In civic circles, a consensus is emerging that the current large language models are too unreliable to use in serious contexts without supervision. With the present state of the technology, anything produced by a large language model should be reviewed by a human before it’s used (e.g. gov.uk, judiciary). They still confidently get things wrong and are vulnerable to prompt injection and other security risks. While this may be very different in 6 months’ time, ambitions of programmatic deployment at scale remain risky for now.
However, this is not to say that the technology is useless! The main applications are in “increased productivity”: on average employees using AI work significantly faster and better, though it’s hard to predict which tasks will benefit from AI and which ones will not, and the exact technique will be quite particular to the piece of work, like it might be with a custom spreadsheet. The implications of this are surprising: likely there will not be any substantial change in the nature of work being done, nor should you start new initiatives or hire new roles specifically related to AI. However, parts of your organisation may suddenly be able to move faster than before, so you might find that your personnel balance has to change. In particular, roles related to managing risk—anything to do with quality assurance, supervision, review, or getting sign-off—may become overwhelmed, frustrating motivated team members. Tech pundit Benedict Evans famously says AI is like having infinite interns; from a strategic perspective, what this actually means is an infinite shortage of managers.
In this issue, we’ll be looking at the generative AI integrations appearing in common enterprise software that your organisation is likely already using. Even if you don’t have a plan to use generative AI in your organisation, it’s coming for you anyway! Most major tech firms have announced a host of new AI integrations into their product lines, launching in various states of usefulness and coherence – however these present an immediate and everyday way that your team will start using generative AI. While the strategic situation may change as the technology and product landscape improve, what we do know is that these tools present immediate opportunities and challenges, and will at least be thought-provoking.
It’s worth spending time reflecting on what uneven changes in productivity might mean for your organisation. Pro-actively discovering and resolving bottlenecks will be critical. Small projects previously shelved may start looking viable. Even modest improvements in the efficiency of very expensive activities—e.g. anything involving lawyers, developers, or senior roles—may well be worth investigating.
If you’ve had any interesting experiences like this, or you strongly disagree with this take, please let us know! [email protected]
In this article:
🏢 Enterprise Suites
For many orgs, their main decision here has been whether to shell out for the AI upgrades of their enterprise suites, chiefly Microsoft 365 or Google Workspace.
Microsoft has now fully rolled out Copilot for Microsoft 365, available for an additional (hefty) fee above the basic 365 subscription, and no longer just restricted to large orgs. At the moment this means new functionality for Outlook, Teams, SharePoint,OneDrive, Word, Excel, Powerpoint, Whiteboard, and OneNote, all confusingly called Copilot even though they behave differently in different apps.
Similarly, Google has begun to integrate AI into most features of Google Workspace under its Gemini model, now called Gemini Enterprise (previously known as Bard as the model and Duet AI for workplace apps). This is lagging behind Microsoft somewhat but presumably not for long.
Reviews of these platforms so far are mixed; they’re expensive and generally considered underdeveloped, but improvements are coming out faster than we can cover them, and we can’t help but admire their vision of end-to-end AI. We’ll cover the changes in different apps in following sections.
It’s the opinion of the Civic AI team that speeding up email will be one of the most tangibly impactful applications of genAI for many organisations: it’s a ubiquitous, non-repetitive, piecemeal task requiring lots of context switching, all things for which language models are especially helpful. Microsoft has released Copilot in Outlook and Google has a feature to draft emails with AI, both of which can summarise email chains and draft emails from prompts. Gmail has third party plugins like Shortwave. Superhuman has every feature you could think of: summarising, drafting, autocompletion, editing; it feels like these kinds of enhancements will become standard in future, and products will be differentiated by what other context they can intelligently draw from (calendars, documents, etc).
💻 Calls
Generative AI as applied to videoconferencing is mostly for making notes and summaries of meetings via transcripts. Transcription technology has been improving quickly, and with a quality transcript, LLMs can provide summaries and answer questions about the meeting, even while it’s happening. It works relatively well, though LLMs can miss tone or emphasis, and get confused if a discussion requires a lot of nuance or additional context not included in the call. In short, automated summaries are better than nothing but are best used for a high-level overview only, and don’t yet take notes that are as useful as those taken by a human.
Microsoft’s Copilot in Teams, Google Meet’s AI extension, and Zoom’s AI companion all offer versions of these capabilities. To set up AI in Teams is fiddly: create a meeting, go to additional settings to allow transcription, send the invite, then once the meeting begins, manually start transcription. With luck you should then be able to start Copilot.
