HelpFirst: October update

We’re chuffed to have received recognition for our work with the wonderful folk at Citizens Advice Scotland! GO Awards named us as Winner for ‘Social Value - Other Organisations’. Digital Leaders 100 named us as Runner Up for both ‘AI Innovation of the Year’ and ‘AI Expert of the Year’. Keep an eye on next month’s Smarter Working awards where we’re shortlisted in 4 categories…

We’re continuing to roll out ‘intelligent case notes’ at Citizens Advice Scotland’s Extra Help Unit, and we’re now up to 170,000 requests to the API since the start of June! Read all about the project here.

Meanwhile, we’re coming to the end of a proof of concept project with another wonderful client. We’ve been exploring AI’s potential to support with:

1. writing up a summary of what happened in an interaction between a counsellor and a service user, and

2. selecting the key topic(s) of the conversation.

We’re seeing exciting indications that large language models, with a human in the loop, could enable these to happen more quickly and accurately. The future impact?

• More people who are struggling get through to a counsellor, more quickly.

• Counsellors better able to focus on what they’re there to do, rather than admin.

• Better insights into where the service is most needed, and emerging trends.

• Better informed advocacy.

Excitingly, we’ve been selected for the funded Exploration stage of CivTech 10. We are partnered with Create/Change to respond to the challenge “How can technology help make teachers’ workload be more manageable, enabling them to focus on the activities that add the most value to learners’ outcomes?” We’ll be developing our ideas together with the Scottish Government (Directorate for Learning), Aberdeenshire Council and Dumfries & Galloway Council in November.

If you’re wondering how AI might support your organisation but not sure how, HelpFirst might be able to help. Here’s how our AI Action Plan works:

1. Talk to us about your big questions around AI. We’ll have a couple of chats with you and your colleagues to get our heads around the context and your early ideas so we can jump straight in on the day.

2. Assemble 3-5 stakeholders from your side for a ~2 hour workshop, either online or (preferably) in person, together with 2-3 of our team. We’ll collectively assess and prioritise potential initiatives against potential time & cost savings, applicability to multiple use cases, size of project for an MVP or full implementation, and risks.

3. We’ll write all this up and send you a report of the workshop outputs, including recommendations based on our relevant experience and your specific needs.

We’ve been receiving rave reviews for these workshops (my colleague Angela is leveraging her McKinsey skills to the max)!

We’ve also been accepted onto the G-Cloud 14 framework, and the NHS Innovation programme! So if you work in the public sector and are interested in working with HelpFirst, it’s now easier to make that happen.

Book a call or email to arrange an initial chat!

As part of Digital Leaders Week, we hosted the session "From Overwhelmed to Empowered: AI's Real-World Impact on Citizens Advice Scotland". Catch up on Andy & Angela’s chat with Natasha Gilmour and George Holmes for free.

We also had the pleasure of participating in LOTI’s Adult Social Care Sandbox Immersive Day, as well as a Children and Adults Services AI Discovery Day hosted by North Yorkshire Council. One comment we loved: “It may be useful to think of AI as part of the team rather than replacing anything - imagine a new member of the safeguarding meeting that knows every bit of guidance and legislation or can read 10 years of case notes in 10 minutes”.

Coming up, catch us at:

• Government Transformation Week: We'll have a stand in their Innovation Alley (5-6th November, London)

• Helplines Partnership Conference: Andy will speak on a panel discussion on "Delivering Services in a Changing Society" (7th November, Birmingham)

• Homeless Link’s Leadership Summit: Andy will deliver a talk (4th December, Stratford-upon-Avon)

Give us a shout if you’re nearby and want to grab a coffee, or look out for us wearing our new HelpFirst hoodies!

Evan Harris (TermScout) has been tracking GPT-4o's accuracy on legal reasoning tasks over time. People say these models feel like they degrade after initial releases. They found the error rate was consistently below 10% for 6 weeks, then between 10% and 20% for a month, then back down to somewhere in between.

Iman Mirzadeh (Apple) and others have published research exploring whether LLMs can truly reason, or are just sophisticated pattern matchers. We wonder whether this is any different with humans - is the ‘reasoning’ we do also basically pattern recognition: observing, inferring and replicating? (See more on this from Santiago Valdarrama, Perplexity, and Google Research.) We asked our genius advisor and neuroscientist; he said ‘good question’, which made us feel we on the right lines.

Drew Breunig suggests thinking about AI use cases in categories of Gods (human replacement), Interns (for supervised tasks), Toys (like interns but error tolerance is higher) and Cogs (autonomous, reliable components).