On May 21, 2026, TD Bank Group launched its first agentic AI model into production, and the metric attached to it is the kind that makes other bank CIOs flinch: the pre-adjudication step for mortgages and HELOCs has dropped from an average of 15 hours per file to under 3 minutes — a 99.7% reduction. Not in a pilot. Not on a demo deck. In a live retail lending workflow at a $2.1 trillion-asset bank with 28.1 million clients. The agent — built by Layer 6, TD's AI research lab acquired in 2018 — classifies client documents, extracts and validates income, runs consent checks, hunts for discrepancies against policy, and writes the summary memo a human underwriter signs off on. Sandra Aziz, the Layer 6 ML engineer who led the build, put it bluntly: "We built where nothing else existed. Everything is new."
This matters for two reasons that go well beyond TD. First, mortgage origination is the textbook hard case for agentic AI in regulated finance — high-stakes credit decisions, heavy documentation, consumer protection rules, and disparate-impact risk. If a Big Six Canadian bank can put an agent into that workflow under a named governance regime, the playbook is now copyable. Second, the deployment lands ten weeks before the EU AI Act's August 2, 2026 full enforcement date, which classifies credit scoring as a high-risk system requiring documented human oversight, explainability, and risk management. TD's "Trustworthy AI" approach is not coincidentally aligned. Every CIO and CFO in regulated lending should be reading this announcement as a forward indicator of what their own examiners will start asking about by Q4.
What Changed: A Production Agent Inside a Regulated Credit Workflow
The headline number — 15 hours to under 3 minutes — is the kind of statistic that gets shared on LinkedIn and immediately demands skepticism. Three details from the TD Stories deep-dive and the official announcement make it credible.
First, the scope is bounded. The agent does not approve the loan. It does pre-adjudication — the document gathering, data extraction, income calculation, consent verification, and discrepancy detection that historically chewed up underwriter hours before a human ever made a credit decision. The output is a structured summary memo handed to a human underwriter. That bounded scope is the most important architectural choice in the entire announcement.
Second, the team and timeline are real. Sandra Aziz joined Layer 6 as Technical Product Owner in March 2025 and led a cross-functional team of research scientists, data scientists, and TD process experts. The launch came roughly 14 months later. That is not a quarterly experiment — that is a deliberate, multi-quarter productization effort using TD's own bank knowledge and Layer 6's research depth. Layer 6 itself is not new infrastructure; it has been operating as TD's AI center of excellence since the $100 million acquisition in January 2018, led by founders Jordan Jacobs, Tomi Poutanen, and Maks Volkovs (who also co-founded the Vector Institute).
Third, the governance regime is documented and pre-existing. TD's Trustworthy AI program was named Best Responsible AI Program (North America) by DataIQ in 2025 — meaning the policy infrastructure, risk-rating taxonomy, and human-in-the-loop discipline were already in place before this agent was scoped. The Trustworthy AI team evaluates models on privacy, security, fairness, accountability, and explainability before customer contact, and continues monitoring post-deployment. The agent is not a bolt-on; it is a graduate of a governance pipeline.
The financial framing is equally specific. Mohit Veoli, TD's SVP of Real Estate Secured Lending, told Fintech.ca that agentic AI delivers "what clients tell us matters most — speed and simplicity." Chief Analytics and AI Officer Luke Gee positioned this within a larger commitment: TD has targeted approximately $1 billion in annual AI value, with roughly $500 million in annualized revenue generation by 2028. This agent is the first production proof point against that number, not a feel-good announcement.
The agent is also explicitly Step 1. TD has signaled an end-to-end rebuild of Real Estate Secured Lending — from document submission through funding release — with agentic AI at each stage. That is a multi-year program, not a single deployment, and it sets a competitive pace that other Canadian and US lenders will now be measured against.
Why This Matters: Two Scoreboards, One Window
For the CIO / Chief AI Officer. TD's announcement is a working answer to the question every bank CIO is currently being asked in board prep: what does a real, in-production agentic AI workflow look like in a regulated business line, and what does our governance regime need to look like to survive an examiner? The answer has three components: (1) a bounded, non-decisioning agent scope; (2) a pre-existing trustworthy-AI taxonomy that the model is graded against before launch; and (3) explicit human-in-the-loop architecture, where the agent writes a memo and the human signs the credit decision. That triad is what makes the deployment defensible in a regulator-friendly way.
