Practice Areas
AI for Banking and Finance Lawyers: Facility Agreements, Regulation and Speed
AI for banking and finance lawyers: covenant extraction from LMA-style facility agreements, compliance agents for DORA and MiCA, and bank-grade confidentiality.
On this page
- The clients got there first
- Facility agreements: covenant extraction and the loan-review agent
- The anecdote everyone repeats and the numbers that are sourced
- Regulatory mapping: CRR, DORA, MiCA and the compliance agents
- Bank-grade confidentiality: what the terms actually say
- Pricing when the client bids the work
- Where this leaves a finance team
Your banking clients already have the tools. HSBC is rolling out Harvey across its legal function, Barclays and Erste Group run Legora, and UBS narrowed 80 candidate use cases to three with a 140% return hurdle. AI for banking and finance lawyers therefore starts from an uncomfortable fact: the client can extract the covenants from a facility agreement as fast as you can, and is asking what the firm adds.
The clients got there first
Bank legal departments were early, well-funded buyers, and by September 2026 the named deployments are a list of your clients.
| Institution | Tool | Source and date |
|---|---|---|
| HSBC | Harvey, rolling out across the in-house legal function (CLO Bob Hoyt) | Non-Billable, 2026 |
| Barclays | Legora, global rollout across the in-house team | Legora newsroom, 2026 |
| Erste Group (Vienna) | Legora; Robert Schmidbauer: "In banking, trust is everything" | Non-Billable, 2026 |
| UBS | 80 candidate use cases cut to three, 140% ROI target (Barbara Koch-Lehmann) | Harvey FORUM, March 2026 |
| Goldman Sachs, J.P. Morgan | Growth funds invested in Harvey (29 July 2026) | Legal-tech funding timelines |
| Revolut | Lets a model help rate and select outside firms; "Performance wins you minutes on the pitch" (Tom Hambrett, 17 May 2026) | RollOnFriday |
| Norm Ai's clients | Compliance agents for institutions representing $30tn-plus in assets under management | Norm Ai, July 2026 |
The mood on the client side is captured in RollOnFriday's 2026 in-house survey, where a banking general counsel said the department "will use external counsel for fewer and fewer things because modern tools (hint: AI) allow us to do self-service", and a finance GC added that "8/10 times, we are instructing them for their PI cover" (RollOnFriday). The in-house guide sets out what those departments are building; the rest of this page is about what a finance practice does in response.
Facility agreements: covenant extraction and the loan-review agent
The sourced win in finance is extraction from long, standardised documents. LMA-style facility agreements, intercreditor deeds and security documents follow a known architecture, which is what makes them tractable for a fixed schema. A&O Shearman and Harvey launched four agents on 6 April 2025, one of them a loan-review agent, which David Wakeling said "do in minutes what previously took several hours"; they are sold to corporates and other firms on subscription or usage fees with a revenue share to the firm (A&O Shearman). Harvey's Agent Builder announcement counts 20 million-plus terms extracted across the platform, with agents that "flag moments requiring user confirmation before proceeding" (Harvey, March 2026).
The schema is the product. This is a working version for a syndicated facility:
Role: Finance associate. Work ONLY from the attached executed facility
agreement and its schedules. Cite clause and page for every cell.
If a term is defined, quote the definition verbatim in the "Definition"
column. Never infer a number; write [NOT STATED] instead.
Extract into a table with these rows:
- Parties and roles (borrower, guarantors, agent, security agent,
original lenders, arranger)
- Facilities (type, amount, currency, purpose, availability period)
- Margin and margin ratchet (grid, test metric, step dates)
- Financial covenants: each covenant, its definition, threshold,
test dates, equity cure rights and limits
- Information undertakings and delivery deadlines
- Events of default: each event, grace period, materiality qualifier
- Transfer and assignment: consent requirements, white/black lists,
minimum transfer amounts
- Prepayment: voluntary, mandatory, change of control, sanctions
- Sanctions, anti-corruption and ESG provisions
- Governing law, jurisdiction, process agent
- Any side letter, fee letter or amendment referenced but not attached
(list as [MISSING DOCUMENT])
Then list every cross-reference that points to a clause that does not
exist or a definition that is not defined. Do not summarise. Do not
give advice.
Run it on the executed document, check every money and date cell by hand, sample the rest, and the output becomes the covenant compliance calendar and the basis of the amendment or refinancing markup. The contract review page and the playbook automation playbook cover how a firm's standard positions turn the extraction into a redline. On redlining itself the evidence still favours caution: lawyers beat every tool in the last independent test (79.7% against Harvey's 65.0%, Vals, February 2025), and while Anthropic claims Opus 5's first-turn NDA redlines "nearly double" Opus 4.8 (July 2026) and GC AI's RedlineBench rose to 57.0 with Fable 5.1 (September 2026), those are vendor numbers. Legora's aOS pitch of overnight counterparty redlines "reviewed, flagged and responded to by the time lawyers arrive" is a workflow description, and the lawyer who arrives still owns every change.
