AI assistants are readily accessible, and can perform a range of mundane and creative tasks, but oversight by domain experts is essential.

For start-ups and scale-ups like us, there just aren’t enough person-hours in the day. Embracing automation where helpful is a must. So I went shopping for AI assistance (tech shopping is the only kind I actually enjoy). Here I’ll describe some of the things I tried, and how that’s going…

Transcribing and summarising meeting recordings

As a scale-up we have many (many) meetings; taking minutes can be tricky and creating summaries takes time. Several products offer meeting transcription, but they’re either slow (Teams takes almost as long as the meeting itself) or expensive. I currently use Cockatoo but am overdue to review that choice (for example, otter.ai looks appealing). These tools are impressive: rare names and abbreviations tend to be misinterpreted, but in general the quality is good.

However, what we really need are summaries and action points. Again, several products do this, but we fall between pricing models (we have a lot of meetings, but we’re small). Many services charge by the token (roughly equivalent to charging by the word), which quickly adds up.
Finding the right tool for the right price is an ongoing process (recommendations in the comments are welcome!) Amongst others, I have tried Notion’s AI option, which didn’t perform as well as I’d hoped, and the free version of Claude, which was better (more on Claude later).

Brainstorming starter lists for new features

Whatever the topic, any of the free AIs (e.g. ChatGPT, Bard or Bing) will generate starter lists of concepts; for example, ‘please list the most important features for an events management system’ (I say please because I was raised right.) However, there is a better way: AI mind maps. I’ve tried Xmind and Miro (through free features), but we use Miro a lot internally, so I’ve stuck with that.

It’s simple: you give it a prompt (by naming the first node), and the AI grows you a mind map. This has left people open-mouthed and is probably the most fun way to show off AI in a work context. Plus, if you live in Miro (like I do) you can re-use and edit the maps for your other work.

Interacting with documents and collections (e.g. user manuals)

Many commercial AIs can be trained on a set of documents for internal and customer support. These work well, but as ever cost is an issue. I’ve tried trialware and free tools (Locusive and Chatbase) with some success, but this is still a work in progress (recommendations welcome here too).

We also work in many languages. Humans are still the only reliable resource for translation, but I was a little freaked out by HeyGen’s ability to morph a video recording of me reading part of ‘Fox in Socks’ by Dr Seuss into German. A native speaker confirmed it’s not bad, barring some pronunciation issues. You can see the results at their site 2 . (Incidentally, my partner said that it isn’t my mouth!)

Generating content for social media

Any of the freely-available tools will do a pretty good job of creating content, and I use them to draft social media posts, though ‘prompt engineering’ (finessing your question) is important to get really good results and any allegedly-factual content must be checked, as we will see later.
For the images to go with the text, many tools are available (including Miro), but you get what you pay for. Again, prompt engineering is key, including specifying the style (cartoon, photorealistic, pixel art, etc.). Perhaps the most famous – DALL-E – is showing its age, but the best are really very good (as long as you count the fingers).

Gathering market intelligence and performing SWOT analyses

All the main AI chat tools will give answers to requests for market intelligence and will even do a SWOT analysis. But this is where the big health warning comes in: confirm everything – never use the ‘raw’ results. Sometimes AIs just give you something that looks like an answer. They improvise without warning, and these ‘hallucinations’ can be dangerously plausible.

I also tried ‘agents’ (child AIs created by a controlling AI to perform subtasks) for more in-depth research, using AutoGPT. This brought its own challenges (knowing when to stop, extracting results). This was also not easy to set up. I’m sure things have improved since Summer 2023 though!

Doing basic data analysis (spoiler alert: this did not go well)

Social media is full of stories of AIs confidently stating that 2+2=5, or not solving the bat and ball test 3 (which to be fair I failed as well – draw from that what you will). So asking the free AIs to do maths is not advised, but even on language problems they can be unreliable:
I had a list of 2,000 names containing several near-identical duplicates, which I needed to clean up. For fun I thought I would also ask Claude, because it allowed me to upload the list (and for no other reason – this is not a critique – in general I like Claude a lot).

Claude reported eight pairs of probably-duplicate entries, and even gave a rationale for some of the matches. So far, so good. But four of the 16 names it had ‘matched’ weren’t even in the original list (though they looked plausible, making them easy to miss).

I replied, ‘Several of these sites were not in the original list – can you explain what happened?’; it apologised and replied with five sites it said were ‘not originally in the provided list’, but three of those five were in the original list, meaning that it also missed two of its own prior mistakes. AI was clearly not the tool for that job, and that’s fair enough. Not everything is a nail.

In closing, I must stress this piece represents my own, limited personal experience. It is not an exhaustive survey, and I am not an engineer. Your mileage will certainly vary.

Conclusion

Small businesses have a lot to gain by using AI tools, within limits and with caveats Domain experts (knowledge workers) aren’t obsolete yet; these tools still require oversight

Footnote: for this post, I thought I’d go meta and get an AI to start me off, but after an hour of prompt tweaking (using the main online bots and a couple of mind mappers) not one had suggested transcribing and minuting meetings – my most frequent use case. Oh well :)

References:

CYTE is a scale-up that has developed an online platform for global clinical trial management (CYTE Connect – https://www.cyteglobal.com/connect/)
2 https://app.heygen.com/video-translate/share/44e2e040cd73427cb0fb2bf2673cda8b
3 https://en.wikipedia.org/wiki/Cognitive_reflection_test

Join Network