If you’re a futurist entrepreneur building next-gen AI for functions the rest of us can’t yet imagine, perhaps stop reading now; this has nothing for you. This is for the rest of us, grateful for the help in our everyday lives. But how should we think of AI and how should we use it?

‘Way back’ in 2018 (!) Benedict Evans wrote that talking about machine learning “does tend to be a hunt for metaphors, but I prefer the metaphor that this gives you infinite interns”.

It now seems fashionable to imagine an AI-assisted future in which certain kinds of work (or indeed humans) are rendered obsolete, but we’re less good at appreciating the opportunities and limits of an AI-assisted present.

What’s in a Metaphor?

Well, this one’s actually pretty layered. (Benedict rows back from the intern analogy saying we’re probably only at the level of infinite fifteen year olds, but that was 2018, it’s safe to say we’re up to interns now.)

If you run a reasonable sized company, hire an intern, obviously the right intern will add value, richness and diversity to your team, and for them it’s a vital entry into professions, but if you were never going to hire an intern anyway, read on…

Interns are great when you want to delegate a certain type of work, but they don’t really know what they’re doing so you absolutely have to supervise and course-correct them. However, most of us are limited by the hours in our day and so delegating a good proportion of the time-consuming activity is incredibly valuable.

Two notes of caution: firstly, don’t get carried away by all the ways we’re told that AI can make us more efficient. The analogy I use here is shopping in the sale: if you didn’t want it in the first place, don’t buy it because it’s 40% off! If you’re changing your workflows to fit in with a new AI product just because it seems cool, it’s not helping you.

Secondly, there are things which interns can’t do (or at least they definitely shouldn’t do!). Remember, they don’t know what they’re doing so you can’t delegate the ‘cognitive load’ of decision-making.

Keep in mind what a large language model is actually doing; it’s not answering your question, it’s pattern-matching and showing you what an answer typically looks like to questions which generally look like yours. It doesn’t know what it’s doing. It doesn’t know the meaning of the words it’s using. Have you met an intern?

Not Creative, but Generative, Writing.

With the explosion of interest in large language models since the release of ChatGPT, their usefulness in writing has become increasingly evident. I have an email address which turns my off-the-cuff bullet points into a professionally worded email; it’s great fun but it doesn’t save me a huge amount of time. What does save me time is using ChatGPT as a writing partner.

If I have an idea for a writing topic but I don’t yet know where to take it, I have two options: I can go for a walk, or I can ask ChatGPT to throw out some ideas. I love walking and it’s hugely thought-productive for me, but I’m not convinced it’s the most time-efficient way to address this problem, not least because my mind tends to wander (again, I love my mind wandering but it’s not time-efficient!). Let’s also be clear, a large language model like ChatGPT is NOT going to do the thinking for me, it’s generating ideas for me to then think about.

This is when I should take my walk. I should think around the ideas, sketch them down in little mind-map diagram, and walk until something useful dawns on me. Then I have my topic, my ideas and my well thought out angle!

Next is the ‘blank page problem’. Maybe you’re already inspired and eight paragraphs deep, but maybe you’re staring at a blank screen still not quite sure where to start. Generations of writers have advised: just write, write anything at all, but again, if we’re shooting for efficiency, just get ChatGPT to write (or Bard or Notion but ChatGPT is by far the best I’ve found). Let it take a first pass, and you can take it from there.

One vital timesaver here is telling the large language model how you want it to write. I invested the time to create a prompt I can copy and paste every time. The prompt tells it to avoid cliches and hyperbole, to use a “fluid, conversational tone”, to use British English spellings. My prompt also asks it to curb some elements of its writing which make it sound like a large language model (or like an intern). Personally, I use the OpenAI Playground for which you need to set up an API key so I can also play around with the ‘Frequency and Presence Penalty’ settings but that’s far from essential.

However good your prompt, the first draft of the output is not going to be perfect. You need to review it and you ABSOLUTELY need to fact-check it. Remember, these large language models are just showing you what an answer typically looks like to questions which generally look like yours. They ‘hallucinate’, i.e. they make up plausible sounding sentences. They don’t care about truth; if you do, FACT-CHECK IT!

