What problem does it solve?
A bank's policy estate is large and layered: group policies, standards, procedures and desk instructions, written by different teams over many years. When a rule changes, someone has to find every document it touches, work out which clauses now conflict or say nothing, and rewrite them consistently. That work is mostly reading and cross referencing, done under deadline by specialists whose time is better spent on judgement.
The result is predictable: documents that contradict each other, procedures that lag the policy they implement, and wording that differs from team to team. When a supervisor asks how a policy implements a rule, the trail from obligation to clause is often reconstructed after the fact.
How does it work?
- Take the obligation. Start from an obligation already mapped by horizon scanning or legal: the new rule text, its effective date and the business it applies to.
- Find what it touches. Retrieval over the policy estate returns every policy, standard and procedure that covers the topic, with the relevant clauses.
- Flag gaps and conflicts. The assistant compares each clause with the obligation and marks it as compliant, conflicting, silent or unclear, quoting both texts.
- Draft the change. For each gap it drafts replacement or new wording using the approved template and style guide, as a redline, never inventing requirements beyond the source.
- Owner review. The policy owner edits and approves; legal or compliance signs off where required.
- Keep the trail. The link from obligation to clause, the drafts and the approvals are stored with the version history.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Emerging
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Compliance quality, Employee productivity, Speed and cycle time, Risk and loss reduction.
Indicative value
A bank that updates 300 policy and procedure documents a year after regulatory change
USD 38,400 to USD 288,000
Specialist time released from policy updates per year
How this is calculated
Formula: documents * hoursPerDocument * timeSaved * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Policy and procedure documents updated per year documents, documents per year | 300 | 300 | The reference bank. Replace with your own volume. |
| Specialist hours per document update (analysis and drafting) hoursPerDocument, hours per document | 8 | 16 | Editorial assumption, replace with your own time records. |
| Share of analysis and drafting time saved timeSaved, fraction of time | 0.2 | 0.4 | Editorial assumption. No public measured benchmark for policy drafting was found; keep this conservative. |
| Fully loaded cost of a policy or compliance specialist hourlyCost, USD per hour | 80 | 150 | Editorial assumption, replace with your own. |
What it leaves out: Time only. It leaves out the value of fewer inconsistencies and findings, faster implementation of new rules, and the cost of building and maintaining the policy knowledge base.
Market estimates (analyst estimates, not deployments)
- In the Bank of England and FCA 2024 survey of UK financial firms, an additional 32% of respondents expected to use AI for regulatory compliance and reporting over the next three years. Artificial intelligence in UK financial services 2024 (2024)
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Administration for Children and Families
United States · Government and public sector · 2025
In March 2025 the Administration for Children and Families, part of the US Department of Health and Human Services, deployed AI to review its existing grants, new grant applications and position descriptions for alignment with HHS Secretarial Directives related to recent executive orders. The AI produces an initial list of documents that may need revision (for grants, with an initial assessment and example passages); program office staff then review, justify and recommend. For position descriptions the agency states that AI was not used to make any final determinations. It is the gap detection half of policy change work: finding which existing documents a new requirement touches.
No outcome disclosed.
Federal Deposit Insurance Corporation
United States · Government and public sector · 2024
The FDIC reported in its 2024 AI inventory a planned assistant for its policy writers: it would check a draft policy against the Plain Language Writing Act for clarity, active voice, concision, jargon and acronyms, and redraft selected sections in plain language. In the 2025 inventory the entry is listed as retired. It is a useful signal that style and consistency checking of internal policy is an early candidate, and that not every initiative reaches production.
No outcome disclosed.
Health Resources and Services Administration
United States · Government and public sector · 2024
The Health Resources and Services Administration, part of the US Department of Health and Human Services, reported in 2024 a Policy Assistant that uses large language models to generate first drafts of key policy documents, funding notices and budget documents from example documents, style guides, key policy decisions and its internal knowledge base, plus an editing tool that checks drafts for inconsistencies and errors. The goal is less drafting time and better document quality. The entry was at the initiated stage and does not appear in the 2025 inventory; no results are published.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A current, versioned policy and procedure library with owners
- An obligation library or the new rule texts with effective dates
- Approved templates and a style guide
- Past redlines and approvals to test the assistant against
Systems to integrate
- Policy management or document management system
- Regulatory change or obligation management tool
- Workflow for review and approval
- Collaboration tools where owners edit drafts
Complexity: Medium
Retrieval and drafting are mature. The effort is in a clean, versioned policy estate with owners, a mapped obligation library and a template the drafts must follow.
