What Is Human-in-the-Loop for AI-Generated Finance Content? What the EU AI Act and the UK Framework Mean for Content Teams
21/11/2026 by Gustav Christopher Wagner, CFA
AI can be used to draft finance content. Nothing in the EU or UK rules prevents it. The deliverable is a claim that a named person has checked and is willing to sign. This guide is for anyone who commissions, approves or produces financial content with AI tools and needs to know what the EU and UK rules require.
AI rules and finance content at a glance
Author's summary. Exact scope should be checked with the relevant compliance team.
| Rule | Who it targets | Does it require a human to review content? | Status (21 Nov 2026) |
|---|---|---|---|
| AI Act, Art. 14 (human oversight) | Providers and deployers of high-risk AI systems | Not for content tools. It covers listed uses such as credit scoring of individuals and life and health insurance pricing | Annex III rules apply from 2 December 2027 |
| AI Act, Art. 50(2) (marking) | Providers of generative AI | No. Providers must mark outputs in a machine-readable format | Applies from 2 August 2026. Systems already on the market have until 2 December 2026 |
| AI Act, Art. 50(4) (labelling text) | Deployers publishing AI text that informs the public on public-interest matters | Human review is the exemption from labelling, according to the Commission's FAQ | Applies from 2 August 2026 |
| AI Act, Art. 4 (AI literacy) | Providers and deployers | No. They must support staff AI literacy, with no guaranteed level | In force |
| UK FCA framework | Regulated firms | Not by an AI-specific rule. Existing rules apply, and the FCA expects risk-based human oversight | In force |
What is human-in-the-loop?
The FCA's glossary defines it as a control where a person reviews, approves, challenges or can override an AI output or decision before it is used. For content, that means a person with the right expertise reads the output, checks the facts and decides whether it goes out. It does not mean someone skimmed it for typos.
The point most guides get wrong: the AI Act does not require this for a blog
Human-in-the-loop for finance content is not an AI Act requirement. It comes from the financial rules firms already follow, and the AI Act mostly confirms the logic.
Article 14 is about high-risk systems. It requires that such systems can be effectively overseen by people, including the ability to override or reverse the output. The finance examples in Annex III are systems that evaluate the creditworthiness of natural persons and systems that price life and health insurance. A tool that drafts a blog post is not on that list. Even for those systems, the oversight rules now apply from 2 December 2027, after the AI Omnibus moved the date.
So where does the human requirement for content come from? From the sector rules. In the UK, the FCA says firms remain responsible for AI systems they develop, procure and deploy, including their outputs and outcomes. In the EU, ESMA said in 2024 that firms' decisions remain the responsibility of management bodies, whether people or AI tools take them. ESMA also said this covers staff using third-party tools such as ChatGPT, with or without senior management's knowledge.
That is a stronger argument than "the AI Act says so". The duty to review sits with the firm, whatever the tool.
The marking rules: what Article 50 does and does not cover
This is where the AI Act touches content. It does two separate things.
Article 50(2): the tool vendor marks, not the content team
Providers of generative AI systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated. That is the provider's job. A content team using the tool does not add that marking itself. The Commission also envisages a narrow exemption for outputs in "business to business" or "industrial contexts", where the guideline conditions are met. Outputs from assistive standard editing are outside the marking obligation too.
Article 50(4): the one that could reach published articles
Deployers must label AI-generated text only if it is published, informative to the public, and on matters of public interest. The Commission's list of public-interest topics includes any economic or financial development that may be relevant to public debate.
The Commission's guidelines give examples. Inside the scope:
Outside it:
- AI-manipulated text that is part of a company's advertisement or product descriptions, as long as it makes no claims on health, consumer safety or sustainability.
- Text that is not "published": private professional correspondence and organisation-internal texts.
One law firm's reading for asset managers is that ordinary fund marketing, fact sheets and investor communications produced with AI assistance will generally not trigger the disclosure duty, while public-interest disclosures, such as sustainability-related ones, may.
"B2B" is not the test
The common shorthand needs tightening. The Commission's FAQ does not categorically exclude marketing, advertising or B2B content. The test is what the text does and who can read it, not who the customer is. So:
- A product page or sales email: almost certainly outside Article 50(4).
