While there’s little doubt there are now plenty of opportunities to save folks time using recent advancements in AI language models, approaches to actually implementing AI to these ends are still evolving. Presently, the most common framing of AI products by far are as “assistants”– an approach directly inspired by the conversational UX of ChatGPT, and one that is particularly potent in quickly communicating the value of AI-powered features to new organizations.

These assistants are remarkable for their ability to handle a wide variety of tasks, with pretty much the same information we’d give quite naturally to a colleague. For ad-hoc needs– for the work that’s the exception and not the norm– this is perfect, but for really making everyday work that much easier– we’re not as sure. There’s a particularly good question, we think, to be asked of how variable the work we do on a daily basis really is, and where the best solutions to push forward with AI really lie. We think it’s quite likely AI assistants are and will remain best for identifying solutions to problems, but not necessarily 

 them– we’ve always had a 10x cheaper, more scalable, more maintainable, and more deterministic way of doing this– code.

Before we continue this thread though, let’s zoom out a bit. Why should we care about exactly 

 work gets done? Whether an AI does it, or a particular script is run, is there really a meaningful difference in the end?

This has to do with how important the proof certain actions were taken is to the end result. Naturally, there’s a great difference here between actions that help read and visualize data, and actions that execute difficult to reverse transactions on outside systems– refunding customers, deleting data, automating correspondence, etc. This is why the vast majority of chatbot assistant products on the market currently are tools for visualizing and understanding information– there’s far less inherent risk in representing a spot of data to you incorrectly as there is in actually taking an action that affects the real world, especially where incorrect actions could butterfly into far larger and complex downstream issues. Conversely, however, this might very well be where some very meaningful applications of AI lie.

, actions that actually impact people and production systems, far more determinism & auditability are called for– a 99% chance of getting the right thing done at the right time usually won’t cut it. To work around this, automation-focused assistant products tend to adopt a simple human-in-the-loop confirmation before such actions. While there’s still always some risk of human error, this kind of “sign off” is without a doubt a necessary step in allowing AI to approach higher value tasks that would otherwise require much more human time and supervision, and allows AI tools to start being adopted in places with high time and effort savings in larger organizations.

But is the future really just asking computers politely every time you need the same old thing done?

, but natural language, no matter how well we master it, is a deceptively unintuitive surface for providing inputs to machines. In the time it takes to explain and approve an action on a particular system– I might as well just log on to that system myself and click a button. Purpose-built software provides me with structured, validated inputs, with no context or response size limit I have to concern myself with. I can take deterministic actions on large amounts of data, and can process charts that don’t have asterisks everywhere warning me of AI unreliability. It is silly to dismiss the permanence of this experience, even in the advent of this exceptional new technology.

This highlights an unobvious, but important upper bound for where AI in automation can be useful– the value of chat might be maximized when the problems, topics, and issues being worked through are in constant change, meriting the activation energy to put into words the context and nature of a task at hand, but is almost certainly at a global minimum when you need to repeat the same thing over and over. Critically, this is in direct odds with how work inside most organizations looks over time– tending toward less chaos usually gets more things done– and it’s always been our prerogative to measure, reduce variability, and make unexpected scenarios less common.

Admittedly thinking particularly far ahead now– we imagine it is for this reason that, 

independent of progress on large language models

, the best version of AI as an force for automation mandates a fluid transition between identifying solutions to problems via a chatbot/assistant-like interface, and making those solutions as reusable and as easy to use as possible via a more traditional & portable software interface. And for as long as we use computers with screens and buttons and mouse cursors, that this is overwhelmingly likely to remain the case. It’s not nearly as sexy a sell as an assistant that just “does the work”, but we think it is magnitudes more valuable, pragmatic, and grounded in long term thinking to use AI to build reusable tools for higher volume tasks than to rebuild the wheel each time.

A critical point, and why it might be time to stop copying ChatGPT.