The architectural lesson is that the model is not the moat — the workflow integration is. Document classifiers, income calculators, and consent checkers are not novel capabilities in 2026. What is novel is gluing them into a single agent that operates inside TD's real document pipelines, with policy validation rules grounded in TD's actual underwriting standards. Sandra Aziz's "we built where nothing else existed" comment is about the integration layer, not the model. Most banks that try to copy this will fail not at the AI, but at the surrounding plumbing — the document ingestion pipeline, the policy rule encoding, the consent and audit logging, the model monitoring. That is also where most of the 14-month build time went.
For the CFO and the Board. The economics of mortgage pre-adjudication are unforgiving. The Mortgage Bankers Association has consistently reported that the all-in cost to originate a loan in the US exceeds $11,000, and McKinsey has estimated that AI-driven workflow automation can deliver up to 20% cost savings, with some lenders reporting 30–50% reductions in operational expenses. Industry vendors are reporting 4x underwriter productivity (Candor) and 7+ hours saved per file (Zeitro). TD's 15-hour-to-3-minute number is on the aggressive end of those benchmarks, but for one bounded step rather than the full origination cycle.
The CFO question is not "should we copy TD?" but "what is our equivalent first-step bounded workflow, and what is the ROI shape across our mortgage volume?" The math is meaningful: at typical mid-size bank volumes of 100,000 mortgages per year, with $200/hour fully-loaded underwriter cost, saving 12 hours of pre-adjudication labor per file translates to $240 million in annualized capacity recovery — most of which gets reinvested as throughput, not headcount reduction. That is the Gartner finding that 90% of finance functions will deploy AI by 2026, but fewer than 10% will see headcount reductions. Capacity is the dividend, not severance. TD's own framing of "helping clients get to a 'yes' faster" is consistent with that — the value goes to throughput, conversion, and competitive win rate, not to FTE reduction.
The dual scoreboard matters because the temptation, for both CIOs and CFOs, is to over-index on the model and under-invest in the surrounding governance and integration work. TD's announcement is a quiet warning shot: the bank that built its AI governance program before it built the agent is the one that gets to ship into production. Everyone else is now playing catch-up against a regulatory clock that closes in August.
Market Context: A Canadian Lead in a Crowded Race
TD is not alone in the agentic-AI-in-banking land grab, but the deployment patterns differ in ways that matter. The competitive picture as of May 2026:
| Bank | Use Case | Reported Outcome | Source / Approach |
|---|---|---|---|
| TD Bank (Canada) | Mortgage pre-adjudication agent | 15 hr → <3 min; ~$1B AI value target by 2028 | Layer 6 in-house; Trustworthy AI governance |
| RBC (Canada) | ATOM credit underwriting | $700M–$1B incremental value target by 2027 | Internal build; $2B/yr tech spend |
| CIBC (Canada) | CRTeX client personalization | $1B+ in new deposits since launch | First Canadian bank to sign federal generative AI code |
| Scotiabank (Canada) | Scotia Intelligence assistive AI | 70% manual work reduction (per company disclosure) | Internal platform |
| BMO (Canada) | Institute for Applied AI & Quantum | $1B+ PPPT target by 2030 | Research-led, multi-year |
| JPMorgan Chase (US) | 500+ internal AI use cases; predictive default | 10%+ pilot reduction in mortgage delinquency | LLM Suite to hundreds of thousands of employees |
| Wells Fargo (US) | Loan re-underwriting agent network | 5-day approvals → 10 minutes (reported) | Google Agentspace; multi-agent system |
Three signals stand out. First, the Big Five Canadian banks are running parallel programs, and they are not shy about putting public dollar targets on the table — collectively, the Canadian banks have signaled at least $4–5 billion in projected AI value over the next 3–5 years. Second, US institutions have leaned heavily into platform partnerships (Wells Fargo on Google Agentspace, JPMorgan on its own LLM Suite, Anthropic into Wall Street), while TD is notable for being fully in-house through Layer 6. Third, the publicly disclosed production deployments — not pilots — remain rare. TD is one of the few to put a specific time-savings number on a named retail credit workflow.
The analyst overlay is telling. Forrester has projected that human visits to financial-institution websites will drop 20% by 2026 while machine-initiated traffic surges 40%, and Gartner expects 90% of finance functions to deploy AI by 2026. Adoption is no longer the question; what differentiates the winners is the depth of process integration and the durability of the governance regime. TD is shipping against both axes simultaneously, which is the meaningful part of the announcement.
There is also a production-gap warning underneath all this. McKinsey, Gartner, and HCLTech have all published variants of the same finding: somewhere between 40% and 95% of enterprise AI pilots fail to reach production. The banks that succeed are the ones that pre-built governance, picked bounded workflows, and treated the model as the easy part. TD's announcement is, in that sense, an existence proof that the playbook works — and an implicit benchmark for everyone still stuck in pilot mode.