Zoom’s AI Companion is worth trying but users have been split on performance: transcription is hit and miss, it doesn’t always understand which points are important, must be activated by the host to work, and is limited to English only. We’ve been recommended fathom.video as an alternative, a Zoom add-on which gives video transcripts, video recording, and AI summarisation.
There’s very little of substance from Google Meet as yet, but updates for Gemini are coming out quickly at time of writing so this may change very soon. Likely similar functionality to the others.
📱Chat
Here we’re referring to where your team sends each other instant messages, not AI Assistants like ChatGPT. It’s no coincidence they look similar though; ChatGPT has leveraged the UX of instant messaging to teach people how to interact with it, an example of skeuomorphism, which can be confusing when it’s put next to things that are actual chats with other humans.
How AI should work with team chat is still being explored: it’s appeared as an integrated AI assistant (Copilot, Discord), a tool to summarise conversations and threads (Slack, Teams), or in the case of Slack, a whole genAI appstore and automation builder. If your team mostly uses chat to communicate, and lots of information comes into the channel throughout the day, features like Slack’s thread summaries can be very helpful.
For Microsoft enterprise users, the situation is confusing. Chat-like experiences have appeared as both M365 Copilot Chat and just Copilot. M365 Chat appears as another conversation within Teams and designed to appear as if you’re chatting with a colleague while the other option appears through a separate portal and acts as a more familiar AI assistant like ChatGPT. Only M365 Chat has access to your chat history with other team members. It also gives a compose button that helps you write messages, and also a copilot button that summarises chat history.
Google’s Gemini will now have access to Google Messages, an app used by over a billion people. Since this is a very recent development, we don’t know how good it is. However, it’s already caught the attention of EU regulators who may see it as a violation of EU law, which prohibits the use of private messages for LLM training.
WhatsApp hasn’t announced much on this front, except the ability to use AI to create new stickers. However, we have seen a few third-party approaches, such as the option to integrate ChatGPT into WhatsApp or create a WhatsApp support chatbot using LLMs, OpenAI, and Python.
✨ AI Assistants
For many people AI Assistants like ChatGPT have become just another kind of work tool, and so merit inclusion in this list as their own category. Most civic organisations are already widely using these for ad-hoc tasks, mostly drafting of text.
While Microsoft and Google’s enterprise suites both have AI Assistants built in—Copilot and Gemini (previously Bard) respectively—many prefer or at least are more habituated to ChatGPT, which is furiously building out its B2B offering with ChatGPT Enterprise, including things like file uploads, data analysis, plus all the privacy and procurement infrastructure you’d expect from an enterprise product. Meanwhile in its mission to integrate-all-the-things, Microsoft has also released Azure ChatGPT competing with their own Copilot, which allows you to build integrations with your internal data if you’re in their ecosystem already. If you’ve tried using it to make anything interesting, let us know!
📅 Calendar
Microsoft Copilot can “intelligently” schedule meetings from Outlook, suggesting available times, relevant people and files, and even draft an agenda. Google Workspace hasn’t announced an AI calendar integration but doubtless will in coming weeks. Calendly is also testing AI integrations, currently available via waitlist.
📦 File Storage
One of the big promises of language models is to be able to help organisations leverage their knowledge base. This is still emerging, but new features from Microsoft, Google, and Dropbox into their existing products will likely be the first taste people will have of that. Gemini’s just released integration with Google Drive is impressive and the company has teased more features are coming. Microsoft OneDrive provides recommendations for files based on associated meetings or colleagues. Specific tools for this, such as Glean, have also appeared. Glean’s very decent list of clients suggest that large organisations are experimenting with it in addition to the AI features embedded in their enterprise suite.
📝 Docs & Notes
Both Google Docs and Microsoft Word can provide drafts based on a single prompt or a prompt that references an existing file. For example, a user could type, “Write a memo to stakeholders based on the ‘2023 Cybersecurity Report’ document” and the new feature would mock up a document according to those instructions. Copilot in OneNote similarly is an AI assistant with access to your notes, letting it do things like create summaries, to-do lists, or give ideas and feedback. The note-taking app Notion also lets you write, refine and summarise text with AI. If you’re familiar with ChatGPT these are all more or less what you’d expect; we at the observatory are more excited about clever AI augmentations of writing and editing that are still being invented.