- A published market-commentary or thought-leadership article that informs a general audience on a financial development: potentially inside it. This is an inference from the Commission's examples, not a ruling.
Human review is the way out, and it has conditions
For text inside the scope, the labelling duty does not apply if the content has undergone human review or editorial control, and a person holds editorial responsibility. The guidelines say what counts:
- It must be substantive. The guidelines require deliberate examination of the substance by people with relevant knowledge and professional judgement, with fact-checking as the minimum.
- Spell-checks do not count. Formal checks, the mere existence of an editorial policy, automated review or cursory approval do not qualify.
- Editing after sign-off voids it. If AI modifies the content after sign-off, it is treated as AI-generated again.
- Someone must own it. A person or entity must hold ultimate legal responsibility, and their identity and contact details should be easy to find, for example on a website's terms or legal pages.
- AI reviewing AI does not count. The guidelines' own failing example is AI-generated articles reviewed and edited by another AI system, with a human doing only a superficial grammar check.
Their passing example is useful for finance: AI-generated sustainability reports on a listed company's website that have undergone human review by professionals in relevant functions, such as compliance.
The rule shapes the workflow. If a qualified person substantively reviews the text and takes responsibility, no "AI-generated" label is needed on in-scope text. If nobody does, a label is.
What it means in the UK
There is no UK AI Act. The FCA says it does not plan to introduce extra regulations for AI and will rely on existing frameworks, including the Consumer Duty and the Senior Managers and Certification Regime (SM&CR). Its 2026 annual public meeting responses repeated that existing frameworks, including SM&CR accountabilities, continue to apply to firms' use of AI, and that the FCA is comfortable they are sufficient.
For content, three existing layers matter.
1. Communication rules. COBS 4.2 says that a firm must ensure that a communication or a financial promotion is fair, clear and not misleading. It applies proportionately, and a communication to a professional client may not need the same information or presentation as one to a retail client. Proportionate is not exempt. Some B2B promotions fall under specific carve-outs, such as non-retail communications, so the exact position of each piece should be confirmed with compliance.
2. Accountability. The FCA does not expect senior managers to be technical specialists, but they should take reasonable steps to understand AI risks in their area, get expertise where needed, and provide effective oversight and challenge. It flagged poor practice where senior managers delegated AI responsibility to technical teams without sufficient challenge.
3. Consumer outcomes. Firms using AI tools remain responsible for outcomes delivered through products, services, communications and support.
What the FCA says good human oversight looks like
The FCA's October 2026 good-practice publication is the most useful document for content teams. Three points stand out:
- Review is a risk-based control, with a qualified person. Good practice includes risk assessments of where human oversight is needed, and an appropriately qualified person with the expertise to spot key risks and the authority to escalate.
- Reading is not enough. Effective oversight is about giving people authority, information, capability and escalation routes to challenge or intervene, not just reviewing outputs.
- A human in the loop can fail. The FCA lists as an area for improvement excessive reliance on human-in-the-loop controls, which may mask model weaknesses, and human review that becomes box-ticking.
The publication also gives concrete content examples. One firm used an LLM to refine letters answering consumer queries and complaints, with staff reviewing and sending them, and tested the new style with consumers for understanding first. Another used AI to propose clearer wording, and staff reviewed all changes before approving the final wording. A small firm also recorded whether staff reviewed AI outputs, and whether they approved, overrode or escalated, which created an audit trail.
The regulator is also checking approval quality. A May 2026 FCA review of ten promotion approvers found some had approved adverts with unsubstantiated claims, and some relied on third-party templates instead of doing proper checks. An AI draft is, in effect, a third-party template.
EU and UK side by side
| EU | UK | |
|---|---|---|
| AI-specific law | AI Act, with Article 50 live since 2 August 2026 | None. The FCA relies on existing rules |
| Must AI text be labelled? | Only in-scope public-interest text without substantive human review | No AI labelling rule found in the FCA material reviewed |
| Human review of financial content | Not an AI Act duty for content tools. It comes from MiFID-type duties and ESMA expectations | Expected as a risk-based control under SM&CR, Consumer Duty and COBS 4 |
| Who is accountable | The management body | Senior managers, and the firm |
| Third-party tools | Staff use of ChatGPT-type tools is in scope | Firms can outsource AI services but remain responsible |
The accountable human: what a workable process contains
Author's checklist. It follows the sources above, but it is not a regulatory template.
