Ask an accountant: How you’re testing out AI in practice

The various uses you’re finding for different kinds of AI, learning to tailor prompts for accuracy, and why the software matters.

We spoke to Alexis Charkiw, Owner, Right Click Accounting about how she uses Artificial Intelligence.

What are you using generative AI for?

I’m using it to vibe-code* tools I’ve wanted to provide to clients for years, but couldn’t. Usually because it wasn’t viable as it was a value-add and not revenue generating, so the cost outweighed the benefits.

I also use it for checking tone of my emails, support with SEO terms, Canva graphics via the Claude connector plus lots of other small but effective tasks.

*vibe-coding is where a user develops ideas in plain language prompts, and the AI creates the actual code needed to deliver the idea.

What are you using agentic AI for?

I’ve recently started using agentic AI, but adding meetings in my calendar with contact, email, video conferencing and locations has been really time saving. I’m not ready for emails as I’m still in the habit of wanting to reply, but I can see how in time this would be really efficient.

Which tasks or areas do you especially recommend AI for, and why?

I suggest using it for areas where you have the most weakness or least confidence. For example, I dislike writing about myself and struggle with anything to do with speaker bios, copy on the website or marketing adverts. So I use AI as a prompt to help generate ideas, and get my writers block off the ground!

Conversely, which tasks/areas would you avoid using AI for, and why?

As mentioned, I’m avoiding emails. I embed things mentally when I write or type so writing emails feels like it helps process my thoughts. I also don’t write any of my speaking content with AI. It’s my raw material and I deliver it in-person so I really need to know the content and context.

Does the AI software matter to you?

Yes, we use Claude and CoPilot for all our AI needs currently. I know the others have their place: I used ChatGPT in the very early days, and I’ve used Gemini perhaps a handful of times.

But for now, I prefer the ethics of Claude and the security measures of CoPilot. We use the Microsoft ecosystem so using CoPilot means we keep as much data in one product as possible.

How do you feel about AI being built into software (such as Xero and alternatives)? Does it help string things together smoothly or produce a bit of a black box effect where results can be hard to fact check?

This is an interesting question because I think it may be too early to tell.

At the moment things like Smart Capture in Xero are reflective of Hubdoc or Dext, whereas the building of Agents in JAX (as announced at XeroCon last week) is a different way of working. Until we build out our own agents and test them, we won’t know how much of a positive impact it will have on our workflows.

I do think any change implementation has to be dealt with carefully and one of the biggest challenges is going to be getting finance professionals to work with it willingly and from a place of positivity as opposed to a place of fear.

How would you go about choosing agentic AI software and setting it up? What’s the process of training it like?

Honestly, I’m just testing it for myself. I do follow a number of AI experts, especially in the US, and watch what they’re using and why.

For my priorities right now, I want to speed up laborious tasks. I have been recommending Anthropic’s free training to anyone who asks. Best to use the training form the platform itself at first, especially as it’s free and the most up to date.

However, I have dabbled in the Canva connector and Higgsfield for marketing recently, and I’m just playing with it to see how it works. It reminds me of when the mobile came out with various apps, there were no proper instructions, it was just ‘plug in and play’. I feel AI is like this in its training nature. 

How do you tailor prompts to ensure accurate and useful outputs? What kind of language do you use?

My style of prompts is to use it like an email. It’s structured, bullet points, expectations and I tell it, ‘if anything is unclear, ask me’. I know others use it in a chattier style but I find this the most effective for the outcomes I’m working towards.

When inputting information, do you find it best to do one large prompt with lots of detail, or break information down across several queries?

I find smaller shorter, more detailed prompts to be the most effective. At times, it’s given me an odd answer and when I’ve read it back, it’s because my prompt wasn’t clear. So I’ve edited to start that part again as opposed to trying to explain myself.

Cat Hall is Content Specialist at AAT.