The world is all a-tither about AI. Nearly every company, product and conference is introducing an AI play or feature. If any product might see an advantage from having something summarized — you can bet that you’ll see a ✨ emoji appear to give you that summary. I’m sure that ChatGPT is bored of summarizing customer support tickets by now, stripping the human context out of the communication in the process.
When it comes to money, doing things carefully is paramount. Proof-of-concepts for “robo-advisors” have popped up frequently, years before the advent of the current large-language-model-based versions of AI. However, the phrase “this is not financial advice” has become a deep part of cultural vernacular in the past few years. And for good reason — bad financial advice leads to people losing money.
So, we’re taking a different approach. We’re still unsure what PocketSmith should use AI for, and we want to do something different from summarizing. So — we’re inviting you to play with the investigative tools we’ve been working on, currently named “AI Tasks” in PocketSmith.
They’re a way for you to interact with OpenAI’s ChatGPT from within PocketSmith. We’ve captured a few obvious use cases: assisting with categorization, cleaning up merchant descriptions, identifying repeating charges, and other analyses.
Further, we’re opening up your ability to interact directly with ChatGPT by allowing you to edit the prompts directly and other key variables, to experiment with your own data as you see fit. This means there are no limitations to what you can do with AI Tasks — edit a preexisting prompt to suit you and send it to ChatGPT.
We want you to explore what AI can do within PocketSmith — so much so that we also feature an “ask anything” AI Task, which allows you to send any text or data to ChatGPT for a response.
So, if you’re a beta user, click the 🤖 emoji on the Transactions page toolbar to access the AI Task view to get started. Yes, we’ve decided to eschew convention and not use the ✨ emoji!
We’ve been working on this gradually for over a year now, and are only happy to release now because of the drastic increase in quality of ChatGPT this year especially — up until now, ChatGPT’s outputs were interesting, but not particularly useful. Then with the release of GPT-4o a couple of weeks ago, we’ve seen a further step-change: AI Tasks produce more accurate and reliable output than ever before.
Previously, a lack of reliability and consistency of output meant that what came back to PocketSmith couldn’t be applied directly — a request to categorize transactions with specific instructions would result in different outputs that couldn’t be properly accounted for.
But now this has changed, so it’s time to release the tools we’ve been building to you all. It’s worth noting that the AI Task user interface is very bare-bones and unpolished, focusing on function over form. Please excuse the rough edges!
Firstly, information sent to ChatGPT via their API is not used for training their large-language models. This means that OpenAI does not use the information you send beyond providing you with the output of the task that you requested.
Secondly, all AI requests and responses are stored solely in your internet browser, in local storage. PocketSmith does not store this information, and it doesn’t appear in our server logs.
Third, you can wholly opt-in and control this feature. If you’re uncomfortable with using AI Tasks, then you do not need to do so. This will always be the case — we’ll never activate a feature where your data is sent to an AI without your express permission. You’re in control — even when things progress, and we automate suitable AI Tasks.
The most important step is receiving feedback from our customers about the AI Tasks as they stand now and hearing ideas about what would be helpful for them for AI tasks. We have achieved consistent results with the prompts we use on our own data. However, the primary reason for getting this very early version of AI Tasks released is to hear from you about how we can use AI to improve your PocketSmith experience.
We’ve set up a dedicated email address for this – just send your feedback and thoughts to [email protected], and your feedback will get directed to the appropriate team.
Once we’ve received feedback, tweaked default prompts, and added new functionality, we’ll likely rebuild the AI integration into something that can be triggered automatically. We’d like to see a world where you can either manage your categorization or never categorize another transaction again because an AI does the heavy lifting for you — and trust that the AI will accurately organize your transactions.
After acting on feedback and introducing automation, we want to expand our horizons. Our users have already been building amazing uses of AI with PocketSmith data, from automatically generated daily budget podcasts to creating a personal CFO using PocketSmith and ChatGPT. As always, we’ll consider ideas carefully to ensure any feature released works well for a broader user base.
Please become a beta user in PocketSmith, hit that 🤖 button on the transactions page to start exploring, and send feedback to [email protected]. We look forward to shaping how AI is used in PocketSmith alongside you!
James is the CTO and co-founder at PocketSmith. He loves tech from software to hardware to music, and is passionate about technology being a net-positive in people’s lives. He lives off-grid with three humans, one axolotl, one rabbit, one dog, and too many possums.