Martech, explained · AI and agents in martech · 2:35
The context layer, explained
The read layer in front of your warehouse, CDP and engagement platform, where AI agents ask about a customer
The short answer
The context layer is the read layer in front of your warehouse, CDP and engagement platform, where an AI agent asks about a customer. It has to be queryable, addressable, current and accountable; most stacks deliver two of the four. And when two systems disagree, no standard decides which number wins. A person does, one fact at a time.
Questions this answers
Does a context layer replace the CDP or the warehouse?
No. The warehouse still stores, the CDP still resolves identity and the engagement platform still sends; the context layer is where an agent reads from them.
What does a context layer have to do?
Four things at once: answer a specific question fast enough to act on, give every customer a stable key, be fresh enough for the decision, and record what the agent read and when.
Why do most stacks score two out of four?
Vendors sold you queryable and addressable. Freshness is often a refresh every few hours, and nothing records what the agent saw, so the last two are nobody's job.
What happens when two systems disagree?
The agent uses whichever number an engineer pointed it at. No standard decides which wins; somebody with the standing to say so has to, one fact at a time.
Where do you start?
With one workflow. Write down what it reads, when, how stale that is by the time it acts, which system each number comes from, and who should own the answer.
Transcript
Every chapter as text. Click a time to play the film from there.
0:00 One question
Before an agent sends a discount, it asks one question.
Is this customer worth it?
Where it looks for the answer is what people now call the context layer.
0:13 Where the layer sits
None of this replaces anything.
The warehouse still stores, the CDP still resolves identity, the engagement platform still sends.
It's the read layer in front of them, where an agent asks about a customer right now.
0:29 Four things it has to do
That layer has to do four things, all at once.
Queryable: ask a specific question, and get an answer fast enough to act on.
Addressable: every customer, cart and ticket has a stable key the agent can fetch.
Current: fresh enough for the decision, which is seconds in a session and a day for a churn score.
Accountable: later on, you can see what the agent read, and when.
0:56 Most stacks deliver two
Most stacks deliver two of the four.
Take a win-back email that goes out after fourteen quiet days: queryable and addressable, two points each.
The activity data refreshes every six hours, so someone who clicked an hour ago can still get it.
One point.
And when someone asks why, nothing records what the agent saw.
Zero.
Five of eight.
Vendors sold you the first two, and nobody sold you the other two.
1:27 When two systems disagree
Look, even all four isn't the whole answer.
Say the warehouse and the CDP both calculate lifetime value.
One nets off returns, the other doesn't, and they're twenty percent apart.
Both are defensible.
Both are fast, both are fresh, both are logged.
The agent uses whichever one an engineer happened to point it at.
1:52 Who decides
No standard decides which number wins.
Somebody at your company with the standing to say so has to decide, one fact at a time.
2:02 Start with one workflow
You don't need a new platform to start, just one workflow.
Write down what it reads, when it reads it, and how stale that is by the time it acts.
Then find out which system each number comes from, and who should own the answer.
2:18 So, is the customer worth it?
So, is this customer worth the discount?
That depends on which number the agent read, and whether anyone decided it should.
Learn more at Martech Stack Builder, or build your own.
Read more
The articles this film is drawn from.