Five ways you can use AI in your digital business right now
Unless you’ve been hiding under the stairs since before Covid, you might think that even pizza now comes with an extra topping of AI.
In fact, I’ve been asked more than once whether I could use AI to produce this tech blog. And it’s true that ChatGPT is marching ever closer to Turing Completeness.
But while certain professions may be transformed quicker than others, small matters like accuracy and nuance mean we aren’t quite ready yet to post content generated solely by the algorithm fka Bard.
No, it’s behind the scenes where AI is really in a position to change the world. Your children may not be taught by robots next week, but their human teacher will certainly use it to take the hard work out of lesson-planning. Creatives may laugh at the current crop of AI-generated ads, but technology is already helping them boost their own productivity.
Testing time
We decided it was time to put today’s AI-powered tools to the real-world innovation test.
If you’re a software engineer you’re probably familiar with the idea of hack days, where teams of programmers compete to create the coolest working prototype in a few hours. Even the NHS is running hackathons these days.
It’s not something that happens every day at a busy SME, though. It’s hard to justify the time and cost of a project that might or might not deliver something of value.
Still, that can make it hard for people to keep their skills current. How do you use your creativity to solve novel problems, if you’re lean, efficient and always-on?
Our Exploration Week brought together 60 developers, data scientists, product managers and UX specialists to get under the hood of AI tech and produce something the business could use.
Here are some of the things they came up with.
1. Write the product manual
Everyone loves well written documentation. Few of us love writing it. AI may not yet be a reliable blogger but surely it can understand and précis some code?
We fed ChatGPT a simple codebase for our newly-launched Product Protection offering, which gives customers the opportunity to purchase supplemental protection for their kit, and asked for a functional description. I’ll let it describe the application we chose:
The provided code defines a set of Python classes that model a system for handling insurance bookings.Sounds good but a little technical — we’d like to create text for our onboarding system.
You: Describe for a non-technical user
© ChatGPT: This Python code defines a set of tools to manage insurance bookings. Imagine it as a system for handling your insurance policies behind the scenes.
1. Booking Details:
The code creates different classes to organize information about insurance bookings. For example, it keeps track of details like names, dates, and policy information.
[etc]ChatGPT generated some pretty decent text-based ERD notation, too, though asking for a complete diagram seemed to bamboozle the bot.
To develop things further — ie ingest an entire codebase — we moved to Vertex AI. This was a quick solution to get up and running because we use GCP hosting and aren’t so rate-limited.
We wanted to create an automated workflow to allow updates and incorporate human subject matter experts to correct inaccuracies.
That involves feeding in prompts via Jupyter Notebook, feeding our repo into Vertex, then generating our documentation in Markdown.
When the Markdown drops into a monitored GitHub repo, our human editors are automatically alerted. The edited Markdown is then fed back to Vertex so it can learn and improve.
And here’s a sample of the end result, ready for use in interactive documentation on a different part of our codebase:
answer = answer_question("Provide a functional description of the Allocation service",
vector_store,
print_prompt=False,
)
print(f"{answer}")
The Allocation service is responsible for managing the allocation of products to transactions. It provides methods for allocating products, deallocating products, and checking if a product is allocated. The service also handles the process of converting reservations to allocations.Looking ahead, we’d like to integrate this into Slack, enabling direct, natural-language querying via a Slackbot.
2. Regression-test your visual design
Can machine learning spot problems with web page layouts? It certainly can.
We puny humans can’t physically look at every page of a website whenever we deploy code. Even if we did, we’d almost certainly miss things.
MPB has plenty of automated tests alerting us to issues with our business logic, but unexpected snarls in page design are trickier to detect.
So we put Vertex AI to the test, training it on screenshots of our home page at various screen dimensions. To keep things simple we overlaid bounding boxes around key elements such as site logo, first-level heading etc.
Training the model was a matter of adding images until we were happy it could identify every element correctly. In this case that took 131 screenshots.
Our model then had to decide whether two alternative versions of the page were OK or wrong. And this it did perfectly, spotting all the missing and moved elements.
There’s an elephant in this room, unfortunately. Training the model to detect changes on this one page took more time and resource than simple visual inspection. Much more.
It might make sense on a page where error-checking is too important to leave to humans alone, but for most purposes we’d probably be looking for an out-of-the-box solution.