- 1
Clean the estate
Retire duplicates, assign an owner and review date to every document and fix the version history. Retrieval over a messy estate finds the wrong clause confidently.
- 2
Start with gap analysis, then drafting
First use the assistant to find affected clauses and classify gaps, and measure how many it misses against expert review. Only then let it draft wording.
- 3
Constrain the drafting
Give the model the template, the style guide and the obligation text, and require every drafted sentence to cite the obligation it implements. Anything without a source is removed.
- 4
Test on past changes
Replay previous regulatory changes where the final policy text is known and compare the assistant's gaps and drafts with what the experts did.
- 5
Embed in the approval workflow
Deliver drafts as redlines into the existing review tool, record who approved what, and keep the obligation to clause link after publication.
Guardrails
- Every drafted clause cites the obligation or source text it implements
- Drafts follow the approved template; the assistant cannot publish or approve
- The policy owner approves every change; legal or compliance signs off where required
- Retrieval is limited to current, approved versions of documents
- Version history and approval trail retained for supervisors and audit
KPIs to instrument
- Time from a rule's publication to approved policy updates
- Recall of affected clauses against expert review on sampled changes
- Share of drafted text accepted without material edits
- Inconsistencies found between policy and procedure in periodic reviews
- Audit and supervisory findings on policy coverage
Human in the loop
The policy owner reviews and approves every change and owns the interpretation of the rule. Compliance or legal signs off material changes. The assistant drafts and flags; it never decides that a policy is compliant.
Common failure modes
- Invented obligations
- The model adds requirements that are not in the rule. Require a citation per sentence and strip anything unsupported.
- Missed documents
- A procedure outside the indexed estate never shows up, so the gap persists. Measure recall and keep the estate complete.
- Style over substance
- Polished drafts get approved without checking the interpretation. Owners must confirm the interpretation separately from the wording.
- Stale sources
- Retrieval returns a superseded version. Index only current approved versions and show the version in every citation.
What are the risks and rules?
EU AI Act
Minimal risk
Drafting internal policy text for human approval is not an Annex III use and has no direct effect on individuals. The Article 4 AI literacy measures still apply to the staff who use it.
Rules that apply
Guidance
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) (NIST, North America). A 2024 companion to the AI RMF that lists risks of generative AI, including confabulation, with suggested actions that apply directly to drafting assistants.
- Article 4, AI literacy (European Union, Europe). Providers and deployers must take measures to support the AI literacy of their staff and others who operate and use AI systems on their behalf.
- Artificial Intelligence (AI) Model Risk Management (Monetary Authority of Singapore, Asia Pacific). Good practices for AI and generative AI model risk management observed in a 2024 thematic review of banks, covering governance, oversight, development and deployment.
Controls to put in place
- Inventory entry for the assistant with an owner, scope and approved sources
- Citation requirement and automated check for unsupported text
- Approval workflow with recorded sign off per change
- Periodic recall testing against expert gap analysis
- Retention of the obligation to clause trail and version history
Frequently asked questions
- Can AI write our policies?
- It can find what a new rule touches and draft consistent wording quickly, but the policy owner must approve every change and own the interpretation. The safe design constrains drafting to approved templates and requires a citation for every clause.
- Where do public deployments stand?
- Mostly early. The Administration for Children and Families, part of the US Department of Health and Human Services, reported using AI since March 2025 to flag grants and position descriptions that may need revision under new directives, with staff making the final assessments. Policy drafting assistants reported by HRSA (initiated in 2024) and the FDIC (retired by 2025) never reported reaching production.
- How is this different from regulatory horizon scanning?
- Horizon scanning finds new rules and maps them to obligations. Policy drafting and gap analysis starts from a mapped obligation and changes the bank's own documents to implement it.
How to cite this page
Blits.ai AI Use Case Library, "AI for policy drafting and policy gap analysis", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/policy-drafting-and-gap-analysis. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published