📈 Sheets
Both Microsoft Excel and Google Sheets boast native AI integrations but they’re very rudimentary, and we don’t recommend them. If you’re adept with spreadsheets these won’t be useful, and if you aren’t, then they might make formulas you don’t understand, perhaps leading you to misinterpret the output. In Excel, you can ask it to, for example, “analyze data in a given document and ‘provide three key trends’”. Google Sheets appears a bit more behind, with only the ability to suggest table headings.
Much more interesting are addons like sheetai.app for Google Sheets and ChatGPT for Excel that let you use language models directly within your spreadsheets as functions. These are more limited than full-on AI assistants as they don’t retain context, and both prompts and responses need to fit within cells, but despite this the flexibility of spreadsheets allow for many creative uses.
🖼️ Slides
Microsoft and Google both offer AI-enhanced slides but differ quite significantly on what they can provide. So far, Google Slides only allows you to use Gemini to create images. With Microsoft Powerpoint, you can create a full draft slidedeck based on a prompt and an existing document. This leaves you with “good bones” for a formal presentation, as one of our interviewees described it. Other tools like gamma.app are also appearing which seem very slick.
🎨 Boards
Microsoft Whiteboard, Miro, and Mural have all introduced similar AI-powered features that suggest new ideas or provide summaries of brainstorm sessions. We have received mixed reviews on these; if you already have templates that you go use regularly you won’t need them, but if you’re making new boards from scratch all the time they might help you get started.
📊 Project Management
Your team may be attached to a certain project management tool, whether that’s Airtable, Trello, Monday, or Asana. Each have announced AI-features that are writing and categorization oriented, though many are still in beta and not widely available. If you’ve tried these and have an opinion on whether they’re useful, let us know.
🙇 Customer Service
AI-enhanced customer service comes in the form of chatbots, or writing assistance for agents. Zendesk, HelpScout, Intercom, Microsoft Contact Centre, and Front have all introduced some version of AI-enhanced features. There are even some creative ways to do things with WhatsApp. These are quite good, though not miles ahead of what customer service chatbots have been able to do before now and can be vulnerable to prompt injection, or at least being made to say silly things. If you’re giving legal advice or helping vulnerable people you should probably steer clear of these, but for less sensitive applications they’re worth exploring.
🔢 Data & CRM
There are all kinds of varied experiments going on in this product area with different levels of sophistication, focus, and customisability; databases are generally integrated with other systems and it’s not very clear at which points of the stack AI functionality ought to sit, so it’s interesting to see how different companies are approaching this. What it will actually mean for you will also depend a lot on how your organisation uses these tools.
Airtable’s AI integration gives you a lot of control, allowing you to create fields that produce an AI response to other fields using a specified prompt. You could do almost anything with this; the obvious use for this is extracting structured data from text, even quite abstract data like sentiment, but usual caveats around model accuracy apply.
HubSpot and Zoho offer an integrated AI assistant that can edit your database for you, as well as generate content drawing from it. If you’re using Microsoft Suite or Salesforce, they claim to allow you to do this directly in your existing apps like your email or customer service channels, which has the potential to be very powerful indeed. Finally CiviCRM hasn’t committed to a product direction yet, but is doing some small experiments to explore the technology.
🎁 Images & Ads
With the proliferation of AI-created images from Mid-Journey, ChatGPT’s DALL-E and Google Gemini, all other creative software packages were quick to offer in-house versions. Canva has launched an AI-powered version of its image design platform, Adobe offers powerful AI-features in Photoshop, and Shutterstock now includes Creative AI, an array of enhanced editing capabilities. Getty Images launched a tool for “commercially safe” generative AI images which claims to exclude known people and copyrighted material.
Both Meta (Facebook) and Google are exploring AI-assisted content generation for their ad platforms. Meta’s will automatically create variations on copy and images for different ad formats, whereas Google Ads will generate a whole campaign based off content from your website. A bit surprising not to see faster innovation here given that these are very much core to both companies, but it’s likely that if you’ve not explored online ads for your campaigns it will soon become extremely easy to do.
🧰 Other
- WordPress has a variety of AI-powered plug-ins.
- Webflow has released various AI features such as generating, altering, translating, and summarising static and CMS content, and AI powered SEO optimisation.
- Buffer introduced an AI assistant that can generate new posts for a schedule and repurpose existing posts.
- SurveyMonkey and Typeform are promising AI powered survey building.
- Zapier and IFTT have introduced AI actions, offering mindbending new possibilities.
- Quickbooks and Expensify can help categorise and reconcile payments, and check invoices and receipts.
- Yoti has released new features to make it more resistant to AI generated faces.
Is there other enterprise software you use all the time that we should be thinking about? Tell us what! [email protected]