起步顺序Order of work 先把价值算清,再让助手升级为办事,最后才把 Prism 推进到高风险决策。Count the value first, then upgrade assistants to tasks, then extend Prism to higher-risk decisions.
公开事实Public facts 已核对Checked
- C$1B2026 年 AI 价值目标AI value target for 2026TD 官网自述Stated on TD's website
- 200+AI 研发中心 Layer 6 的团队规模People at Layer 6, its AI research centre2018 年收购,当时约 20 人Acquired in 2018, about 20 people then
- 25,000+配备生成式 AI 生产力工具的员工Employees with generative AI productivity tools含 GitHub Copilot、Microsoft 365 CopilotIncluding GitHub Copilot and Microsoft 365 Copilot
- 40%+TD 称 TD AI Prism 使营销响应率提升Campaign response lift TD reports for TD AI Prism自研预测基础模型,2025 年 6 月In-house forecasting model, June 2025
需要完成的三个转变Three shifts to make 判断Judgment
- 助手只回答问题Assistants that answer questions → 智能体办完任务,由人确认Agents that finish tasks, with a person confirming
- 只做营销预测Forecasting for marketing → 分级之后,进入信贷与合规Credit and compliance, after risk tiering
- 两万多个工具席位25,000 seats of tools → 一本能证明 10 亿加元的台账A ledger that proves the C$1B
专家组裁决Expert panel 依据公开观点的推演Inference from public views
Foteini Agrafioti
每个智能体都挂到业务目标上;Layer 6、平台与业务条线的权责写成一页纸。Tie every agent to a business target, and write one page of ownership between Layer 6, platform and lines.
Erik Brynjolfsson
先设基线和对照组再数 10 亿;预算要含流程再造,而不只是工具。Set baselines and a control group before counting the billion; budget for process redesign, not just tools.
Ziad Obermeyer
信贷与反洗钱:先写清真实目标与代理指标;所有模型先建清单。For credit and AML, name the real target and the proxy first; inventory every model.
Ethan Mollick
在真实分行任务上先测出智能体会在哪失败再放大;保留人工确认。Test where agents fail on real branch tasks before widening; keep a person confirming.
Andrew Ng
先选一个任务(如改地址),三到六个月内出数字。Pick one task, such as address change, that can show numbers in three to six months.
Thomas Davenport
这是服务与信贷流程的改变,不是再加一个工具;点名流程负责人。This is a process change in service and credit, not another tool; name the process owners.
Ethan Mollick:尽早上真实任务,摸清边界。Get agents onto real tasks early to find the edge.
Ziad Obermeyer:信贷模型先过目标与清单关。Gate credit models on target and inventory first.
→ 低风险任务走快速通道;信贷与合规走完整审查。Fast lane for low-risk tasks; full review for credit and compliance.
六位专家是谁、怎样思考、有什么铁律 →Who the six experts are, how they think, their rules →
三个目标Three priorities 建议Recommendation
把 10 亿加元一行一行算清Prove the billion, line by line
按业务条线建价值台账,写明基线与对照组。A value ledger by business line, with baselines and a control group.
让助手办完任务Let assistants finish tasks
先在客服中心与分行各做两个简单任务,由人确认。Start with two simple tasks in the contact centre and branches; a person confirms.
给 Prism 加一道关再推进Extend Prism with a gate
信贷与合规须先过风险分级与「目标 vs 代理指标」检查。Credit and compliance only after risk tiering and a target-versus-proxy check.
