What TD automated
The system comes from TD's AI research centre, Layer 6. It coordinates multi-step document verification, connects information across internal systems and drafts case summaries for a human to review. TD expanded the deployment on 13 October 2026 into mortgage funding and mortgage discharge, two stages that sit after the initial credit decision.
In funding, the agent writes the case summary. TD says initial review times there fall by up to half. In discharge, it assembles the documents that back discharge statements, which is clerical work with almost no room for a wrong figure. Neither workflow asks the model to decide who gets a mortgage.
That split is deliberate. The underwriter keeps the cases that need judgment, and TD describes its productivity goal as a doubling in selected mortgage processes rather than a reduction in headcount.
The figures TD published
| Metric | Figure |
|---|---|
| Average pre-adjudication processing time | 15 hours, now under 3 minutes |
| Mortgage funding, initial review | up to 50% time saved |
| Colleague productivity, selected processes | target of 2x |
| Enterprise-wide AI value | $1 billion per year, roughly €0.93 billion at 1.08 USD/EUR |
Two rows are measurements, two are targets. The 15-hour figure and the 50% cut describe what happens in named steps of named workflows. The productivity multiplier and the $1 billion sit in a multi-year plan, and a plan is not an audited result.
What a pipeline like this costs on public APIs
TD runs Layer 6 models on its own infrastructure, so the numbers below are not TD's bill. They are what a European lender would pay if it built the same document pipeline against commercial APIs. Prices are per million tokens, EUR converted at an assumed 1.08 USD/EUR.
| Model | Input $/1M | Output $/1M | Input €/1M | Output €/1M |
|---|---|---|---|---|
| Mistral Large 4 | 1.36 | 4.18 | 1.26 | 3.87 |
| Gemini 4 Argon, introductory | 2.00 | 10.00 | 1.85 | 9.26 |
| Gemini 4 Argon, standard | 4.00 | 20.00 | 3.70 | 18.52 |
| Grok 4.7 | 2.00 | 6.00 | 1.85 | 5.56 |
Take a mid-sized file: 50,000 tokens of documents in, a 5,000-token case summary out. On Mistral Large 4 that is about €0.06 for input and €0.02 for output, so under ten cents per case before retries, retrieval, embeddings and orchestration. At Gemini 4 Argon's standard rate the same job lands near €0.28. Claude Haiku 5.5, DeepSeek-V4.1-Flash and GPT-6.1 Sol are positioned at or below this band, and GLM-5.3 Flash ships under an MIT licence for anyone who wants to host it.
A real agent makes many tool calls per case, so multiply the estimate by the number of steps in your workflow. Even generously, ten cents becomes a euro, and the arithmetic against 15 hours of underwriter time is not close. The constraint on these projects was never the token bill. It is the audit trail.
How EU rules land on a pipeline like this
Under the AI Act, Annex III point 5(b) classifies AI systems used to evaluate the creditworthiness of natural persons or establish credit scores as high-risk. Fraud detection is the listed exception, and document assembly for a discharge statement is not credit scoring. A summary that feeds a lending decision is a different matter, and that is where a European version of TD's system would land.
For a build team the classification is a requirement before it is a filing: it decides what has to be logged, versioned and reviewed. Standalone Annex III obligations now apply from 2 December 2027 and product-embedded Annex I systems from 2 August 2028 under the Digital Omnibus, Regulation (EU) 2026/1744. The general-purpose model grace period is finished: since 2 August 2026 the Commission's AI Office supervises model providers directly, with fines up to €15 million or 3% of worldwide annual turnover. Article 50 transparency duties are already enforceable, which means telling people when they are dealing with an AI system and marking synthetic output in a machine-readable way.
GDPR Article 22 adds a second constraint. A mortgage refusal has legal effects, so a decision made solely by automated processing needs a legal basis and safeguards, including meaningful human involvement. TD's underwriter-in-the-loop structure happens to be the cheaper way to satisfy that, and it is also the structure a European bank would have to document in detail: which model version, which prompt, which record, which human reviewed what.
What a European lender can copy
Start with the two stages TD added in October, not the one it started with. Document assembly and draft summaries are the parts least likely to be classified as creditworthiness evaluation, and they are where the time savings are easiest to measure. Keep the summary as a draft with a named owner. Log every tool call, because a supervisor will ask how a figure reached a funding statement.
Hardware decisions differ by bank. A 16 GB consumer card, the class of GPU we test in our AI Arena benchmarks, is not the machine that runs a mortgage document agent. Model choice matters less than the logging and review layer around it, and that layer is where a European deployment will be judged.
TD's rollout is North American. No part of the system is sold in Europe, and there is no published EU availability plan.
Does TD's agent approve mortgages on its own?
No. TD states that underwriters keep the decisions requiring human judgment. The agent handles document verification, cross-system lookups and the drafting of case summaries for review.
Could a European bank run the same pipeline today?
Yes, for document-heavy steps. A system that influences a creditworthiness assessment moves into Annex III high-risk territory, where the compliance date for standalone systems is 2 December 2027. GDPR Article 22 still requires meaningful human involvement in decisions with legal effects.
Where does the $1 billion figure come from?
It is TD's own target for annual value from its enterprise AI programme over the coming years, not a reported result. The 15-hour to under 3 minutes change and the up to 50% cut in funding review time are the operational figures the bank has published.
