Govern Agentic Banking With Real-Time Active Context

00:05 [music]
00:13

00:13 >> Globally, regulators are already
00:15

00:15 preparing for agentic AI.
00:18

00:18 Across many financial markets,
00:19

00:19 supervisory bodies are building active
00:21

00:21 runtime frameworks to keep track of
00:23

00:23 artificial intelligence outcomes.
00:27

00:27 The intense across the Monetary
00:28

00:28 Authority of Singapore's safeguards for
00:30

00:30 agentic finance at runtime, [music]
00:32

00:32 SAFR, guidelines from Hong Kong's
00:34

00:34 Monetary Authority, the UK Financial
00:36

00:36 Conduct Authority, and the well-known EU
00:39

00:39 AI Act are similar.
00:41

00:41 Financial institutions must govern
00:43

00:43 automated decisions at runtime and
00:45

00:45 ensure that every single agent decision
00:47

00:47 remains fully auditable. [music]
00:50

00:50 To handle this, the Monetary Authority
00:52

00:52 of Singapore created SAFR as a reference
00:54

00:54 framework.
00:56

00:56 It sits directly between an AI's
00:58

00:58 decision and the actual transaction.
01:01

01:01 It protects the bank in four main steps.
01:04

01:04 First, it identifies exactly who the
01:06

01:06 agent is.
01:08

01:08 Secondly, it checks the action against
01:10

01:10 the bank's internal rulebook or controls
01:12

01:12 repository.
01:14

01:14 Third, the agent makes a final
01:16

01:16 definitive decision.
01:18

01:18 And finally, it locks that entire
01:20

01:20 [music] history into a permanent audit
01:22

01:22 log.
01:23

01:23 Every action resolves to one of four
01:25

01:25 outcomes: deny, escalate, auto execute,
01:28

01:28 and observe.
01:31

01:31 This is all about governing agentic
01:32

01:32 actions at runtime and keeping decisions
01:34

01:34 auditable.
01:36

01:36 But the question remains, [music]
01:38

01:38 how do you handle this engine data
01:39

01:39 that's connected, governed, live, and
01:41

01:41 semantically [music]
01:42

01:42 coherent, and always ready for an agent
01:44

01:44 to act on?
01:46

01:46 The Denodo platform provides the AI data
01:49

01:49 layer, enabling active context for these
01:51

01:51 agents.
01:53

01:53 Four pillars make it work: universal
01:55

01:55 connectivity, zero copy delivery,
01:57

01:57 governed access, and semantic trust.
02:01

02:01 We'll see all the four pillars in action
02:03

02:03 in the next few minutes.
02:05

02:05 Let's start with the Denodo data
02:07

02:07 marketplace and click on this customer
02:09

02:09 risk profile data product.
02:12

02:12 This is a certified business data
02:13

02:13 product and unifies know your customer
02:15

02:15 risk tier, AI generated risk score, and
02:18

02:18 sanction status in the Denodo semantic
02:20

02:20 layer.
02:21

02:21 It's governed and published once, then
02:23

02:23 usable by any consumer, human, or AI.
02:27

02:27 That's semantic trust.
02:29

02:29 As you will see shortly, each agent will
02:31

02:31 only have access to the specific
02:33

02:33 information in this data product [music]
02:34

02:34 that it is entitled to.
02:37

02:37 For the purposes of this video, here is
02:39

02:39 a simulated runtime governance center
02:41

02:41 for a bank following the SAFR guidelines
02:43

02:43 we described earlier.
02:46

02:46 AI agents are actively initiating
02:48

02:48 transactions in the bank, as you [music]
02:49

02:49 can see in this table.
02:52

02:52 And payments are being approved, risks
02:54

02:54 are being scored, and a simulated
02:56

02:56 runtime disposition engine is used to
02:57

02:57 evaluate each transaction with all the
02:59

02:59 evidence federated from Denodo.
03:02

03:02 This live dashboard shows us that 15
03:04

03:04 records are being monitored currently,
03:06

03:06 and the current status is that there are
03:08

03:08 three denials, three risky transactions,
03:10

03:10 three [music] which have escalated their
03:12

03:12 limits, and six which are auto approved.