场景组合Scenarios 建议Recommendation
| 场景Scenario | 做法Approach | 状态Status |
|---|---|---|
| 客服中心任务智能体Contact-centre task agent 提效Efficiency体验Experience | 端到端办理换卡、改地址,由人确认。Replace a card or change an address end to end, with a person confirming. | 建议 · 基于已上线助手Proposed · builds on live assistants |
| 分行任务助手Branch task assistant 提效Efficiency体验Experience | 分行助手从答问升级为准备并提交业务。Move the branch assistant from answering to preparing and submitting requests. | 建议Proposed |
| Prism 进入信贷与反洗钱Prism for credit and AML 降风险Risk增收Revenue | 分级与目标检查之后,扩展到信贷与反洗钱筛查。Extend forecasting to credit and AML screening after tiering and target checks. | 建议 · 完整审查Proposed · full review |
| 财富与保险Wealth and insurance 增收Revenue降风险Risk | 利用 2,810 万客户规模做个性化投顾辅助与核保自动化。Personalised advice support and underwriting automation across a 28-million customer base. | 建议Proposed |
12 个月路线12-month roadmap 建议Recommendation
奠基Foundation
- 价值台账与基线Value ledger and baselines
- 模型清单与风险分级Model inventory and risk tiers
- 一页权责One page of ownership
试点Pilot
- 两个客服任务上线Two contact-centre tasks live
- 对照组测量Control group measured
- 首份三层指标报告First three-layer report
上线Go live
- 分行任务上线Branch tasks live
- 关键步骤人工确认Person confirms key steps
- 上线即监控Monitoring from day one
扩展Extend
- Prism 信贷准入评审Prism gate review for credit
- 跨条线复用Reuse across lines
- 年度重排优先级Annual re-ranking
治理与合规Governance 建议Recommendation
- 模型清单对齐 OSFI E-23(2027-05-01 生效)Model inventory aligned to OSFI E-23 (effective 2027-05-01)
- 信贷与反洗钱模型:完整审查Credit and AML models: full review
- 涉及客户权益的步骤由人确认A person confirms customer-affecting steps
三层指标Three-layer metrics 建议Recommendation
| 采用Adoption | 各业务条线周活跃使用率Weekly active use by business line · 智能体发起的任务数Tasks started by agents |
|---|---|
| 过程Process | 任务完成率Task completion rate · 处理时长Handling time · 人工确认通过率Human confirmation pass rate |
| 结果Outcome | 对照 10 亿台账的价值Value against the C$1B ledger · 单任务成本Unit cost per task |
目标值由企业自定。Targets are the company's to set.
风险Risks 判断Judgment
同一份收益被重复计算Counting the same win twice
没有基线与对照组,工具席位与智能体会重复计算价值。Without baselines and a control group, tool seats and agents will double-count value.
智能体超出能力边界Agents act beyond their edge
能办事的智能体会以问答助手没有的方式出错;先在真实任务上测试。Task agents fail in ways Q&A assistants never did; test on real tasks first.
信贷里的代理指标Proxy targets in credit
模型可能优化的是代理指标而非目标;上线前写清两者。A model can optimise the proxy, not the goal; name both before launch.
前 90 天First 90 days 建议Recommendation
| 第 1–2 周Weeks 1–2 | 点名台账负责人;定基线Name ledger owners; fix baselines |
|---|---|
| 第 3–4 周Weeks 3–4 | 盘点所有在用模型与助手Inventory every live model and assistant |
| 第 5–10 周Weeks 5–10 | 风险分级与快速通道初稿;选定两个任务Risk tiers and fast lane drafted; pick two tasks |
| 第 13 周Week 13 | 首次复盘First review |
我们可能错在:We could be wrong if: 客户不接受由智能体办完的任务,或对照组显示这部分价值本来就存在。customers reject agent-completed tasks, or the control group shows the value was already there.
出处Sources
本页是基于公开资料的分析,与所述企业没有合作关系,不代表其官方立场或内部计划;企业数字多来自其自述或媒体转述;「专家组」是依据六位专家公开观点的推演,不是他们的原话,也不代表他们或其雇主的立场。资料整理日期:2026-11-14。This page is an analysis based on public information. It has no affiliation with the company and does not represent its official position or internal plans. Company figures come mostly from its own statements or media reports. The expert panel is an inference from six experts' public views, not their words, and does not represent them or their employers. Compiled 2026-11-14.
想给你的公司也做一份?Want one for your company?
同样的方法:公开事实、专属判断、12 个月路线和前 90 天。公司大小都可以聊,中文沟通。Same method: public facts, a company-specific judgment, a 12-month roadmap and the first 90 days. Any size of business; we can talk in Chinese or English.