3. Give Tech Support a makeover
What if a chatbot could give your people an IT helpdesk that’s always on, happy to help and knows exactly what they need? One that learns and improves with every call?
And before the tech support team starts looking worried, what if they had more time to focus on fixing, planning, specifying, installing … or even firefighting?
Our IT knowledgebase isn’t a single entity — there are different repositories depending on the context — but a chatbot offers a single front desk accessible to all.
We tested Custom GPT against Vertex Conversation, creating custom support chatbots based on MPB’s internal data sources. Each was tested against:
1. Typical IT support questions from our knowledgebase
- How do I raise a ticket?
- How can I book a desk?
2. Less straightforward questions
- How can I access dashboards?
- Where can I learn about fishing [sic] emails?
3. Completely unrelated questions
- How do I cook an egg?
GPT gave strikingly natural-sounding answers in its familiar breezy tone of voice, which were helpful, accurate and on-topic. Bot behaviour and rules are customisable and highly flexible. It dealt well with vague or misspelt questions, and politely directed us elsewhere on the unrelated questions.
Vertex is hugely customisable, but requires more manual input to set up. Its answers were a little techie and less natural-sounding than GPT’s. It’s aimed at developers, which probably explains its different approach — it shines on very specific questions and provides lots of detail in its answers, but struggles with vagueness.
Unless you have an enterprise level of support requests, this isn’t going to be an expensive way to run your first-line support. GPT can be integrated via API, so there are options for bringing its functions into tools like Confluence or Slack.
4. Give your helpbot sales training
Online chatbots have been around on the web since at least the olden days but it takes time to fine-tune them.
We wanted to increase the resolution rate of MPB’s chosen AI chatbot, Fin, and reduce the burden on our wonderful Customer Experience humans.
We started simply, using the bot’s Custom Answer capability to create a workflow for questions about transaction status. The user inputs their transaction code and Fin provides an update, where previously a support call might have been needed.
Time to up the ante. Could MPB generate an offer to buy some kit from the user, entirely within Fin?
Again, yes. All the prompts appear within the bot and link up with our platform’s back end, known as Re.Serve, to complete and record the transaction. The user receives natural-language answers, complete with transaction ID, and email confirmation works as usual.
For an even more complex (and real-world) scenario we trained the bot to rebook a missed collection. That involves arrangements between three parties (user, MPB and courier) and often requires human intervention in today’s world.
Once again all stages of the process were completed, succinctly and successfully.
Security matters here and we’d need to add customer verification before this is ready for release. And there will still be tasks that need human intervention.
But we did enough work to predict a huge (~66%) improvement in Fin’s resolution rate from these few changes.
5. Take control of your tickets
Even for the best Product Managers, staying on top of every feature request, bug fix and enhancement in the ticketing pipeline is hard work. At worst, a chaotic Jira queue can lead to lost requests, duplicated work and a whole spongy morass of wasted time.
One of the pain points in any such system is getting tickets to the right team for resolution. Surely a well constructed AI could cure our pain in a nothingth of a second?
We trained an AI model on historical Jira data, based on words used in the Ticket Summary and Description fields for each ticket and using BERT to perform sentiment analysis.
This resulted in 90% accuracy when assigning tickets to one of four Scrum teams — a definite time saver for the human making the ultimate decisions.
And finally …
If you’re looking to dip a toe in the waters of AI, I hope this has helped demystify the topic a little.
It’s certainly not yet a mature set of technologies. We have some very capable AI engines at our disposal but some of the interfaces have yet to catch up, while rate limits make everything harder to explore for non-enterprise users.
All the same, there are some very usable technologies here. I suspect the real bang-per-buck will be behind the scenes, not just for our company but in the wider world. We are also looking at how we can use emerging technologies for the benefit of our customers on the mpb.com platform.
For example, I just created a job description via AI. It was about 80% right, so I still need to review and rewrite; I just have more time to focus on other things. Hospital consultants won’t be replaced by AI diagnosis any time soon, but they should be able to help more patients, more efficiently.
It’s worth reflecting on the downsides of this technology, too. It requires huge compute power, which will have an impact on energy use. And LLM-trained bots are creative by nature, which means they often prefer a wrong answer to no answer.
Sophie Davies-Patrick is Chief Technology Officer at MPB, the largest global platform to buy, sell and trade used photo & video gear. https://mpb.com
