problem with this practice is the notebook and relevant information doesn’t travel with the model to production. The context is lost, and business partners don't have awareness of this documentation, nor the ability to easily navigate or understand, even with access.

It might sound dramatic, but going beyond the basics with your machine learning documentation can set you on a path toward excellence - while crushing your competition in the meantime. 

Reaping the rewards of good AI documentation

Below are some of the benefits to using a scalable software solution built for ML documentation.

An organized, clear system of record

The problem with using any of the five options above for ML documentation is that those tools are not enabling for users and stakeholders. Plus, they’re not a real system. Enterprises using disjointed documents, spreadsheets, one-off chats, email, or notebooks, are not only putting the business at risk. Undervaluing the importance of ML documentation also slows down teams and cuts productivity across the company (data science, risk management, IT).

Implementing a documentation system can help teams govern AI and ML by enabling the following (at scale):

  • Create, organize, and share the key information business and risk partners want
  • Proactively share expectations as far as what documentation or governance practices need to be followed
  • Have a centralized, standardized process for good documentation
  • Have consistency in how documentation is created and maintained
  • Maintain tracking of user activity to support accountability

Tools for collaboration and automation

Using the five tools above for ML documentation and model governance usually means opting for a short-term convenience over a long-term strategy geared towards excellence. When doing so, teams are missing an opportunity to implement a more robust tool that provides:

  • Alerts and notifications, for example, when a model is due for review
  • Ability to create tasks and workflows
  • A system that is scalable with options for automation
  • Options to easily share documentation and evidence across teams, and with third parties
ML documentation concept - image

Integration with other business applications

One main benefit of integrating other applications with your ML documentation is the ability to provide (and easily access) key evidence related to models. For example, integrating tools for data governance, model monitoring, and model serving helps build important proof that best practices are being followed. 

Leveraging these integrations also enables automations and workflows, and decreases having to rely on manual input.

Using one of the five tools above prevents teams from creating the connective tissue between ML documentation and other key technologies that relate to and complete documentation for models. 

A competitive edge 

“Better visibility” is a phrase that’s probably used too much in the realm of enterprise software - but it’s mentioned so often because it’s a central business need. Leveraging a software system for model governance and documentation gives teams:

  • Clearer insights into ML performance
  • Faster production times, as a result of streamlined communication and governance documentation
  • Audit trails to easily share information cross-functionally and with key stakeholders

Building models is already hard enough - documentation and visibility into not just the systems, but the thought process and sets of decisions made when building the model are so important. Model building teams deserve purpose-built solutions to enable capturing these key pieces of the journey. Good AI deserves great governance tools.

Early adopters that get model governance right will enable faster processes and better visibility - and ultimately gain a leg up on their competition. 

Getting in front of AI Governance

AI Governance is coming. Like most enterprise initiatives, the project of implementing good governance will be what your team makes of it. Taking the time to implement today’s top technology - for ML documentation and beyond - alongside AI strategy best practices will be game-changing for the companies that get it right. With the challenge of governance comes the opportunity to manage bias and risk, streamline processes in order to optimize for excellence, and set your team up to have strong advantages above your competition.