Executive summary
Two marquee deployments signal how large enterprises are operationalizing generative AI beyond pilots: (1) Bank of America’s AskGPS—an in-house payments intelligence assistant for Global Payments Solutions (GPS)—and (2) Deloitte’s global roll-out of Anthropic’s Claude across ~470k professionals. Both projects emphasize governance-first architectures, domain-tuned knowledge, and clear productivity goals (turnaround time, onboarding speed, and scaled advisory). Together they offer a living blueprint for AI-enabled knowledge work in regulated industries.
1) Bank of America AskGPS: payments expertise on demand
What launched & why it matters
On September 30, 2025, Bank of America (BofA) unveiled AskGPS, a generative-AI assistant for its GPS division that answers complex client and product questions in seconds—work that previously required combing internal repositories and coordinating experts across time zones. BofA expects “tens of thousands” of employee hours saved annually and faster, higher-quality advisory to its 40,000+ business clients. Bank of America+2PR Newswire+2
Training corpus & scope
AskGPS was built in-house and trained on 3,200+ internal documents (product guides, term sheets, FAQs, presentations). Responses synthesize vetted institutional knowledge rather than merely surfacing documents. PYMNTS.com+1
Capabilities & early impact
- Speed: Queries that took up to an hour now return almost instantly. Bank of America
- Multilingual reach: 29 languages, supporting globally distributed teams and easing new-hire onboarding. ciodive.com
- Client value chain integration: Complements GPS tools such as CashPro Chat/Erica, forecasting, and intelligent receivables to accelerate advisory and operations. American Banker
Why this is notable in banking
BofA has already reported broad workforce AI adoption and ongoing investment in AI tooling; AskGPS extends that momentum into domain-specific, grounded retrieval-augmented generation (RAG) for transaction banking—one of the most documentation-heavy and compliance-sensitive workflows in finance. Bank of America+1
2) Deloitte × Anthropic: scaling AI across 470k professionals
What launched & why it matters
On October 6, 2025, Deloitte announced it will deploy Anthropic’s Claude organization-wide (~470,000 employees). Deloitte will establish a Claude Center of Excellence and train/certify 15,000 specialists to industrialize deployments, especially for regulated sectors (FS, healthcare, public). It is Anthropic’s largest known enterprise deal to date. Anthropic+2IT Pro+2
Program contours
- Center of Excellence (CoE): Implementation frameworks, best practices, support for moving from pilots to production. Anthropic
- Workforce enablement: Certification pipeline (15k+) to embed AI in client service and internal ops. IT Pro+1
- Context: Announcement coincided with scrutiny after an AI-tainted report prompted a refund to an Australian government client, underscoring the need for model governance and citation discipline. TechCrunch
Ecosystem signal
The move complements Anthropic’s broader enterprise push (e.g., alliances, enterprise IDE integrations with IBM). It positions Deloitte to package domain personas, compliance tooling, and agentic workflows atop Claude for large clients. The Wall Street Journal
3) Design patterns both deployments share
- Governance-first architecture
- Grounded RAG: connect models only to approved, versioned corpora (policy docs, term sheets, playbooks). AskGPS’s curation (3,200+ vetted sources) exemplifies narrow, high-quality grounding. PYMNTS.com
- Controls: access gating, audit logs, and answer traceability (document links, provenance). Deloitte’s CoE indicates systematic controls to reduce hallucinations and ensure regulatory compliance. Anthropic
- Operational KPIs over vanity metrics
- Time-to-answer: shift from “find the doc” to “deliver the answer,” measured in seconds, not minutes. (BofA.) Bank of America
- Onboarding speed: language coverage and playbook-style answers shorten ramp time. (BofA 29 languages.) ciodive.com
- Advisory quality: structured, strategy-oriented replies rather than raw excerpts; Deloitte’s persona approach points to role-specific quality bars. IT Pro
- Scale via enablement
- Centers of Excellence & certification: codify patterns (prompting, evaluation, red-teaming, data contracts) so each team doesn’t reinvent the wheel. (Deloitte.) Anthropic
4) Reference architecture (high level)
Data plane
- Curated knowledge sets (policies, terms, product guides) in a content store with document lineage + retention; embeddings index; PI/PHI redaction where applicable. (Pattern illustrated by AskGPS’s curated corpus.) PYMNTS.com
Control plane
- Access control (role, geo, client restrictions), policy enforcement, prompt/response logging, and evaluation harness (automated tests for accuracy, safety, PII leakage). (Mandated by Deloitte’s governance narrative and regulated-industry focus.) IT Pro
Model/agent layer