Bottom line: artificial Intelligence can’t (yet!) do your writing for you, it’s only as good as that intern, but it’ll turn a 90 minute job into a 30 minute job. Here’s the crucial bit: you need to do the FIRST fifteen minutes and the LAST fifteen minutes, the hour it saves you is in the middle.

If you’re writing content like corporate blogs, ChatGPT is also great at creating a handful of engaging tweets and LinkedIn posts (including the #hashtags) to promote it. An extension of this is A/B testing, e.g. creating calls to action; ask ChatGPT to write three or four CTAs optimised for effectiveness, test which works best and go that way in the future.

I also use it to overcome a couple of my cognitive biases: the first is to tell me what I’ve overlooked, I’m fine at reviewing what I’ve written, but awful at realising what I’ve completely forgotten to include. Secondly, I use it to create cold emails. I HATE writing cold emails and I hate sending them too. This may just be my personal quirk but once I’ve used AI to write them, I feel much better about sending them out; somehow it validates the approach!

Beyond Writing: Visible and Invisible Productivity.

AI-Assisted Search.

Since Microsoft released the AI-enhanced Bing, the combination of machine learning and web search has been promising. Remember I said above that large language models “don’t care about truth”? Well, surely in search results they must?! It seems not!

The two great advantages of Bing over ChatGPT are that firstly, it’s live on the web, as opposed to ChatGPT which was only allowed access to an historic training set with a cutoff date, and secondly, it cites its sources and includes their links, so it helps with fact-checking.

A great use case for this kind of intelligent searching is competitor research where you can not only ask Bing to compare you to your peers on whatever metrics matter to you, but also to format its results as a table which you can paste straight into your pitch deck… once you’ve fact-checked it, of course!

Text-to-Speech.

My life is such that it’s generally easier to spend an hour listening than it is to spend an hour reading. I might be travelling, walking between meetings, working out, or folding the laundry…  Speechify is incredible at reading text aloud, even in heavily formatted PDFs and even from printed books by taking a photo in the app. I’ve tuned the voice and reading speed to best suit me and it’s an incredibly effective way for me to absorb information.

Sure, an intern could do that too but it would be kind of awkward to have them stand beside me while I’m cooking the kids’ dinner, reading out IPCC reports.

I also use it to read back anything I’ve written (like this) as a quick and effective way of proof reading.

Document Querying.

I deal with a LOT of legal agreements. The thing about legal agreements is you generally have to read them all, otherwise you might find what you’re looking for in clause 12, and completely miss the fact that clause 23 starts “Notwithstanding clause 12…”. (Arrrgh!!)

Now my infinite interns, or in this case, AskCorpra can do a first pass, even if the document is the scanned execution version instead of being natively digital. Not only that, I can chat to the AI and ask it to perform tasks, like “list out the steps I’d need to take to terminate this agreement and summarise the consequences”. I can even take it a step further and ask it to “draft a notice of termination meeting the conditions laid out in the agreement”.

Of course, it’ll need fact-checking!

Hunting Down Contact Details.

This one feels like a dirty secret but I have an AI-powered plug in for my Chrome browser which finds the email addresses for people whose LinkedIn profiles I’m visiting. I’ve already said how much I hate cold emailing people, but it’s nothing compared to my dislike of connecting with strangers on LinkedIn; I get so much junk that way, I don’t want people to assume that’s what they’ll get from me (because it’s absolutely not!). So, I use FinalScout to find their email addresses and write to them properly.

Taking Meeting Notes.

This one still blows my mind! Whenever I’m on a video call, SuperNormal joins me, transcribes the discussion, and writes up the meeting notes, highlighting everyone’s actions. This leaves me free to concentrate on what’s being said and think up my high-value contributions! I only wish there were an IRL equivalent (which wasn’t an intern).

Calendar Shuffling.

Reclaim.ai is a handy little tool which I forget I even have running (surely, the dream!). It lets me schedule time in my calendar to get my head down on deep work, to take stock, to chip away at my to do list but if someone wants to meet me, it’ll just rearrange these personal slots to fit. I can tell it how ‘defensive’ I want it to be, e.g. if I have a very urgent task, and if I’m running out of time in a day, it’ll eventually refuse to move it. With team mates using the system too, it’ll arrange one-to-ones, and again we can define how much of a priority they are relative to all the other demands on our time (which are hopefully a bit less demanding now you’ve read all this!).