Illustrative worked example

    AI readiness in commercial-loan document intake

    Assess the handoffs before automating the work.

    The workflow being assessed

    In this hypothetical bank, loan operations receives a commercial-loan document packet, checks it against bank-approved product requirements, resolves missing or inconsistent items, and records clearance. The assessment covers document completeness and exception coordination, not underwriting or credit approval.

    1. 01

      Receive

      Record the packet

    2. 02

      Compare

      Use the checklist

    3. 03

      Flag

      Draft exceptions

    4. 04

      Review

      Human decision

    5. 05

      Resolve

      Retain evidence

    Where assistance could fit

    • Intake & review

      AI may assist with
      Classify documents; extract visible fields; link each item to its source and version.
      People retain responsibility for
      Confirm access rights, checklist applicability and ambiguous or unreadable content.

    • Exceptions

      AI may assist with
      Draft missing-item flags, a review summary and a proposed follow-up.
      People retain responsibility for
      Verify every flag; decide exceptions; approve any borrower communication.

    • Resolution

      AI may assist with
      Maintain a draft queue and prepare a source-linked handoff.
      People retain responsibility for
      Clear exceptions, record rationale and approve any change to the system of record.

    A worked exception: the reporting period does not match

    Assumed input

    The checklist asks for a signed statement for the required reporting period. The sample packet contains a prior-period statement, and no signature is located in the reviewed material.

    Draft output

    “Reporting period differs from the checklist; signature needs review.” Each flag identifies the file, version and page. Failed extraction is marked “unable to determine,” not treated as proof that an item is missing.

    Human action

    The reviewer confirms the requirement, checks the source, approves any request for updated material, and records the resolution. AI does not waive the requirement or clear the exception.

    Illustrative findings

    Readiness comes before the build

    These are assumed findings for the example, not findings about any real bank. Each gap has an owner and evidence required before a test can proceed.

    • Ownership

      Illustrative
      Assumed condition
      An operations sponsor exists; authority to clear exceptions is unclear.
      Required action / evidence
      Loan operations names a workflow owner, reviewer, backup and exception decision-maker.

    • Information

      Illustrative
      Assumed condition
      Checklist versions vary; packets can contain superseded or unreadable files.
      Required action / evidence
      Workflow owner approves one product-specific checklist and a versioned document inventory.

    • Access & privacy

      Illustrative
      Assumed condition
      Borrower data is sensitive; the proposed tool and access scope are not approved.
      Required action / evidence
      System and data owners approve access, retention and logging rules. Start with synthetic records.

    • Human control

      Illustrative
      Assumed condition
      Flags may be uncertain; uploaded documents may contain misleading instructions.
      Required action / evidence
      Reviewer checks source evidence. Treat document text as data, not instructions. Uncertainty goes to manual review.

    • Adoption

      Illustrative
      Assumed condition
      Reviewers have no tested review routine or fallback instructions.
      Required action / evidence
      Operations lead tests the review queue, trains users, records feedback and preserves the manual process.

    • Value & proof

      Illustrative
      Assumed condition
      No reliable baseline or agreed acceptance thresholds exist yet.
      Required action / evidence
      Sponsor and reviewers define baseline, test cases, acceptable error levels and decision rules before build.

    Illustrative decision: fix the prerequisites, then test a bounded assist

    Do not approve production automation. Resolve checklist, ownership, access and measurement gaps first. Then consider a Build & Prove Sprint limited to source-linked completeness flags and draft follow-up, with every output reviewed. No credit scoring, underwriting, lending approvals or autonomous borrower messages are in scope.

    What a useful test would have to prove

    Compare AI-assisted review with a reference sample checked by reviewers: net reviewer handling time including corrections; missed exceptions and false alarms; accuracy of source references; and reviewer use of the queue and fallback. Define the sample and acceptance thresholds before testing. No targets or results are asserted here.

    Proceed only if quality, controls and operating value meet the agreed criteria. Refine if errors are correctable; pause for access or control failures; stop if a simpler process change solves the problem better.

    Have a workflow worth examining?

    EAIS’s AI Workflow Readiness Audit helps identify what is worth testing, what needs fixing first, and what evidence leadership needs for a build/no-build decision. Scope and recommendations depend on the workflow.

    Explore the AI Workflow Readiness Audit