Framework #1: The Mortgage AI Pre-Adjudication ROI Calculator
Most bank CIOs will read TD's announcement and immediately ask: what is the ROI shape for my book? The TD-style deployment has a deceptively clean economic model — labor hours saved per file × annual file volume × loaded underwriter cost — but the realistic build cost and time-to-value vary by institution scale. Here is the model, parameterized for three bank sizes.
Inputs (industry-typical, 2026):
- Pre-adjudication labor saved per file: ~12 hours (assuming TD's 15-hour baseline, conservative 12-hour realization for a copyist deployment)
- Fully-loaded cost per underwriter hour: $200/hr (US/Canada blended)
- One-time build cost: $8M–$50M (model dev + integration + governance instrumentation)
- Annual run cost: 15–25% of build cost (model ops, monitoring, retraining, audit)
Scenario A — Regional bank (10,000 mortgages/year):
- Annual labor capacity recovered: 10,000 × 12 hr × $200 = $24M
- Build cost: $8M one-time; $1.6M/yr run
- Year 1 net: $24M − $9.6M = $14.4M
- Payback: ~4 months post-deployment
- 3-year ROI: ~580%
Scenario B — Mid-size bank (100,000 mortgages/year):
- Annual labor capacity recovered: 100,000 × 12 hr × $200 = $240M
- Build cost: $20M one-time; $4M/yr run
- Year 1 net: $240M − $24M = $216M
- Payback: ~5 weeks post-deployment
- 3-year ROI: ~3,500%
Scenario C — Tier-1 national bank (1,000,000 mortgages/year, TD/RBC/JPM scale):
- Annual labor capacity recovered: 1,000,000 × 12 hr × $200 = $2.4B
- Build cost: $50M one-time; $10M/yr run
- Year 1 net: $2.4B − $60M = $2.34B
- Payback: ~9 days post-deployment
- 3-year ROI: ~14,000%
Reality check (read this before you put it in a board deck):
Three corrections most CFOs miss. First, the capacity recovered is almost never converted 1:1 to cost. Banks typically reinvest 70–80% of recovered hours into throughput (faster decisioning, more files per underwriter), conversion (faster yes-rates win business), and quality (more time on exceptions). The reported P&L impact is therefore closer to 20–30% of the gross savings number, with the remainder showing up as revenue acceleration. TD's own framing — "$500M annualized revenue generation" as part of the $1B AI value target — is consistent with that pattern. Second, the 12-hour labor savings assumes the bank already has a TD-style document pipeline, policy rule library, and consent infrastructure. If those are missing, build cost climbs 2–3x and Year-1 ROI compresses sharply. Third, the governance cost (Trustworthy AI evaluation, audit logging, ongoing fairness monitoring) is not a one-time bill — it is the run cost, and it grows with model surface area. Banks that under-budget the governance line item are the ones whose models get pulled by Legal six months in.
For most CIOs, the takeaway is: the gross savings math is huge, the realized savings math is real but more modest, and the path to capturing it depends on integration discipline rather than model selection. You can read deeper benchmarks across other banking workflows in our 10 Banking AI Use Cases with Real ROI Benchmarks for 2026.
Framework #2: The Regulated Lending Agent Deployment Checklist
If TD's 14-month, Layer-6-led, Trustworthy-AI-graded build is the reference architecture, the operational question is: what does the equivalent program look like for our bank, and what do we need to be true on Day 1 to ship inside twelve months? The checklist below distills the TD pattern into a regulated-lending agent deployment playbook, grouped by readiness pillar. Treat each item as binary (yes / not yet); fewer than 18 of 22 yeses is a red flag for a 2026 production launch.