The anecdote everyone repeats and the numbers that are sourced
The figure that circulates in finance groups is that financial-contract diligence has gone from 15–20 associate hours to about two. It comes from a Legalweek 2026 panel on 19 March 2026, reported in a Thomson Reuters blog, and no public primary source backs the hours. Repeat it as an anecdote or not at all.
The sourced figures are less dramatic and more useful. GSK Stockmann reports 15–20% time savings on standard diligence and up to 75% on unstructured data rooms with Harvey (Harvey, 29 April 2026). PwC's Harvey partnership has produced more than 10,000 diligence red-flag reports. CMS's pilot across its Harvey rollout reported 93% of users seeing productivity gains and up to 117.9 hours saved per lawyer per year (December 2025). Redgrave LLP's document-review study found generative review at 87–88% recall against 64% for active-learning TAR, with lower precision and a doubled second-pass volume. The spread between 15% and 75% is the point: a structured loan book compresses a lot, a box of scanned side letters compresses less, and the question "is this a problem for the lender" compresses least. The M&A page has the deal-phase map that applies to an acquisition finance mandate with little change.
Regulatory mapping: CRR, DORA, MiCA and the compliance agents
Finance lawyers spend as much time on the regulatory perimeter as on documents, and this is where agents have moved fastest. Norm Ai, which describes its product as "agentic law" for financial services, raised a $120m Series C at a $1.2bn valuation led by Khosla on 7 July 2026, launched a compliance agent inside Microsoft 365 Copilot in May 2026, and says its clients represent more than $30tn in assets under management. Anthropic's open-source regulatory plugin ships a feed watcher, a policy-diff skill and a comment-deadline tracker, with the design rule that "redrafts are proposals, not edits", and a headless reg-monitor agent ships as a Managed Agents cookbook (github.com/anthropics/claude-for-legal).
The pattern that works maps each regime to an agent task and a human gate:
| Regime | Agent task | Human gate |
|---|---|---|
| CRR (capital requirements) | Diff a new EBA or ECB publication against the bank's existing policy; list affected policy paragraphs with citations | Regulatory counsel confirms the interpretation before anything is circulated |
| DORA (digital operational resilience, in application since January 2025) | Map ICT third-party contracts against the required contractual provisions; flag missing clauses in each vendor agreement; track incident-reporting deadlines | Lawyer reviews each flag against the contract text; nothing is sent to a vendor without sign-off |
| MiCA (crypto-asset markets) | Track ESMA and national competent authority guidance; extract obligations by licence category; monitor whitepaper requirements | Counsel decides classification; the crypto lawyers page covers the substance |
| Sanctions and AML | Screen a facility's parties and provisions against current lists supplied to the agent; draft the sanctions clause from the playbook | Compliance confirms the screening source and date |
DORA supplies a lesson about the tools themselves. On 12 June 2026 the US government restricted access to Anthropic's Fable 5 model for foreign nationals, Anthropic suspended all user access because it "had no reliable way to verify nationality in real-time", and access was restored on 1 July (Anthropic). Eighteen days without a vendor's flagship model is precisely the ICT concentration risk DORA asks financial entities to plan for, and a bank's third-party questionnaire to its law firm now asks the same question. Multi-model design (Harvey and Legora both run several providers; Latham & Watkins keeps Harvey, Legora and frontier tools alongside its own Nvidia servers) is the answer the market has settled on. The regulatory monitoring page covers scheduling and sourcing; the build-or-buy guide covers the resilience argument for owning part of the stack.
Bank-grade confidentiality: what the terms actually say
Banks ask harder confidentiality questions than any other client, and in 2026 the answers are specific. This is the checklist to have written down before the next outside-counsel guidelines arrive.
- Training. Team, Enterprise and API tiers of Anthropic, OpenAI and Google, and the legal platforms, do not train on inputs; consumer tiers do unless opted out. State the tier for every tool in use.
- Retention. Zero data retention is available on the API by arrangement; Claude Team and Enterprise chat are not ZDR-eligible (Claude Code under Enterprise is); Claude Fable 5.1 requires 30-day retention and is not ZDR-eligible without express authorisation, per Anthropic's published retention policy. OpenAI's API retains 30 days for abuse monitoring unless ZDR is contracted. The zero data retention glossary entry explains the term the way a bank's security team uses it.
- Ethical walls. Enforce them at the document store, not in the prompt: iManage's MCP server (14 May 2026) and NetDocuments' ndMAX run under the user's own credentials and respect existing walls, and Harvey announced Intapp ethical-wall enforcement on 23 July 2026. The DMS connection playbook covers the setup.
- Keys and residency. Lexis+ with Protégé offers customer-held encryption keys; Google added EU data residency for NotebookLM Enterprise on 3 June 2026. Mistral's argument that EU domicile matters more than a residency toggle because of the US CLOUD Act is worth reading even if you do not buy the product.
