TD Bank Group

Mixed evidenceStep before the decisionDeterministic control

Mortgage and HELOC pre-adjudication and application memo preparation

That preparing a mortgage or HELOC file for a credit decision requires a human to read the document package end to end - classifying documents, extracting figures, calculating income, checking consents and writing the summary - because borrower documents are too variable for a rules engine to handle reliably.

Credit and lending · Canada (TD Real Estate Secured Lending operations)

Collections: Human still decides · Regulated autonomy · Embodied work · Negative results

A mortgage analyst surrounded by document queues contrasts with a human credit adjudicator receiving one prepared dossier.

Executive brief

The operating-model shift, in one view.

TD picked the step before the decision, not the decision. The agent does the document work - classify, extract, calculate, check consents, find discrepancies, draft the memo - and the credit adjudicator still owns the lending call. That boundary is why this shipped in a regulated process. Two design choices are worth copying regardless of industry. First, TD refused to let the language model do arithmetic and forced it to call a deterministic calculator, turning a well-known failure mode into an explicit control rather than a residual risk. Second, TD chose a first use case with clear inputs, clear outputs and a human already in the loop, and said so - the value was real but the point was to build an operating model it could reuse.

AI value · Pre-adjudication processing time per mortgage or HELOC application (document review through summary memo generation)

Company-reported

Average preparation time: 15 hours → under 3 minutes

Before

Submits the purchase agreement, ID, account statements, and proof of income. → Classifies documents, finds relevant figures, verifies details, calculates income, and writes the summary. → Reviews the human-prepared summary and makes the lending decision. → Waits for the pre-adjudication result, which took an average…

After

Submits the same document package. → Classifies documents, extracts key data, calculates income, checks consent, finds discrepancies, and drafts a summary. → Receives the AI-generated summary memo and makes the lending decision. The decision right is unchanged. → Tests privacy, security, fairness, accountability,…

Human boundary

Why it matters

Note what TD has not published: the number of applications treated, any accuracy figure, and where the freed reviewer capacity went.

How the work changed

Before

How the work ran before the change.

Step 1 of 4
Borrower
Submits the purchase agreement, ID, account statements, and proof of income.
Step 2 of 4
Human reviewer / mortgage operations analyst
Classifies documents, finds relevant figures, verifies details, calculates income, and writes the summary.
ControlManual verification against policy requirements; consent checks performed by hand.
Step 3 of 4
Credit adjudicator
Reviews the human-prepared summary and makes the lending decision.
ControlCredit policy and adjudication authority.
Step 4 of 4
Borrower
Waits for the pre-adjudication result, which took an average of 15 hours of processing.

What changed

Preparing a decision file does not require a human to read every page.

Decision rightAI prepares; the human adjudicator decides

After

How the same work runs now.

Step 1 of 4
Borrower
Submits the same document package.
ControlDocument checklist unchanged.
Step 2 of 4
Autonomous AI agents built by TD's Layer 6
Classifies documents, extracts key data, calculates income, checks consent, finds discrepancies, and drafts a summary.
ControlDeterministic tools (calculator or spreadsheet) are mandated for arithmetic rather than the language model; TD-specific term and acronym definitions supplied; guardrails restrict the model to an approved tool library.
Step 3 of 4
Credit adjudicator
Receives the AI-generated summary memo and makes the lending decision. The decision right is unchanged.
ControlCredit policy and adjudication authority retained by the human; an independent commentator describes the design as one where 'a human still needs to be accountable for the final decision'.
Step 4 of 4
Trustworthy AI team (led from Layer 6)
Tests privacy, security, fairness, accountability, explainability; monitors performance, accuracy, answer relevance.
ControlRigorous pre-deployment check series; post-deployment monitoring with defects routed to future releases.

Process model built from the published workflow evidence for TD Bank Group. Every step, actor, and control appears in full below.

Every step, actor, and control

Exception path

Not published as a discrete exception workflow. The structural exception path is the human adjudicator, who receives the memo and decides; TD chose the use case specifically because it has 'clear inputs, clear outputs, and a human in the loop'. Post-deployment, the Trustworthy AI team monitors models and routes observed mistakes to future releases. TD does not disclose what happens to an individual file when the agent cannot classify a document, cannot compute income, or flags a discrepancy it cannot resolve.

