Exploring AI in the West Oxfordshire communications team

Time to deliver the quantity and quality of communications our organisation requires was becoming an increasing challenge. Like many council comms teams, we have limited budget and staffing resources, high expectations and high ambition in the team. We wanted to look at AI and assess how it could support the team to deliver the best service possible. 

We had used some elements of AI beforehand as individuals but never looked at it strategically and across all our work. We also didn’t have a proactive organisational approach to AI so we decided to look at comms as a specific service. This included reviewing all our processes and outputs, to see where AI could add value. We were open minded but realistic in our approach, knowing that early tests had shown potential for value add but also limitations.

We were also conscious of the governance, data security and other aspects of AI use that would need to be addressed through the work.

We worked with AI specialist, Ben Verinder, to analyse all the tasks we do during the day to see where AI could be applied to improve efficiency and support the team in delivering their work.

The analysis highlighted key areas where AI had the potential to significantly improve our efficiency and speed up our processes, such as:

• Producing first drafts of press releases
• Drafting copy for newsletters and social media
• Producing communications plans in line with the OASIS framework
• Writing speeches and presentations

The aim was not to replace the human role, but to free up time for specialist officers to focus more on strategy, evaluation, and creative work – tasks that often get deprioritised when the team is overloaded with their usual day to day tasks. We didn’t have an ‘in-house’ AI solution so we couldn’t explore some of the productivity benefits some platforms provide such as minute taking, etc.

After looking at options we settled on ChatGPT as the platform of choice. This was due to its ability to set up GPT’s that are tailored to specific tasks. With good quality examples for reference and clear prompts to instruct, it can produce a much better results in many writing tasks than you would get from a standard GPT model.

The team has a shared account where we develop, test, refine and share GPTs for a range of applications including press release writing, comms planning, writing speeches, newsletter content, social media and more. Our approach is to test any normal task through the AI to see what value we can get. We then run different tasks through the GPTs and amend the prompt and reference material to get better outputs. It enables us to produce GPTs specific to our council using the house style, our key messages, that reference our council plan, etc

We worked with colleagues to put our own procedures in place to manage data and security risks while the corporate policies were being developed. We also used the ‘Team’ accounts of ChatGPT that also provides additional data and security protections to help manage risks.

The team are now regularly using AI to support their work. We estimate it will add around 0.5 - 1 FTE by speeding up tasks (in a team of 4 FTE total). We have done this through calculating how much time has been saved per task compared to frequency of delivering those tasks. It does vary though depending on the type of work being delivered at any one time as the AI adds more or less value in different areas.

We are now able to better spend time on the strategic aspects of our work and deliver more work in some cases. For an example on a specific topic - a draft press release on an executive report that may have taken an hour or two to write, now takes 20-30 mins.

Taking a more strategic approach to AI means we are all talking about the approach, sharing learning and growing as a team. There is still a lot more to test and refine over the coming months, however, as a team we now have a shared goal and we are learning together.

Given the pace AI is improving, it will be crucial to keep up with advancements, test new capabilities and adapt process to be able to get the best out of it. Having a strategic approach is helping us keep up to date.

While this has largely been a positive experience we have certainly taken some valuable learnings:

• Our corporate IT team had valid concerns around data and security. A corporate/ team AI policy is crucial to ensuring the risks of AI are managed. In the absence of a corporate policy, a local one for the team is important
• An overall understanding of the different types of AI is important to understand what they are and aren’t capable of. Also how to use the different types of model effectively
• AI certainly isn’t ready to replace comms roles. It can speed up processes, and enable higher productivity, but it still has many limitations. We found many of the known issues with AI applied to us, even with refined models. We got the press release GPT to get us to 90-95 per cent on most press releases where it had a report/ good source material to work from compared to 60-70 per cent on standard models. However, where you don't have a source material it is far less beneficial.
• It is also crucial that any content produced by AI is proofed thoroughly as sometimes the work it produces is poor quality and we have even experienced occasional inaccuracies where something has been included that wasn’t in the original source material. Therefore, it is a tool to be used carefully, not as a replacement for officers.
• Not all tasks benefit from AI. The real benefit lies in repeatable tasks where you can have a standard prompt/ custom GPT. There are many tasks, especially one off tasks, where AI takes as long to prompt to get a good quality output as it would take to just do the job yourself. There are also communications tasks where AI is simply not appropriate - aside from areas with obvious GDPR concerns - particularly image and voice generation. Not only is ChatGPT’s capability for this still limited in places (although improving quickly), it also raises important questions around ethics, authenticity, and public trust.
• Test and getting stuck in. AI isn’t simple and there isn’t catch all training. There are people who know a lot and there are good guides online, however, there are lots of gaps given how quickly AI is improving. Having a go is just as valuable as any training (although learning the basic concepts such as how to effectively write a prompt is important).
• The quality of the output is entirely dependent on the quality of the prompt and the reference material the AI is given. We learned that when building a GPT, it needs to be provided with multiple good and bad examples to help it understand the tone and style it needs to replicate. Additionally, we found that prompts work best when instructions are presented in the RACE format (Role, Action, Context, Execution).
• It takes time and investment to train staff, set up prompts/ GPTs, buy a platform and monitor trends, however, this investment has more than paid for itself in increased productivity
• Don't rely on AI and think about the perception of using it. We are starting to notice the trends in AI produced work, both in what we ask it to do and also what we are seeing produced by other people. It uses certain formatting, language, etc. As AI use becomes more common, more people will start to notice the traits. It can devalue work if people think what you are presenting them was produced by AI alone.

It is important to note that we are far from experts and are fully aware that we are learning as we go and what we are doing can be improved. Even now some of the drawbacks we have found may have been solved in more up-to-date models so its crucial to test, refine and adapt as you learn and test capability.