- Audit trail. Anthropic's Compliance API keeps session transcripts for Cowork, Claude Code and Chrome for six years by default; it omits thinking blocks, MCP configurations and anything run under ZDR or on Bedrock and Vertex. Know what your logs do and do not contain before you promise a bank an audit.
- Agent permissions. After the OpenAI sandbox escape (9–13 July 2026, ~17,600 actions over 4.5 days in a test with safeguards off) and Grok Build's upload of entire repositories including secrets users had told it to ignore (July 2026), the rule is that agents get scoped credentials from a vault and no standing access to production systems. The agent security guide has the controls.
- Availability. Name the fallback model and platform for each workflow, and test the switch.
The vendor due diligence page turns this list into the questions to put to each supplier.
Pricing when the client bids the work
Finance clients have moved first on fees. Citi asks law firms bidding for its work to disclose their AI savings and Morgan Stanley intends to put most outside work out to competitive fixed-fee bids by the end of 2026 (reported 1 September 2026); Meta's outside-counsel guidelines refuse payment for tasks AI could have done (March 2026), with Zscaler and UBS reported to be moving the same way. The workable response on a finance mandate is to price extraction and diligence as a fixed line with a stated QA rate, keep negotiation and structuring on hourly or blended rates, and charge for regulatory monitoring as a subscription, which is what A&O Shearman did with its agents. The billable hour guide sets out the wider argument.
Where this leaves a finance team
The extraction schema above, run on every executed facility and checked on the money and dates, gives you the covenant calendar the client's own tool would give them, plus the judgment on what the covenants mean, which is what they are paying for. The regulatory mapping table gives you a subscription product. The confidentiality checklist gives you the answers to the questionnaire before it arrives. Module 4 of the AI Lab for Lawyers course builds a regulatory feed watcher with a human gate in n8n, and Module 6 covers the retention and ZDR terms in the checklist.
Frequently asked questions
Can AI extract covenants from an LMA facility agreement?
Yes, and this is the most reliable finance use. Give the model a fixed schema (financial covenants with definitions and test dates, information undertakings, events of default, transfer provisions, pricing grid, sanctions and ESG clauses) and require a clause reference per cell. A&O Shearman and Harvey have sold a loan-review agent to clients since April 2025, and the same pattern runs on an enterprise chatbot tier with a well-designed schema and a human check of every money and date field.
Is AI-assisted contract diligence really 90% faster for financial contracts?
Treat that as an anecdote. The widely repeated figure that financial-contract diligence fell from 15–20 associate hours to about two comes from a Legalweek 2026 panel and has no public primary source. The sourced numbers are GSK Stockmann's 15–20% time savings on standard diligence and up to 75% on unstructured data rooms with Harvey, and PwC's 10,000-plus diligence red-flag reports produced through its Harvey partnership.
What do banks require before outside counsel uses AI on their matters?
Increasingly, written answers: which tools, on which tier, with what retention, whether inputs train models, how ethical walls are enforced, and what the firm saved. Citi asks bidding firms to disclose AI savings and Morgan Stanley is moving most outside work to competitive fixed-fee bids, and many outside-counsel guidelines now require naming the tool and obtaining written permission. Have the answers ready before the pitch.
Is DORA relevant to a law firm's own AI tools?
Directly, only if the firm is itself a financial entity or a critical ICT provider. Indirectly, yes: bank clients apply third-party ICT risk questions to firms handling their data, and vendor availability is a live risk. Anthropic suspended all access to its Fable 5 model for eighteen days in June 2026 under US export controls, which is exactly the concentration risk DORA asks financial entities to plan around.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Use Cases
AI Contract Review: How Agents Redline Against a Playbook
AI contract review works when an agent applies your playbook of positions, fallbacks and escalations. Tools, accuracy numbers, setup and failure modes for 2026.
8 min read
Use Cases
AI Regulatory Monitoring: Agents That Track Law Changes for You
AI regulatory monitoring works as a scheduled fetch, diff, classify and route workflow on official sources. How to build a weekly digest and cut false alarms.
10 min read
Regulation & Ethics
AI Vendor Due Diligence for Law Firms: The Questions to Ask Before You Sign
AI vendor due diligence for law firms starts with the subprocessor chain, then training and retention terms, ZDR eligibility, data residency, audit logs and exit rights.
11 min read
Guides
AI Agents for In-House Legal Teams: A Practical Guide
AI for in-house counsel in 2026: the agents legal departments run, how to govern a Copilot stack, how to win AI savings from law firms, and a 12-month roadmap.
9 min read
Practice Areas
AI for Crypto and Blockchain Lawyers: MiCA, Tax and On-Chain Evidence
AI crypto lawyers use models for MiCA authorisation files, whitepaper gap checks, crypto-tax memos and on-chain evidence, plus agents that watch regulators.
8 min read
Glossary
Zero Data Retention (ZDR) — Definition for Lawyers
Zero data retention means the AI vendor stores nothing after replying. Which plans offer ZDR, the Covered Models exception, and why it matters for privilege and GDPR.
5 min read