Work removed

Manual classification of the borrower's submitted document package
Reading and searching documents of up to 100 pages to locate specific figures
Manual extraction and transcription of key application data

Decision authority

The lending decision right remains with the human credit adjudicator. American Banker states plainly that the agent 'handles these steps and presents the results to a credit adjudicator to make a decision'. The agent's own authority is bounded in two published ways: it must route arithmetic to deterministic tools rather than compute it, and it may only use an approved library of tools ('here is the library of things that you can use, here is what they are for'). Validation is against 'select policy requirements' - TD does not disclose which policy checks are in scope, so the boundary of the agent's validation authority is only partly public.

Before

01
Borrower
Submits the purchase agreement, ID, account statements, and proof of income.
Control: Document checklist; incomplete packages returned.
02
Human reviewer / mortgage operations analyst
Classifies documents, finds relevant figures, verifies details, calculates income, and writes the summary.
Control: Manual verification against policy requirements; consent checks performed by hand.
03
Credit adjudicator
Reviews the human-prepared summary and makes the lending decision.
Control: Credit policy and adjudication authority.
04
Borrower
Waits for the pre-adjudication result, which took an average of 15 hours of processing.
Control: Service-level expectations only.

After

01
Borrower
Submits the same document package.
Control: Document checklist unchanged.
02
Autonomous AI agents built by TD's Layer 6
Classifies documents, extracts key data, calculates income, checks consent, finds discrepancies, and drafts a summary.
Control: Deterministic tools (calculator or spreadsheet) are mandated for arithmetic rather than the language model; TD-specific term and acronym definitions supplied; guardrails restrict the model to an approved tool library.
03
Credit adjudicator
Receives the AI-generated summary memo and makes the lending decision. The decision right is unchanged.
Control: Credit policy and adjudication authority retained by the human; an independent commentator describes the design as one where 'a human still needs to be accountable for the final decision'.
04
Trustworthy AI team (led from Layer 6)
Tests privacy, security, fairness, accountability, explainability; monitors performance, accuracy, answer relevance.
Control: Rigorous pre-deployment check series; post-deployment monitoring with defects routed to future releases.

Work that left the path

Manual classification of the borrower's submitted document package
Reading and searching documents of up to 100 pages to locate specific figures
Manual extraction and transcription of key application data
Manual income calculation and annualisation from pay stubs and rental income
Manual consent checks
Manual inconsistency and discrepancy searching across the package
Hand-writing the application summary memo for the adjudicator

Human role before

A human reviewer performed the entire pre-adjudication package: document classification, extraction from long documents, verification, inconsistency checking, income calculation and summary writing - work TD's own AI lead characterised as 'fairly arduous tasks, looking through various IDs, hunting through 100-page documents for that key piece of information'.

Human role after

The preparation work moves to the agents. The credit adjudicator's role narrows to judging a complete, more accurate summary and making the lending decision. TD frames the intent as reshaping reviewer work 'into something that's much more closely aligned with delivering on the brand promise', and describes a hybrid model where colleagues and AI work together; TD has not published a redeployment or headcount outcome, so the destination of the freed reviewer capacity is not established.

AI roleAutonomous preparation agent for the pre-adjudication step. It performs document classification, information extraction, income calculation, validation against selected policy requirements, consent checks and discrepancy detection, then drafts the application summary memo for the adjudicator. It does not make or recommend the credit decision.

Outcomes

Pre-adjudication processing time per mortgage or HELOC application (document review through summary memo generation)

Company-reported

An average of 15 hours→An average of less than three minutes

Early results from production deployment in January 2026, disclosed 2026-05-21 and reported 2026-05-28 - an observation window of approximately four months. TD does not state the measurement window explicitly. · TD Canada Real Estate Secured Lending mortgage and HELOC pre-adjudication, in production. The treated application count is not disclosed. For context on the workflow's population, TD states it underwrites hundreds of thousands of mortgages annually and recorded $826 million in US residential mortgage origination volume in 2025 - both are book-size figures, not the treated cohort.

TD explicitly labels these 'Early results', and that hedge must travel with the figure. Both the baseline and the result are TD's own averages; TD does not publish the measurement definition, the sample size, or the variance. American Banker restates the figures rather than measuring them independently, though it adds independent expert commentary. The treated population size is undisclosed, so the aggregate hours saved cannot be derived. The metric covers pre-adjudication only - not application-to-funding cycle time - and TD states this is 'an impactful but small chunk of the overall journey towards funding a mortgage'.

