LensReading which lens this session carries.
Research & Development/Data Science & Modeling

Sub-function

Data Science & Modeling

Owns model work end to end inside Research & Development. Runs 3 lines across 19 stations, drains the repeatable share into the DATA-ANALYTICS tower, and holds 4 human gates.

orchestrator Modeling Orchestratorhuman owner Director, Data Science routes into data-analytics

Estate rating

BBB

72 / 100

autonomy 75%cap 83%

Training, validation and drift detection are fully agentic. Promoting a model to production requires a named reviewer, because a model in production is a decision-maker.

Agents

8

Lines

3

Units / wk

15,032

Timetable vs actual

13.8h → 12.4h

Cost / unit

$14

was $46

Run-rate saving

$24.6M

annualized at this volume

The 8 agents running this sub-function

Grouped by what they are for, not by where they sit. The agent org is process-shaped.

OrchestratorOwns the queue and arbitrates between desks.

Modeling Orchestrator

Owns the data science & modeling lane end to end. Sequences the other agents, holds the timetable, and decides what surfaces to a human.

ObserveA3195 / 24h95%11% ovr$8.26/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

AI GBSRuns the shared transaction spine.

Feature Engineering Agent

Runs the repeatable portion of data science & modeling inside the shared transaction spine at spine cost and spine controls.

AI GBSA3664 / 24h98%5% ovr$0.34/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

TaskExecutes stations on a line.

Model Training Agent

Executes the data science & modeling step it owns, posts its evidence, and hands the unit to the next stage.

WorkflowA32,874 / 24h98%6% ovr$0.60/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

Validation Agent

Executes the data science & modeling step it owns, posts its evidence, and hands the unit to the next stage.

WorkflowA33,505 / 24h89%9% ovr$0.36/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

Explainability Agent

Executes the data science & modeling step it owns, posts its evidence, and hands the unit to the next stage.

WorkflowA34,173 / 24h92%4% ovr$0.58/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

ObserveWatches signals and watches the agents.

Drift Monitoring Agent

Watches data science & modeling continuously. Detects drift, quantifies it, and routes what matters without waiting for a reporting cycle.

ObserveA3155 / 24h87%15% ovr$4.50/task

ceiling A3 · escalates to Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.

PolicyAuthors and versions the binding rules.

Model Governance Agent

Holds the rules that bind every other agent in data science & modeling. Versions them, tests them against live decisions, and blocks work that breaches them.

PolicyA221 / 24h92%8% ovr$6.97/task

ceiling A3 · escalates to Escalates on any decision that would breach a live policy version, with the clause cited.

ChallengerPaid to disagree before a human has to.

Data Science & Modeling Challenger Agent

Adversarial reviewer for data science & modeling. Argues the opposite case on every position and flags where the evidence does not carry the claim.

StrategyA2191 / 24h93%10% ovr$2.94/task

ceiling A3 · escalates to Escalates when a ratified position is still running against a break condition it flagged.

Estate rating breakdown

Seven control dimensions. Judgment readiness runs against autonomy on purpose.

Control Coverage69
Evidence Quality74
Override Discipline76
Data Integrity77
Recovery Readiness79
Cost Transparency77
Judgment Readiness54

People

What the humans stopped doing, and what they do now.

headcount on this work69.250.6 FTE

Redeployed into scenario work and strategy authoring review.

inference cost per unit$3.42

Lines running in this sub-function

Each line is a workflow. Click a line to walk it station by station.

13 clear4 evidenced3 held
RD5AModel DevelopmentTransactional · 8 stations

Frame problem

0 q

Engineer features

2 live

Train

13 q

Validate

2 live

Explainability pack

2 live

Reviewer approve

2 live

Register

2 q

Deploy

1 live

RD5BDrift ResponseTransactional · 6 stations

Detect drift

2 live

Diagnose

7 q

Retrain

1 live

Validate

2 live

Approve promotion

3 q

Deploy

8 q

RD5CModel RetirementTransactional · 5 stations

Detect obsolescence

1 live

Assess dependency

2 live

Owner approve

2 live

Retire

18 q

Archive

1 live

Strategy positions this sub-function holds

An agent that only executes is a robot. These are the calls it made, the argument against each one, and what would prove it wrong.

RD5-POS-100

Data Science & Modeling Operating Position

In challengev2

Where should the autonomy ceiling for data science & modeling sit over the next four quarters, and what evidence would justify moving it?

Hold the ceiling at A3 for data science & modeling. Training, validation and drift detection are fully agentic. Promoting a model to production requires a named reviewer, because a model in production is a decision-maker. Raise it only after two consecutive quarters where the override rate stays under five percent and every override has a written cause.

confidence

59%

The challenge — Data Science & Modeling Challenger Agent

Data Science & Modeling Challenger Agent argues the recommendation leans on 5 quarters of data from a period with no volume shock. The confidence band overlaps the alternative, and the position does not say what it would take to be wrong. Recorded as a dissent, not a block.

Break conditions

  • ·Override rate rises above 13 percent for two consecutive months
  • ·Cost per model stops falling while volume keeps rising
  • ·A control failure in this lane reaches a customer or a regulator

Evidence gaps

  • ·No external comparator on a like-for-like unit definition
  • ·Exception cases under $22k are sampled, not fully measured
Alternative consideredCostRiskVerdict
Hold the current position$850k run rateKnown and priced. Cedes ground if comparators move faster.live
Raise the ceiling one level now$723k to build controlsOverride rate is 7 percent; raising the ceiling before that settles imports the error into production.rejected on evidence
Move the exception tail to the spine$860k transitionLoses local context. Rework risk on the cases that are hardest to recover.under review
authored by Modeling Orchestratorratifier Director, Data Science3 stated assumptions

Policy register

The rules the agents above are bound by. Version, owner, approver, and the agents each rule constrains.

RefRuleScopeOwner agentApproved byVersionStatus
RD5-POL-200

Data Science & Modeling Decision Envelope

Agents in this lane may act without a human when the model sits inside the stated value, risk and confidence envelope. Outside it, the unit holds at a gate with a named approver and a running clock.

binds 4 agents · EU · UK · US

envelopeModel Governance AgentDirector, Data Sciencev2.7Active
RD5-POL-201

Data Science & Modeling Evidence Standard

Every autonomous decision writes inputs, the rule version applied, the model and prompt version, the output and a reversal path. An action with no evidence record is treated as a control failure, not a fast decision.

binds 4 agents · US · Canada

evidenceModel Governance AgentDirector, Data Sciencev2.3Superseded
RD5-POL-202

Data Science & Modeling Escalation Rule

Escalation is mandatory when confidence falls below the floor, when two options sit inside the confidence band, or when a break condition on a ratified position fires. Director, Data Science owns the response clock.

binds 4 agents · US · Canada

escalationModel Governance AgentDirector, Data Sciencev5.0Active
RD5-POL-203

Data Science & Modeling Autonomy Ceiling

The ceiling for this lane is 83 on the platform scale. Training, validation and drift detection are fully agentic. Promoting a model to production requires a named reviewer, because a model in production is a decision-maker. The ceiling is a governance decision, not a technical limit, and only Director, Data Science can propose moving it.

binds 4 agents · LATAM · US

ceilingModel Governance AgentDirector, Data Sciencev1.7Active