The mistake most SA agencies make with AI
When marketers ask how I use AI, they expect me to say “for writing captions.” That is the loud use case. It is also the shallow one.
Here is what actually changed my agency: I stopped asking AI for answers, and I started asking it about my own client history.
Every campaign you have ever run. Every creative test. Every winning subject line. Every underperforming ad. Every objection a client raised, every result you delivered, every “we tried this and it flopped” from two years ago. All of that already exists in your business. It sits in Slack scrollback, in your sent folder, in a screenshot from GA4 you dropped into a Google Doc last April. It is written down. You just cannot get to it fast enough for it to matter in the pitch, the strategy call, or the crisis Slack at 22:00.
That gap between “we know this somewhere” and “we can act on it now” is your biggest hidden cost. AI closes it.
What this looks like at agency scale
I run performance marketing, SEO and content for about twenty active clients. On any given Tuesday I might be looking at Meta underperformance on a financial services client at nine, a GA4 attribution question from a wholesaler at eleven, and a new-business pitch to a hospitality group after lunch. That Tuesday is impossible for one operator carrying it in their head. The only way it works is that I do not carry it. I carry a system that surfaces it.
Concretely: everything I write about a client lives in one place. Meeting notes, campaign summaries, ad account changelogs, screenshots of dashboards, small observations like “the client mentioned their compliance officer changed jobs.” That folder is indexed by a retrieval layer. When I open a conversation about that client, an AI assistant sees the last six months of context before I say a word. It knows their brand voice. It knows which channels work. It knows what we already tried and why we stopped.
That is not a productivity hack. That is a different kind of agency.
Why "just use ChatGPT" does not get you here
The version of AI most agencies use is the browser tab. Open ChatGPT, paste a request, get a reply, close the tab. That interaction has amnesia baked in. It knows nothing about your client, their sector, their compliance posture, or your last three quarters of results with them.
So the output is generic. It uses phrases nobody in the sector actually uses. It contradicts a positioning decision you agreed six weeks ago that you now have to re-explain to the account team. You end up rewriting most of it, and you conclude “AI is not there yet for this.”
The problem is not the AI. The problem is that you are asking a stranger to write about someone else’s business.
The three-layer setup that changes this
You do not need my exact stack. You need the pattern.
Layer 1: One home for every piece of client knowledge. Pick one place where every meaningful thing about a client lives as plain text. For me that is a folder of markdown files per client. For you it might be Notion, or a shared Drive with a strict rule that everything worth keeping goes there. The requirement is not the tool. It is that “written down somewhere” and “in that place” mean the same thing.
Layer 2: A capture discipline. After every meaningful client interaction, three lines get written down. What we agreed. What is still open. What surprised me. If those three lines never get captured, the conversation effectively did not happen. This is the discipline nobody wants to hear about, and it is the discipline that makes the rest work.
Layer 3: A retrieval layer over the whole thing. This is where AI enters. There are now several ways to point an AI at a folder of your own text and let it answer questions using only what is in there. Some are technical. Some are as simple as pasting the last month of client notes into a chat and asking. The specifics will change every six months. The principle will not.
Once these three layers are in place, the change is not that you save an hour a week. The change is that you start asking questions you would never have bothered to ask before. “Which of my clients has flagged the same issue this year?” “What did we quote the last three prospects who asked about paid social?” “Which creative angle worked across three unrelated clients?” Questions that used to take an afternoon of digging now take seconds. Because they take seconds, you actually ask them.
Three failure patterns that kill this
Trying to build the perfect system first. Six weeks of choosing tools, arguing about tagging conventions. By the time it is ready, nobody trusts it. Start with a single folder and a single rule. Improve later.
Asking AI to organise the knowledge for you. AI is good at retrieval. It cannot invent institutional memory you never captured. If your team has never taken meeting notes, no assistant will magically produce them. The capture is on you. The retrieval is where AI earns its keep.
Confusing the AI answer with the source. An AI answer is a starting point. It is not a citation. Before you act on an AI-surfaced fact in front of a client, click through to the actual note it came from. If the note is thin, your answer is thin. This one habit prevents most AI-related mistakes.
What this changes in a client conversation
When a prospect asks “have you seen this work in our sector,” you should be able to answer in under a minute, with three concrete examples pulled from your own history. Not a generic case study. Your actual history. That is a completely different sales conversation to “let me get back to you next week.”
When an account manager takes over a client, the handover stops being a two-week ramp of scrolling old email threads. It becomes a single conversation with the retrieval layer.
When a strategy call needs the answer to “what have we already tried on paid social for this client,” you have it before the meeting starts.
None of this requires new headcount. It requires captured knowledge and a retrieval layer that reaches it.
The unfair advantage nobody talks about
Every agency in your market has access to the same public AI. The public AI gives the same generic answers to everyone. That is not a moat. That is a tie.
Your moat is the knowledge that only exists inside your agency. Your specific clients, your specific mistakes, your specific wins. That knowledge is already yours. It is already written down, badly, in six different systems.
The agencies that compound the fastest over the next three years will be the ones that take that pile seriously enough to make it queryable. Not because AI is magical. Because for the first time, the cost of turning a mess of client notes into a working brain is close to zero.
Do the capture. Build the folder. Point the retrieval layer at it. That is the whole play.
Everything else is caption generation, which is nice, and completely beside the point.
Want us to audit what your agency already knows?
If this landed, the practical next step is not more tooling. It is auditing what your business has already written down and turning that into a queryable knowledge base your team actually uses. GKnect Digital does this for agency operators who want to stop losing pitch history and start compounding it.
Reply to this post or reach out via gknectdigital.co.za.
Part 1 of a 4-part series on how Marcos Diez uses AI to run GKnect. Coming next: AI without guardrails is a compliance lawsuit waiting to happen.
