MMasters in Business
← All frameworks
Innovation

Governed Human-in-the-Loop AI Adoption

Automate repeatable work while humans retain judgment and ownership

Difficulty
Advanced
Time to result
~months to results
Steps
6
Confidence
98%

Start AI adoption with repetitive processes that draw from a stable information base, such as proposal responses, commentary drafting, or document review. Before deployment, establish infrastructure, privacy, cybersecurity, access, and governance controls suitable for the organization’s risk level. Let AI retrieve, synthesize, or draft, but require a qualified human to verify factual accuracy, relevance, and policy compliance. The human remains accountable for the final product and any portfolio or business decision. Measure both efficiency and failure modes, then widen scope gradually. More advanced tools can augment research by surveying many reports, but they should support—not replace—the professional’s pressure testing and judgment.

Origin

Heinel described State Street’s progression from long-standing machine-learning use to governed generative-AI applications in repeatable operations and research augmentation. Extracted from Masters in Business.

Core principles

  • 01Begin with repeatable, information-rich processes.
  • 02Governance and infrastructure precede scaled deployment.
  • 03AI should free people for higher-order work.
  • 04Humans must pressure-test outputs and own final decisions.
  • 05Augmentation is safer than premature autonomous judgment.

How to run it

  1. 1

    Establish the Control Foundation

    Define approved data, access rules, privacy protections, cybersecurity controls, audit requirements, and accountable owners before scaling AI use.

  2. 2

    Choose a Bounded Workflow

    Select a repeatable, information-heavy task with clear inputs, outputs, reviewers, and quality criteria. Favor work that currently consumes skilled time without requiring final human judgment at every intermediate step.

  3. 3

    Ground the AI

    Connect the system to approved source material and instruct it to retrieve or synthesize only within the permitted scope. Preserve traceability to underlying information.

  4. 4

    Keep a Human Owner

    Assign a qualified person to inspect the output, correct errors, and approve the final product. Make ownership explicit rather than treating review as informal cleanup.

  5. 5

    Measure the Pilot

    Track time saved, factual errors, reviewer edits, escalations, user satisfaction, and incidents. Compare results with the previous process.

  6. 6

    Expand Through Augmentation

    Extend proven controls to adjacent tasks and higher-order research tools. Continue to reserve consequential interpretation and final decisions for accountable professionals.

In the wild

AI-Assisted RFP Responses

A client question arrives in unfamiliar wording. AI searches an approved database, retrieves the most relevant prior material, and drafts a response. A human checks accuracy, suitability, and compliance before owning the final answer.

→ The team responds faster while preserving professional accountability.

Research Copilot for Portfolio Managers

A portfolio manager uses a research copilot to survey dozens or hundreds of reports and surface relevant findings. The manager tests whether the synthesis makes sense and independently decides whether it should affect the portfolio.

→ Research breadth and speed increase without delegating the investment decision to the model.

Common mistakes

Starting with Autonomous Decisions

Consequential judgment is a poor first deployment target. Prove reliability on bounded augmentation tasks before increasing autonomy.

Measuring Only Time Saved

A faster process can still be worse if errors, privacy exposure, or weak review increase. Track quality and risk explicitly.

Using Human Review as a Label

The reviewer must have enough expertise, time, and authority to reject the model’s output rather than merely approve it.

Is it for you?

Best for

Regulated or high-stakes organizations introducing AI into operational and research workflows.

Not ideal for

Unbounded autonomous decisions where errors cannot be reviewed, reversed, or attributed.

From the transcript

“the biggest places that we're seeing AI support our business are in more things that are operational in nature that are repeatable processes where we…”

Lori Heinel

“you have a human in the loop always in our environment today that ultimately owns the final product.”

Lori Heinel

“they still have to you know pressure test whether the results are getting back make sense and they still ultimately make the decision about what…”

Lori Heinel

From the episode

Balancing $5.7T in Active and Passive Management with Lori Heinel

→