Intelligence Layer — Plain-English Guide
Audience: anyone deciding what this system should become. No technical background needed. Engineering version: technical.md. Back to the Marketing OS index.
In one line
One shared store of everything the business knows — rules, brand, keywords, personas, competitors, performance — that every agent reads from, instead of each agent gathering its own facts.
Why this matters
Today each agent assembles its own picture of the world before it does anything. The blog writer separately fetches the brand voice, the compliance rules, the competitor brief and the SEO keywords, trims each to fit, and builds a prompt. The next agent does its own version of the same thing.
That works with three agents. With nine it becomes the main source of bugs: change one fact about the brand and you have nine places where it might not have taken effect.
The intelligence layer inverts it. The facts live in one place. Agents ask.
You have already built the prototype
The Compliance Intelligence Centre is exactly this pattern, at small scale:
- Rules are authored by your team, not baked into code
- Each rule carries its own reasoning and the source it came from
- The agent must cite the rule it acted on
- Citations are checked against the store — if the model invents a rule, the citation is thrown away
That last point is the whole idea. The model does not get to be the authority. The store is the authority; the model reasons over it and must show its working.
Generalise that shape across everything else the business knows, and you have the layer.
What belongs in it
Things that already exist, currently scattered:
| Knowledge | Lives today in |
|---|---|
| Compliance rulebook | Compliance Intelligence Centre ✅ already right |
| Brand voice, avoid list, blog master prompt | Guidelines page |
| SEO keyword inventory and gaps | Keywords page |
| Audience personas | Audience page |
| Hook formulas and creator voice | Learning profile |
| Competitor topics and positioning | Competitor Analysis |
| Article scores, categories, geography | Discovery feed |
Things that do not exist yet and would land here:
- What the team actually chose — queued, dismissed, published, edited
- What performed — after publishing, not just before
- Do-not-translate glossary — product names, regulatory terms
- Customer data — the
IIFL DATA → CRMbox on page 1, still unconnected
The one decision worth making deliberately
"The data becomes the LLM" can mean two very different things, and they behave nothing alike:
| Retrieval — agents look things up | Fine-tuning — facts baked into the model | |
|---|---|---|
| Changing a rule | Takes effect on the next call | Requires retraining |
| "Why did it say that?" | Point at the record | Cannot be answered |
| A wrong fact | Delete the row | Stays until retrained |
| Regulatory review | Auditable | Not auditable |
| Cost to maintain | Low | High |
For regulated financial services, retrieval is almost certainly the answer — and your compliance rulebook shows the instinct already went that way.
Fine-tuning is worth revisiting later for style — house voice, script cadence, the shape of a hook. It should never hold facts or rules, because neither the auditor nor the reviewer can inspect what's in the weights.
What changes for the agents
| Agent | Today | With the layer |
|---|---|---|
| Discovery | Scores each item in isolation | Scores against what this team has historically valued |
| Compliance | Already reads a rulebook ✅ | Same pattern, more rule types |
| Creative | Gathers its own context per draft | Asks once |
| Learning (proposed) | — | Writes back into the layer rather than owning a private store |
| Localization (proposed) | — | Reads the shared glossary |
| SEO Scale (proposed) | — | Reads geography, keywords and published pages from one place |
The shift worth understanding: once this lands, output quality is mostly determined by what's in the store and how well it's retrieved — not by prompt wording. That moves the work from prompt tuning to curation, which is a job your team can do without engineers.
Where this sits on the roadmap
This is deliverable ④, "Marketing OS — Architecture", on page 5 of the
whiteboard — and page 1's data layer (IIFL DATA → CRM, Outside world DATA,
Manual) is its input side.
It is the one deliverable every other item quietly depends on, which is why it looks unglamorous next to "video" and "social automation" and matters more than both.
What it is not
- Not a data warehouse. It holds what agents need to reason, not everything the business records.
- Not a replacement for the CRM. It would read customer data; it should not become the system of record.
- Not automatic quality. A badly curated store produces confidently wrong agents faster than no store at all.