Governance & Risk (must be in place before model work begins)
- Named AI risk taxonomy with documented risk tiers (low / medium / high / restricted)
- Trustworthy / Responsible AI policy approved by board risk committee
- Pre-launch evaluation gate covering privacy, security, fairness, accountability, explainability
- Post-deployment monitoring program (model drift, fairness drift, error rates, escalation rates)
- Documented human-in-the-loop authority — what the agent cannot do, written in code, not just policy
- EU AI Act / OCC / OSFI mapping for the specific workflow (high-risk Annex III if credit scoring)
Data & Document Pipeline
- Centralized document ingestion with classification taxonomy (income docs, ID, consent, policy artifacts)
- OCR / IDP layer with measured accuracy benchmarks per document type
- Income calculation rule library aligned with current underwriting policy
- Consent and entitlement service with audit trail
- Discrepancy detection rules that map to your live policy, not a generic template
Model & Integration
- Bounded agent scope — no credit decisions, only memo generation for human underwriter
- Structured output schema (so downstream systems can consume the memo deterministically)
- Versioned prompt, model, and tool registry with rollback capability
- Side-by-side evaluation against the human baseline before any auto-routing
- Latency and cost SLOs documented and monitored
Talent & Operating Model
- Cross-functional pod: ML engineering, data engineering, business policy, risk, legal
- Technical product owner with end-to-end accountability (Sandra-Aziz-equivalent role)
- Underwriter feedback loop — measured weekly, with structured failure-mode tagging
- Quarterly governance review with risk committee
- Internal communication plan that addresses fear-of-replacement before deployment
- Vendor and model independence — no single-supplier lock-in on the critical path
Most banks today will check 10–14 of these. That is also a reasonable explanation for why the pilot-to-production failure rate in banking remains so high. The list is the work — the model is the easy part.
Case Study: 14 Months Inside Layer 6
The most interesting datapoint in the TD announcement is not the 15-hour-to-3-minute headline; it is the 14-month build. Sandra Aziz joined Layer 6 as Technical Product Owner in March 2025; the agent shipped in May 2026. Inside those fourteen months, Layer 6 ran a cross-functional team of research scientists, ML engineers, and embedded TD operators from Global Technology & Solutions, Real Estate Secured Lending, and risk management.
What stands out is what TD did not do. There is no public mention of an external consulting partner on the build, no signed-up systems integrator, no AI-platform vendor in the credits. Layer 6 is the moat. The acquisition cost in January 2018 — approximately $100 million according to Globe and Mail reporting — looks, eight years later, like one of the better technology acquisitions in Canadian banking. Layer 6's three founders (Jordan Jacobs, Tomi Poutanen, Maks Volkovs) brought research credibility from the University of Toronto deep-learning ecosystem and co-founded the Vector Institute, giving TD continuous access to a research bench other banks have to rent.
The takeaway for institutions without an in-house Layer-6-equivalent is sober. The TD deployment time was 14 months with a mature research lab, a mature governance program, and a mature talent base. A peer bank starting from a generic GenAI center of excellence and a Big Four implementation partnership should plan for 18–24 months and a higher integration bill. The Sandra Aziz quote — "we built where nothing else existed. Everything is new" — applies to every bank that tries this, regardless of who built Layer 6 or what vendor logo is on the slide. That is the durable lesson.
A second lesson concerns the human side. Luke Gee's framing of "a hybrid future where our colleagues and AI work together" is not a PR line; it is the deployment topology. Underwriters did not lose their jobs to the agent. They gained a structured memo and 12 hours of reclaimed file capacity. The TD agent does the gathering, sorting, and summarization. The human signs the loan. That topology is what makes the EU AI Act's Article 14 human-oversight requirement structurally satisfied rather than something that has to be retrofitted after the regulator calls.
What to Do About It
For the CIO / Chief AI Officer (next 60 days). Pick one bounded, non-decisioning workflow inside a regulated business line — pre-adjudication, KYC summary generation, policy validation memo writing — and stand up a Trustworthy-AI evaluation gate for it before the model work begins. If your current governance framework cannot answer the EU AI Act Article 14 human-oversight question for that workflow today, fix that first. Treat the model as the easy 20% of the program.
For the CFO / Risk Officer (next quarter). Pressure-test the ROI math against your own file volume and underwriter cost. Be skeptical of any vendor pitch that promises a TD-style number without specifying realized vs. gross savings, run cost as a function of scale, and the integration cost in your specific document and policy environment. The TD case is real; the temptation to over-quote its savings on your own deck is also real.
For Business Leaders (next 6 months). The competitive window in regulated lending is closing. Of the seven banks in the comparison table above, six have public AI value targets. Customer expectations are being reset by the leader's deployment ("yes faster"), and the EU AI Act enters full enforcement on August 2, 2026, forcing the laggards onto the same governance footing as the leaders. Build the pod, name the technical product owner, and ship the first bounded agent before the year-end planning cycle starts. The cost of waiting is no longer the model price — it is the conversion-rate gap against banks that are already shipping.
The TD announcement is a small one in terms of dollar spend, but it is the kind of milestone the rest of the industry will measure itself against. Three minutes is the new floor. Fifteen hours is now visible debt on every other bank's balance sheet.