Accuracy and cost of the pre-adjudication output

Company-reported

Accuracy and unit cost of the previous manual human review, neither quantified by TD→TD states the agent provides 'more accurate results and thereby reducing both risk and the unit cost on adjudication'; the mortgage officers 'get a more complete package, they get a more accurate summary at the very end'

Early results from the January 2026 production deployment, disclosed 2026-05-21 · TD Canada RESL mortgage and HELOC pre-adjudication; no quantified accuracy delta, error rate or unit-cost figure disclosed

Directional and entirely unquantified. TD asserts improved accuracy and reduced unit cost without publishing a baseline error rate, a post-deployment error rate, or any cost figure. This outcome is recorded because the accuracy claim is central to TD's own framing, but it cannot support a numeric claim and should not be presented as a measured result. An independent expert quoted by American Banker separately notes that human verification of AI output carries its own burden and that reviewers 'can get bored and fail to check everything carefully' - a risk TD does not address publicly.

What leaders can reuse

Anti-pattern

Multiplying a per-application average by the institution's total book. TD discloses 15 hours to under three minutes per application and, separately, that it underwrites hundreds of thousands of mortgages annually - but the treated application count is not published, and those two numbers must not be combined into an aggregate saving. The second anti-pattern is treating the human-in-the-loop as a control without funding it: an independent expert quoted by American Banker warns that verification can take longer than expected and that reviewers 'can get bored and fail to check everything carefully'. An organisation that automates preparation, keeps a nominal human approver, and then measures only cycle time has moved the risk rather than managed it.

Questions

01In our highest-volume document workflow, which step precedes the decision - and could we automate only that step without touching the decision right?
02Where are we letting a language model perform precision-critical operations it is structurally bad at, and could we force a deterministic tool call instead?
03If we published a per-unit improvement, would we also be able to publish the treated population, or would the number be unusable?
04Who verifies the agent's output, how long does that verification actually take, and have we measured whether verification quality degrades over time?
05Which policy checks are we comfortable letting an agent validate, and which must stay human - and is that boundary written down?
06Is our first agentic use case chosen to prove an operating model we can reuse, or to chase the largest available benefit?

Portability conditions

The workflow is document-heavy with relatively standardised inputs and a clear response deadline
A human accountable decision-maker already exists at the end of the process and stays there
Arithmetic and other precision-critical steps can be delegated to deterministic tools the agent calls rather than performs
Domain terms, acronyms and policy definitions can be supplied to the model as reference material
A standing responsible-AI function can gate release and monitor the model after deployment
The organisation is willing to publish a per-unit metric with an honest 'early results' hedge rather than waiting for a programme-level number
Cross-functional build capacity exists - business subject-matter experts, scientists, engineers and risk partners together

Reputation risk

low - TD is describing its own regulated process with a named executive, a primary release and independent trade-press corroboration, and the decision right demonstrably stays with the human adjudicator. The residual risks are presentational: the 'early results' hedge must be carried, the undisclosed treated population must not be used to imply an aggregate benefit, and the unquantified accuracy and unit-cost claims must not be presented as measured. No customer, employee or confidential information is involved.

Evidence and authority

What the public record supports.

Current · updated 2026-08-22

1 independent, 1 primary; publication outcomes are verified and reported.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 8574a74693efd62c

Sources

TD Launches Agentic AI to Transform Real Estate Secured Lending from End to EndTD Bank Group (TD Stories / CNW) · 2026-05-21 · primary
TD's new AI agent shaves 15 hours off mortgage decisionsAmerican Banker · 2026-05-28 · independent

Read the evidence, freshness, caveat, and version policy.

Mortgage and HELOC pre-adjudication and application memo preparation

The operating-model shift, in one view.

How the work changed

Outcomes

Pre-adjudication processing time per mortgage or HELOC application (document review through summary memo generation)

Accuracy and cost of the pre-adjudication output

What leaders can reuse

What the public record supports.

Related transformations

Smart Gating machine-learning gate assignment

Teacher-controlled ChatGPT lesson planning with implementation guide

GenAI Quick Quote preliminary assessments

AI-automated first-line credit analysis in Home Equity origination

Sources