- Enterprise LLM (e.g., Claude) with system prompts encoding tone, disclosure, refusal rules; tools (calculators, policy lookups, FX calendars); RAG with citations and fallbacks (answer abstention when confidence low). (Consistent with Deloitte’s plan to build industry-specific solutions/personas.) IT Pro
Experience layer
- Chat-first UX embedded in existing portals (e.g., CashPro, internal hubs), multilingual support, structured outputs (term-sheet comparisons, fee tables), and feedback loops (thumbs up/down + explain). (Reflected in AskGPS goals to deliver answers, not documents.) American Banker
5) Risk register & mitigations
| Risk | Example | Mitigation |
|---|---|---|
| Hallucinations & incorrect advice | Deloitte refund incident shows reputational and financial exposure. | Grounding + must-cite policies; answer abstention on low confidence; human verification for high-impact outputs; evaluation gates before new corpora go live. TechCrunch |
| Out-of-date content | Policies/fees change frequently in payments. | Freshness SLAs for corpora; content owners sign off; automatic re-indexing; response banners with “effective as of” dates. |
| Privacy/PII leakage | Client-specific details surfacing across accounts. | Tenant isolation; PII redaction; scoped retrieval by client/region; differential logging; red-team tests. |
| Over-reliance by staff | “Automation bias” in advisory. | Training on verification; UX nudges (confidence, sources); periodic skills re-certification. |
| Model drift / vendor lock-in | Changing model behavior or pricing. | Abstraction layer for models; golden prompts; continuous eval; data portability. |
6) Measuring value (suggested KPI set)
- Time-to-answer (TTA) distribution (p50/p90) for top use cases. (AskGPS targets seconds.) Bank of America
- Deflection rate: % of queries resolved without SME escalation.
- Onboarding time: days to proficiency by role; language-coverage utilization. (AskGPS: 29 languages.) ciodive.com
- Answer quality: rubric-based scoring (accuracy, completeness, citations), periodic blind reviews.
- Safety incidents: hallucination reports, data-handling violations, and time-to-remediate.
- Business impact: win-rate lift on RFPs; cross-sell/upsell insights surfaced; reduction in SLA breaches.
7) Implementation playbook (90 days)
Days 0–30: foundations
- Define top 10 intents (e.g., treasury onboarding, fee schedules, cut-off times).
- Curate corpus with owners, effective dates, retention schedules; strip stale docs.
- Stand up eval harness: gold Q&A sets; hallucination and citation tests; PII probes.
- Security reviews: data residency, role-based access, logging scope.
Days 31–60: pilots
- RAG pipeline with chunking tuned to domain (tables, rate cards, term sheets).
- Agent tools: currency calendars, calculators, product comparators.
- Human-in-the-loop for high-impact tasks; measure TTA, deflection, quality.
Days 61–90: scale
- CoE + certification (prompting standards, incident runbooks, change control). (Deloitte pattern.) Anthropic
- Rollout plan by region/language; publish adoption dashboards; create “knowledge-owner of record” workflow for updates.
8) Competitive and ecosystem context
- Banking peers are stepping up AI for transaction services; BofA’s move arrives alongside other big-bank and vendor announcements, signalling a race to embed AI into treasury and payments advisory. American Banker
- Consultancies & SIs are formalizing AI deployment factories (personas, playbooks, controls). Deloitte’s certification and CoE strategy is likely to become a de facto norm. IT Pro
- Toolchain convergence: Anthropic’s enterprise push includes alliances (e.g., IBM IDE integration, agent frameworks), hinting at a near-term standard stack for secure enterprise agents. The Wall Street Journal
9) What to watch next (Q4 2025–H1 2026)
- Deeper integration of AskGPS with client-facing portals (e.g., templated proposals, auto-cited advisory notes). American Banker
- Persona libraries from Deloitte (auditor, risk analyst, solution architect) with industry-specific guardrails and evidence packages. IT Pro
- Agentic workflows spanning multiple systems (KYC/AML checks, counterparty onboarding) with audit-ready traces.
- Third-party assurance: independent audits of AI controls becoming a sales differentiator.
Appendix A: Snapshot facts & sources
- AskGPS launch date: Sept 30, 2025 (Bank of America newsroom; PRNewswire). Bank of America+1
- Corpus size: 3,200+ internal docs (product guides, term sheets, FAQs). PYMNTS.com
- Client reach: 40k+ business clients via GPS. Stock Titan
- Time savings: “tens of thousands” of hours; hour-long queries now seconds. Bank of America
- Language support: 29 languages. ciodive.com
- Deloitte × Anthropic: global rollout to ~470k, CoE and 15k certifications, regulated-industry focus. IT Pro+2Anthropic+2
- Governance lesson: refund incident underscores stakes of oversight. TechCrunch
- Anthropic enterprise momentum (context): alliances/integrations with IBM and others. The Wall Street Journal
