AI in public affairs: what we've learned, and what we're still asking

This article has been co-authored with Stonehaven Head of Data Luke Betham.

From precision tax enforcement in Italy, machine-driven fraud detection in Britain's benefits system, and AI assistants handling citizen queries in Estonia, AI is reshaping the way governments work. 

It is also redrawing how people and organisations interact with governments. This is the first in a series we are publishing on this. 

While governments are human-led, so is engaging with them

Alongside care, physical work and novel thinking (though this last one is changing fast), accountability is one of the things humans still do better than AI. Elected representatives hold democratic accountability for what gets decided, so those who seek to influence what gets decided hold accountability too.

That gives a simple rule. Where a human is accountable for the work, AI should support the human but not replace them. Public First's own AI Global Report (polling 18,000 people across 15 countries) puts around 45% of occupations in the 'complementary' category, where AI accelerates the work around a task but humans remain accountable for the task itself. Government engagement sits squarely in that group.

This is already happening in other fields. Harvey and Legora, for example, are AI systems built to accelerate contract review, due diligence and research, with a qualified lawyer reviewing every output and carrying the accountability. The FCA’s recent Mills Review into AI and the future of retail financial services makes the same case for finance: AI can widen access to expert help, but a human still owns the recommendation.

Doing nothing on AI is an economic choice 

The AI Global Report also found that 47% of the potential economic prize from AI over the next five years depends on the choices firms make. The technology sets the ceiling. Human decisions about how deeply and how sensibly to integrate it decide who captures which share. That gap between meaningful integration and casual subscription is where the 47% is being earned or lost. 

Sitting on our hands has a cost.

Three practical uses of AI we’re using 

The integration of Stonehaven and Public First a year ago combined polling of people (rather than pure synthetic data), economic modelling, policy experts, campaigning, and political advocacy. These tools gave us an outsized advantage for AI experimentation which has proved rich and surprising. Sharing best AI practices internally has proved to be the best way for us to experiment. 

Now we want to share our learnings: what we’ve tried that didn’t work, what we are still unsure about, and concrete practical examples of how AI is transforming our work: 

• Our best use case has been the AI-assisted development of internal software. We’ve accelerated our work by replacing SaaS products with custom workflows and gaining increased control of our research. Examples include our internal project management software, our finance system workflows and dashboards, and a complete overhaul of our survey fielding and MRP software. 
• Bringing together to-dos and project statuses. We use Granola for note-taking in internal and (where we have permission) in client conversations. The point is to free us to deeply listen and be present on calls rather than take notes. We connect Granola to either Claude or ChatGPT, which are connected to our emails and messaging system. This gives us a holistic picture of our to-do lists project by project. It is the foundation of a ‘second brain’ which we are now building internally.  
• Democratising internal knowledge and insights. It now takes just seconds instead of hours for our consultants to identify the existing relationships and knowledge held in our firm to benefit our clients. Our internal AI tool, ‘Monty’, enables our team to easily access our polling, focus group transcripts and project archives through our internal chat system. But also from a security standpoint, Monty can access databases that not all staff have full access to, democratising our data in a way that was not previously possible. Typical uses:

‘@Monty, I have a meeting this afternoon on devolution. What data or research have we done on it in the last 2 years?’ 

‘Who in the company holds the best relationship with individuals in the new Government and how strong are the relationships?’

‘From our polling, how have attitudes to economic growth shifted in the last 5 years?’

The judgement, the call, and the accountability stay with the human. AI merely assembles the material they request. Half a day of pulling material together becomes a starting point in minutes. 

Repeatable workflows for the work we do most.

We have built custom GPTs in ChatGPT and 'skills' (pre-set workflows for the tasks we do most often) in Claude. Some examples include a quality assurance skill to enable junior team members to quickly produce high-quality briefings and reports, a speech writing skill tailored to the voice of a principal trained on their previous speeches, and a visuals studio which turns raw data and simple prompts into clear, presentation-ready visuals in our brand. 

These examples enable us to work faster and more efficiently. Crucially they are freeing up our employees – including our junior staff – to do human-accountable government engagement better. 

Where experiments have failed is when we have tasked an AI tool with doing the parts of government engagement that humans do best. For example, AI cannot produce a perfect briefing in one go. Drafts look plausible but are lengthy and points don’t land. However, AI can do things to help the human writing process. I.e. skills for quality control and writing style, critiquing writing, offering counter arguments are all helpful. 

We also have big open questions we’re working through. How do junior people learn when the work that used to shape their judgement – like writing briefings and producing stakeholder maps – is now the work AI can do? How will government engagement work change as monitoring, briefing and stakeholder analysis become faster and cheaper? What happens to the integrity of government consultations when AI tools can flood consultations with synthetic submissions that a government AI tool then reads? 

Done badly, AI risks lowering the quality of government engagement. Done well, it can strengthen it.

This is what we're committed to building.

If you are part of an in-house government engagement team, we’re inviting you to a small in-person event on Wednesday 9 September where (over drinks) we’ll have a short fireside discussion with people doing interesting things with AI in human-led industries, share practical AI tools that are working for us, and discuss what AI uses aren’t and are working in your team. Our tech team will be on hand to help you build things (big or small) to enhance your workflows. If you’d prefer to arrange a 1-1 session, we’re always open to doing this with our clients. 

No technical knowledge is needed, just a laptop. To ensure we can keep the session practical and personal, spots are limited to 20. If you would like to be on the list, click here to request a place.