03:15

03:15 As we have seen in the SAFR framework,
03:17

03:17 there are four possible outcomes for
03:19

03:19 each transaction.
03:21

03:21 For example, if a particular transaction
03:23

03:23 has an anti-money laundering flag, it is
03:25

03:25 automatically denied.
03:28

03:28 We can also see that around $500 million
03:30

03:30 worth of transactions are being
03:32

03:32 evaluated in this live dashboard.
03:35

03:35 >> [music]
03:35

03:35 >> All the data that you see in this
03:37

03:37 governance telemetry system has been
03:38

03:38 federated live from data sources [music]
03:40

03:40 such as Databricks, Snowflake, Postgres,
03:43

03:43 an API, or any system that banks may
03:46

03:46 use.
03:48

03:48 This demonstrates the [music] universal
03:49

03:49 connectivity capability of Denodo. Now,
03:52

03:52 let's demonstrate governed access.
03:55

03:55 As you can see here, we are in a full
03:57

03:57 visibility mode, but each agent should
03:59

03:59 only see the field that it needs.
04:02

04:02 For example, the payments agency's
04:04

04:04 account, amount, limit, and verdict.
04:08

04:08 And nothing about risk or sanctions,
04:10

04:10 which are grayed out.
04:12

04:12 The risk agent, on the other hand, who
04:14

04:14 should, of course, have visibility of
04:16

04:16 all these risk metrics and may not need
04:18

04:18 to have visibility on account ID.
04:21

04:21 The security travels with the data
04:23

04:23 enforced by the semantic layer.
04:25

04:25 Now, what if rules suddenly change or
04:27

04:27 this bank needs to test a new threshold?
04:31

04:31 For example, a new spending ceiling of
04:33

04:33 $40,000 instead of a standard spend.
04:37

04:37 Let us apply this threshold. As you can
04:39

04:39 see, the governance verdicts have been
04:41

04:41 recomputed [music] in real time.
04:44

04:44 Many transactions that were previously
04:46

04:46 escalated are now auto executed.
04:49

04:49 This is an example of zero copy
04:51

04:51 delivery. Live data, no movement.
04:54

04:54 A change in the rules engine or the
04:56

04:56 rules repository did not
04:59

04:59 need to be conveyed through complex
05:00

05:00 pipelines or batch jobs for the agent to
05:03

05:03 take note of the change and the decision
05:04

05:04 to be applied.
05:06

05:06 Every verdict of every agent is
05:08

05:08 auditable as well. Let's pick a
05:10

05:10 transaction.
05:12

05:12 For example, number three. The verdict
05:14

05:14 for transaction three is observe, and we
05:17

05:17 can see exactly why.
05:19

05:19 We can see the KYC risk tier, the AI
05:21

05:21 risk score, the AML sanction status, and
05:24

05:24 the limit as well.
05:26

05:26 So, we know exactly which data has been
05:28

05:28 used to arrive [music] at this verdict
05:30

05:30 of observe, and all the data has been
05:32

05:32 federated through Denodo.
05:34

05:34 Finally, let's see an agent querying
05:37

05:37 Denodo in real time.
05:39

05:39 This is a simple agent, and all we are
05:41

05:41 asking it is to analyze transaction ID
05:43

05:43 one according to the [music] SAFR
05:45

05:45 guidelines and provide conclusions.
05:48

05:48 Active context from Denodo is providing
05:50

05:50 this AI agent the real-time information
05:52

05:52 to make the evaluation that it needs.
05:56

05:56 And here we go. Transaction ID 1 was
05:58

05:58 denied due to AML match sanctions
06:00

06:00 despite moderate [music] AI risk score
06:02

06:02 and transaction amount below the
06:03

06:03 applicable limit.
06:06

06:06 This is active context [music] working.
06:08

06:08 Any institution governing autonomous
06:10

06:10 agents will definitely need the same
06:11

06:11 foundation.
06:13

06:13 Denodo, the AI data layer, which creates
06:15

06:15 active context across your full
06:17

06:17 enterprise data estate with universal
06:19

06:19 connectivity, zero copy delivery,
06:22

06:22 governed access, and semantic trust.
06:28

06:28 >> [music]