How to Successfully Deploy Generative AI into a Small Team: Best Practices for Maximising Efficiency and Collaboration

Starting off

AI has perhaps generated more hot-air and LinkedIn posts than any other technology in the last 20 years. The rapid evolution of generative AI tools and techniques in particular has brought with it a whole raft of promises of workplace revolutions, together with dire warnings of redundancies and workforce disruption.

To make this all a bit more tangible this article relates recent experience of actually deploying generative AI tools into a small programme management team in a large organisation, and the pros, cons and "mehs" that came along with it. I'll run through some of the things we learned, actionable steps if you're looking to go down the same path, and some gotchas to look out for.

The takeaway: the team loved it, it saved time, the return on investment probably stacked up, but as yet, probably not a revolution.

Build the team with space and permission to think and innovate

To set the context here, the team in question were sat in the middle of a large programme of delivery, which in turn consisted of many other smaller agile and less-agile delivery teams. They were responsible for both supporting senior management and stakeholders to track progress and set direction, and also in supporting the delivery teams to unblock issues and generally enable delivery. The team were working in a hybrid way, with the majority of working time done remotely with regular time co-located in the office.

The team were already set-up to work in an agile way, and we were ensuring that we took time to both reflect on how we were working and the wider challenges we were facing. This happened in a number of forums including regular retrospectives, but there was also ad-hoc sharing of ideas, work in progress and experiments in scheduled time-slots like our weekly demo session.

I'd also taken time with the team to reflect on their role within the organisation, and how the team should support the wider strategic direction of the programme. In quarterly strategy and planning sessions we'd set out our aspiration as a team to evolve away from "shuffling data around" towards "spending more time on adding value and providing insight" to the teams and stakeholders we were supporting.

Foster a culture of knowledge sharing

Some of the team were already reasonably familiar with ChatGPT, and the other OpenAI tools in prior roles within the business and also outside of work, and they'd played around with various free generative AI tools, so they had a good idea of the "shape" and the general capabilities of AI. We took time to share some of this knowledge around the team in demos, written materials and conversations, and provided pointers for other team members to find and experiment themselves.

This way of working also helped us accelerate once we had the AI tools in place. A strong collaborative mindset help us share the things that worked well around the team.

Identify the needs first, not the solution

Our overall team set-up then helped us join the dots between some of our problems (e.g. "too much to do", "too much repetitive manual work") and the opportunities that might come from AI tools.

It's probably pretty important to note here that at no point did we imagine that AI tools would be a "magic bullet". There were plenty of other approaches we were taking to solve those problems (for example looking at how we could better link and share data, and how we could lean our processes), but we had a hypothesis that generative AI tools might be able to be part of the solution to our needs.

For productivity applications the organisation was co-existing in a hybrid state with both Google G-Suite and MS365 being used across the programme. Although Google is catching up fast, Copilot, Microsoft's AI tooling developed with help from OpenAI, is still some way ahead and is pretty well integrated into most of MS365 - particularly MS Teams.

Copilot is sold as an add-on to existing MS365 licenses, so comes at an additional cost, and typically will need to be enabled by whoever is maintaining the IT licenses for your org.

Handily, our technology department were looking for guinea pigs to experiment with AI, and not finding that many takers. So as I could demonstrate we'd done our homework, understood the tools and the risks, had sketched out roughly what benefit we'd be expecting to get, and how we would measure success, it only took some gentle persuasion to get the appropriate Microsoft licenses added to MS365 to enable Copilot integration.

Give yourself time and integrate gradually

Once everyone was set-up we observed that after a fun few days of "playing" with the tools, usage actually initially dropped off, before gradually starting to pick up over time as teams understood better how the tools could be used. We were a busy team and spending the effort needed to shift ways of working to consistently and satisfactorily use the AI tooling was not always straightforward.

Sometimes there were easier wins. For example transcription tools in MS Teams together with Copilot allowed meeting notes and summaries to be automatically generated and more quickly shared. These weren't always perfect, but as these kind of things were shared on Slack for review and comment anyway, our existing processes had existing quality checking processes built-in. The Copilot AI summaries were certainly no worse than the manually recorded ones and in fact they were usually better and often more comprehensive.

There are other advantages too, for example rather than waiting for notes to be cleansed and published manually, these could be shared nearly immediately, when the actual conversations were fresher in people's minds, which actually helped to produce a more accurate and useful result. Additionally as Copilot can summarise meetings without meeting recording being activated, meetings felt a little less rigid as people weren't as conscious that every sneeze and twitch was being preserved for posterity.

Summarising longer materials proved very useful, and again, surprisingly accurate. Being able to take a long group email thread in Outlook or Gmail, summarising it either directly in the tool or by cutting and pasting text into a fresh document, and then passing it on to somebody else for comment could definitely lead to a more succinct conversation.

Drafting of documents also proved a quick win, although we found that often rather than directly asking Copilot to draft in full and then re-writing, it could be a useful mid-step to ask for suggestions at a higher level. For example by suggesting it produced a document structure, outline or bullet points, this then then let the team more accurately include domain knowledge that Copilot didn't have.

Monitor results carefully

As we'd started with the needs, this also helped us later on, as we were able to refer back to our original intent and assess how successful we'd been. Of the use-cases we'd identified we successfully achieved around 60% usefully, another 20% less usefully, and the remaining 20% were a complete blow-out. In particular we found less use cases in Excel and Powerpoint, where we either lost trust in the accuracy of the results Copilot produced, or found it easier to produce from scratch ourselves.

The benefits gained varied across the team with those whose roles were a close fit to the "useful use-cases", for example more summarisation of documents, conversations and in particular transcription of meetings, finding the most benefit. Time saved across the month for those members outweighed the likely ￡25 a month per user cost but for others it was more marginal, and interestingly the feedback from senior managers in particular was that they found that they used it less.

For admin users Microsoft actually provide a useful set of reports. This both enable you to pre-identify users who will likely benefit from Copilot, by analysing their usage of different apps in the MS365 suite, but also allow you to look at their true usage when Copilot licenses are deployed. This is a good way to cross-reference qualitative feedback from teams and the actual use. We found that a majority of people across the organisation thought they wanted to use Copilot and wanted to join us on the trial, our experience though suggested some would benefit more than others.

Conclusion

Starting with a well set-up team, with a general understanding of AI tools capabilities, and a clear problem to solve which fits that capabilities leads to a higher chance of initial success. Starting small, rather than a big bang approach made it easier to experiment, iterate and find where the tool was most useful. Lastly, make sure to reflect and reflect on adoption, including hard data to back that up.

When we compared members across the team we found that co-pilot did save time, however depending on the team members role this varied a lot. The capabilities of Co-pilot are strongest and most useful currently around MS Teams: if you're a heavy Teams user right now you're likely to see the most immediate benefit.

Obviously Microsoft, Google and others are pushing the development of their products hard, so this picture might well improve further, and in particular the possibilities for data analysis in for example Power BI and Excel look to have strong potential.

Overall though, it was a positive trial and the team continues to use and expand their use of AI.