Picture the scene: You sit down to draft a proposal for a client. Or a pitch deck. Or a strategic framework. You know how your agency prefers to do this type of work. There may even be a template. But you can't shake the feeling that you've made something very similar to this before. Something great that you were proud to put in front of the client. But you can't remember which client it was for, or what project it was part of. Or maybe you can, but you still can't find the files. So you slog through the hours to get the work across the line again, suspecting that the last time you did this work it was even better, and the time spent starting from scratch is a huge waste.
This is what knowledge management actually looks like at an agency. It’s not an abstract strategy, but the daily cost of not making the most of what you already have. Making the most of past work isn't a new problem for agencies, but AI is making it one that is both more important and much easier to solve. Everyone has access to the same AI tools, but those tools are only as good as the context you feed them, so a repository of more, better reference work creates a sustainable competitive advantage in a market that's never been more competitive. Previously, building a database of past work relied on everyone at the agency diligently filing work in the correct way. It became a to-do item that often slipped off the list when more urgent priorities came along. This article is a practical guide to unlocking the value of your previous work for current clients, so you and your clients save time, and you maximize the value of every AI tool credit you spend.
What actually is context?
If knowledge management used to mean folders, naming conventions, and a wiki nobody updated, in the AI era it means something more useful: context. "Context" and "taste" are hot buzzwords these days as everyone races to find ways to use AI to do distinctive work without starting from scratch every time, or prompting for longer than it would have taken to do the work yourself. Practically, there are three types of context most AI tools need to create outputs:
The prompt itself. The instructions, examples, and framing you give in the moment. "Write a LinkedIn post" versus "Write a LinkedIn post for agency leaders, in my voice, arguing X". The second provides more context.
Supplied materials. Documents, data, brand guidelines, past work, meeting transcripts. Anything you feed in so the AI has background on your situation rather than the average background it can access (usually from the internet).
Ambient/system context. Things the AI has access to without you pasting them: connected tools (your CRM, docs, calendar), memory of past conversations, knowledge of your role and preferences.
Writing a great prompt is a real skill, and even if you and your agency colleagues happen to be very good at it, it's almost impossible to provide all the nuanced information needed to deliver a high-value, relevant output from a prompt alone. Supplied materials make it more efficient to provide a lot of rich information, quickly. However, it takes work to gather all the relevant documentation and feed it to an agent. Doing this every time you prompt is time-consuming, and repetitive if what you're supplying is relevant to every task for that client.
The real unlock is in the third type of context: ambient/system context. By giving AI access to connected tools, reference documentation about specific clients and projects, and agency-wide guidelines and frameworks, you can keep output high-quality and on-brand without supplying new context every time.
What needs to change to unlock context with AI
So how do you get your agency consistently working from ambient/system context in your tools? Lots of individuals invest in creating a "personal knowledge management system" to help them with daily tasks, but creating one for an entire organization can be a little more complex. Here are the steps to take to lay a foundation for more effective use of context:
Step 1: Audit tools
Where does your agency's work live? These are the typical categories of tools where your context already exists:
i) Document repositories, such as Google Drive, Notion, or Dropbox. Hopefully, a lot of the documentation of your past work is there. Don't worry if the data is in multiple formats like video files, PDFs, documents, and spreadsheets. AI can usually read all of these.
ii) Written client comms. The agency probably has a standard email platform, often Gmail or Outlook. Usually you'll also have a messaging tool, such as Slack or Microsoft Teams. These are excellent sources of more qualitative information, like feedback on past projects that can inform current work.
iii) Meeting notes. If you're reading this article, you hopefully already use an AI notetaker tool, like Supernormal's Notetaker. If not, perhaps you have an alternative way of storing meeting records? Meeting notes are often an even richer source of qualitative context than email, as they contain information both about the delivery of the work, and the reaction clients had to it.
iv) CRM and/or accounts. This may be no more complex than a spreadsheet, or maybe you have a full Salesforce instance. Whichever suits your scale, you almost certainly have tools to manage relationships and track revenue and costs. This context allows you to overlay business context onto your work, so AI can help you prioritize.
v) Design tools and brand guidelines. Maybe you use Figma, Canva, or specialist presentation tools like Pitch. Tone-of-voice guidelines for your agency and clients may live in a standard document, but images and design standards are often in dedicated design tools.
vi) Project management tools. Systems like Asana, ClickUp, Monday.com, or something you built yourself in Notion hold all the information about how your agency's work actually gets done. Who works on what, how you break down projects into tasks, and which deadlines are hit and missed.