- A named approver. One named person signs each piece. For published pieces in EU scope, the responsible entity should be easy to find.
- A reviewer who can judge the substance. A finance subject matter expert, not a proofreader. The AI Act guidelines and the FCA both stress competence and authority.
- A substantiation file. Every number and claim gets a source, created at review time.
- A freeze after sign-off. Any AI rewrite afterwards triggers a new review, because modified content is treated as AI-generated again.
- A review log. It records who reviewed the piece, what was changed and what was rejected. The FCA's good-practice examples include audit trails of reviewer decisions.
- Permission to say no. Authority to reject is what separates oversight from rubber-stamping, as the FCA's oversight guidance makes clear.
- Reader comprehension testing. The FCA lists assuming AI output is clear, without testing, as poor practice.
- Vendor and tool control. The firm knows which tools staff and freelancers use. ESMA expects firms to control third-party AI use by employees.
Which route fits which content?
Author's judgement. Compliance has the final word.
| # | Content | Article 50 risk | Human review needed |
|---|---|---|---|
| 1 | Product pages, brochures, sales emails | Low. Ads and product text are outside 50(4) | Yes, as a regulated promotion or client communication |
| 2 | Fund factsheets, fund marketing | Generally low | Yes, SME plus compliance |
| 3 | Published market commentary, thought leadership | Possible if aimed at a general audience on financial developments | Yes. It is also the exemption |
| 4 | Corporate reports with investor information | Yes, listed in the guidelines | Yes, with a named responsible entity |
| 5 | Private client letters, internal notes | Low. Not "published" | Yes, under FCA or ESMA rules |
| 6 | Content that makes health, safety or sustainability claims | Higher | Yes, and labelling rules need checking |
Using AI for finance content and needing a human who can sign it off? G.C. Wagner helps investment firms, FinTechs and financial services businesses turn complex ideas into clear, accurate content. With 15+ years in financial services and CFA charterholder credentials, the writing is done the way a specialist reads.
Contact Me | Connect on LinkedIn
What would change this analysis
- A regulator or court reads "public interest" more narrowly. The Commission's guidelines are guidance, not a binding interpretation. If they are read more narrowly, fewer finance articles fall under Article 50(4).
- The FCA introduces AI-specific content rules. Today it says it will not. However, the Mills Review looked at how AI could reshape retail financial services, and the FCA's position could change.
- The firm is not regulated. An unregulated B2B firm faces fewer of these rules. What matters is whether it communicates or approves financial promotions.
Takeaway
The AI Act does not tell anyone to check AI drafts. The regulator does, in effect, because the firm is accountable for what it publishes. The AI Act adds one useful detail: where a label would otherwise apply, a substantive human review with named responsibility can remove it.
The first question is therefore not "does this need a label?" but "who has checked it, and who will sign it?" The same argument runs through the guide to German financial content writing: a text can read perfectly and still be wrong.
Frequently asked questions
Does the EU AI Act require human review of AI-generated marketing?
No. Article 14 covers high-risk systems such as credit scoring of individuals. For content, the human duty comes from financial regulation and firm accountability.
Must AI-generated finance content be labelled?
Only if it is published, informs the public on a public-interest matter, and has not had substantive human review with a person holding editorial responsibility, according to the Commission's FAQ. Advertising and product text is generally outside this.
Is B2B content exempt from Article 50?
Not as a category. The Commission's FAQ does not exclude B2B content categorically. The 50(2) exemption is narrow and targets providers. The 50(4) test is about published, public-interest text.
Does the UK have an AI content labelling rule?
None was found in the FCA material reviewed. The FCA says it will not introduce extra AI regulation and relies on existing rules.
Is a quick proofread enough?
No. The Commission says formal checks and cursory approval do not count. The FCA warns against box-ticking oversight.
Who can check AI-assisted finance content?
A finance subject matter expert or a specialist financial writer who can judge the substance.