Step 2: One-off data organization
As you audit your tools, it's likely you'll find opportunities to consolidate where data lives, or realize that certain types of information haven't been consistently captured. Now is a good time to do a one-off cleanup. Don't worry too much about perfectly structuring where data lives. AI is often better than humans at working with unstructured data. However, you will likely want a clear way of differentiating which context relates to which client, and which relates specifically to your agency. That way you can ensure that you're never using inappropriate context from one client to inform work for another.
After Step 2 you should know both where your context is, and roughly how it's structured.
Step 3: Automate future capture
As you clean up your different "buckets" of context, it's normal to feel a bit frustrated about how messy everything is, or where there may be gaps because some things weren't consistently captured. There are diminishing returns in trying to fix this, so focus instead on how and where information gets captured from now on. The rule that matters: knowledge you have to file by hand never gets filed, but knowledge captured automatically as you work does. Make sure everyone at the agency is using the same meeting notetaker and storing meeting notes consistently for example.
How to use an AI agent with your agency's past work
Now that you know where your context lives and have a plan in place to keep it fresh, you can start putting it to work. Every AI model or tool will have a different name and way to set up the connection between the tool and your context, on both an individual and an organizational level, but broadly speaking there are three ways to feed AI context:
How to do something. Think of this as a kind of permanent prompt. You give the AI context on how your agency prefers to do things, for example tone-of-voice guidelines, or a strategic framework you always use. Claude calls these "Skills", but other tools may use other wording, for example "Design Guidelines" in a design tool. Think of this type of context as a way of avoiding having to write the same thing in the prompt box over and over again.
Examples to inform outputs. When you want the tool to reference past work, you can connect a broad set of context to inform the output. You aren't telling the AI how to do something, but giving it examples to help guide its work, just like you would a colleague you're briefing on a new project. This is where finished work earns a second life. Treat your best briefs, proposals, and recaps as templates and reference material, not one-offs.
Permission to execute in tools. Beyond supplying guidelines and background on the work, you can also give AI permission to edit things in other tools. For example, you can tell the AI to update your CRM based on a recent customer call. With this kind of connection you close the loop, and agents can continue building your context as well as referencing it to produce better work.
For each AI tool your agency is working with, decide what type of connector you want in place, and which context should be referenced where. Think through which connectors should be in place for all team members, and which will be the responsibility of individual users to set up. The heading illustration for this blog post may help you think this through.
Most AI tools have a way of keeping certain workstreams, and therefore work context, separate. In both ChatGPT and Claude, this is called a "Project". For agencies, it is often extremely important to keep context from one client separate from another's. Savvier clients may even have clauses about this in their contracts with you. Determine who at the agency is responsible for maintaining and monitoring Project setup in your tools, and ensure the walls between client contexts are as immune as possible to natural human error.
How to actually implement this change at your agency
As with any change in workflow, best practice is easier to design than it is to implement. You and your colleagues have existing habits, and these may have to change for you to start reaping the benefits of referencing agency context with AI. Agency leaders we've spoken to had the following advice for anyone setting up new ways of working with AI:
Start small. Run a pilot project with one client or team to test what works before rolling new processes out agency-wide.
Work with, not against, team preferences. If team members love using a certain tool, forcing them to adopt a new one is unlikely to be successful. Help people use context more effectively where they already like working.
Make things visible, and social. Create team rituals that demonstrate how work was made with agency context, not just the end result. Encourage people to show the failures as well as the successes so the whole team can learn.
Leaders have to walk the walk. Make sure leadership is also demoing their workflows, and following new ways of working that integrate context-powered AI.
Be careful what you measure. Billable hours are no longer the prime measure of value in the agency world, and the whole point of using AI more effectively is to improve agency margins by decoupling client value from time spent. Tracking credit usage can also be counterproductive. Think about what you're celebrating at your agency, and the practical implications for team progression and financial rewards.
Your agency’s opportunity to make more of AI
AI is disrupting the agency business model, and creating a lot of uncertainty. When margins are squeezed and growth is hard to find, it is natural to fear new ways of working. But AI offers agencies a huge opportunity: to turn past work into a sustainable competitive advantage. Case studies of previous campaigns may have won you pitches in the past; now that work can help you deliver higher-quality, more distinctive work in the future as well. There's a reason Y Combinator, the world's most prestigious startup accelerator, is now requesting applications from AI-native agencies. The asset is already there, sitting in your Drive, your inbox, and your meeting notes. Now you can turn those files into fuel.
Looking for the easiest possible way to get started? Try Supernormal's free meeting notetaker, then use an agent with access to all that meeting context to complete your next steps for you.
