Proofreading, style & brand-voice rewriting

Every organisation writes thousands of texts a week: emails, web pages, manuals, offers, support answers. Many people write them, and many of those texts are now LLM drafts. Spelling mistakes, inconsistent terms and an off-brand tone cost credibility, and human copy-editing does not scale. AI writing assistants combine rule-based checkers (spelling, grammar, terminology, house style) with LLM rewriting (clarity, tone, length, brand voice). They sit in the tools people already use and give suggestions that the author accepts. Basic proofreading is now built into office suites and operating systems. What sets products apart in 2026 is how well they enforce a company's own style guide, terminology and voice, with explainable and repeatable checks.

Problem & value

• Today: spell-checkers in Word/Outlook catch typos. Style guides live in a PDF nobody reads, terminology lists sit in Excel, and brand and legal review happens late and by hand. With generative AI in use, much more text gets written, much of it in a generic "AI voice".
• AI value: consistent quality across all authors and languages, fewer review cycles, enforced terminology (important for technical documentation and translation cost), and a consistent brand voice in customer-facing text. A non-native writer can produce business-ready text.
• Evidence: Writer reports a 70 % time saving on derivative content at KPMG and 57 % faster time-to-market at Vanguard (vendor-reported case figures covering broader content workflows, not only proofreading) [5]. Independent, measured benchmarks of brand-voice compliance do not exist. Grammatical error correction (GEC) itself is a well-studied academic task [9].

Variants

• Private users / students: free tiers of LanguageTool, Duden Mentor, DeepL Write, Grammarly. Writing tools are also built into the OS (Apple Writing Tools, Copilot in Windows/Word, Gemini in Google Docs). For use with exams and graded essays see EDU-003. For CVs and cover letters see PERS-012.
• SMB: team plans with a shared dictionary, simple style rules and tone profiles (LanguageTool Team, DeepL Write Pro, Grammarly Pro).
• Enterprise marketing & communications: brand-voice profiles, governed terminology, quality scores and quality gates in CMS/DAM, plus a check of LLM output before publishing (Markup AI, formerly Acrolinx; Writer). This links to SEO and marketing content generation (LANG-001, LANG-005).
• Technical documentation: controlled language (ASD-STE100 Simplified Technical English), translation-oriented writing and terminology management (Congree, Markup AI). This is the check layer for LANG-022 and cuts translation cost in LANG-009.
• Docs-as-code / developers: prose linting in CI with Vale (YAML rules, GitHub Action) [7], optionally with an LLM step for rewrite suggestions.
• Public sector: plain-language and gender-inclusive writing rules for official letters and websites. Full Leichte Sprache conversion is LANG-010.
• Boundary: checking regulated claims (HWG, financial promotions, green claims) is LANG-024. Translation is LANG-009. Fact-checking is LANG-021. Generating new text from scratch is LANG-001/002/003.

How it works

1. text from editor / add-in / CMS / API / CI
2. segmentation + language detection
3. deterministic layer: spelling, grammar rules, terminology DB, style-guide rules (regex/NLP)
4. ML layer: GEC model (seq2seq or encoder tagger) for errors rules miss
5. LLM layer: clarity, tone, length, brand-voice rewrite (style guide + example texts in prompt or retrieved)
6. diff + explanation per suggestion, score per rule category
7. author accepts / rejects (human-in-the-loop)
8. optional quality gate (CMS publish, CI build)
9. feedback (accepted/ignored rates)
10. tune rules and prompts

• Hybrid is standard. Rules and terminology are deterministic and explainable, so the same text gets the same result. Markup AI advertises deterministic scores for quality gates [2]. LLMs handle open-ended rewriting, where explanations and repeatability are weaker.
• Brand voice is passed as a structured profile (do/don't lists, tone attributes, approved examples). For large example sets, the most similar approved texts are retrieved with text embeddings and given to the model as few-shot examples.
• Minimal vs. fluency edits: proofreading should change only what is wrong. Rewriting deliberately changes wording. The two modes should be kept apart, as DeepL's API does with separate /write/correct and /write/rephrase endpoints [1].
• Human-in-the-loop: the author keeps editorial control. Automatic rewriting without review only suits low-stakes text, such as normalising internal notes.

Data preparation

The basis is rarely data in the classic sense but company writing knowledge: a style guide as PDF or Word, product names in the ERP or PIM, approved texts in the CMS, and brand guidelines held by marketing. Much of it is implicit, living in the heads of two senior editors.

Typical gaps are outdated or contradicting style guides, no list of forbidden and preferred terms, scattered example texts, and no defined brand voice. Drafts often contain customer data or unpublished figures, so the tool choice depends on what leaves the company.

Preparation steps: 1. Condense the style guide into 1 to 2 pages of checkable rules (spelling variants, tone, number formats, inclusive language) (PREP-014). 2. Build a terminology list with approved, deprecated and forbidden terms (PREP-015). 3. Collect 10 to 20 gold-standard texts per channel as brand-voice examples (PREP-016). 4. Define which text classes may go into which tool (PREP-024). 5. Capture editors' accept and reject decisions to refine rules (PREP-021).

Effort: low, about 10 percent, mostly consolidating the style guide and terms. Start with the standard spellchecker and add custom rules step by step.

Implementation paths

Which path for whom

• SMB: Path A, because mature, EU-hosted writing assistants with team glossaries exist at low per-seat prices and the process is standard.
• Enterprise: Path B, because brand voice, terminology and compliance checks should be enforced in the CMS and editor landscape through one shared API layer.
• Developer: Path C, because docs-as-code linting with open-source tools such as Vale or a self-hosted LanguageTool runs in CI without sending text out.
• Private: Path A, because browser and office add-ons are sufficient for personal writing.
• Public: Path C, because drafts of notices and press releases are often confidential and a self-hosted checker on EU infrastructure avoids procurement hurdles.

Path A — Buy (recommended for most SMBs)

• General writing assistants: LanguageTool (EU, team dictionary and style-guide rules, DPA under GDPR) [3]; DeepL Write Pro (EU, Cologne, styles and tones, Pro data not used for training, ISO 27001 / SOC 2 Type II) [4]; Duden Mentor (DE, strong for German); Grammarly (now part of the Superhuman suite since Oct 2025, with brand-tone profiles per team and agents in Superhuman Go) [6][8]; Microsoft 365 Copilot / Editor, Google Gemini in Workspace.
• Enterprise content governance: Markup AI (formerly Acrolinx, Berlin roots; rebranded Sep 2025 with 15 "Content Guardian Agents" for style, terminology, brand voice, claims and AI-search readiness, reachable through API and MCP) [2]; Writer (brand voice, terminology, Palmyra models, agent platform) [5]; Congree (Karlsruhe, DE; rule-based authoring with terminology, STE, gender-inclusive and easy language, combined with external GenAI) [10].
• Choose A when: writers work in Office/Google/browser and need a shared dictionary and style rules. This covers nearly every SMB. Check EU hosting, a no-training clause, SSO and terminology import (TBX/CSV).

Path B — Build on cloud APIs

• Specialised APIs: DeepL API /v2/write/correct and /v2/write/rephrase with writing_style or tone, 11 languages, Pro plan only [1]; LanguageTool HTTP API (cloud or self-hosted); Markup AI API.
• LLM rewriting: Mistral (EU provider; e.g. Mistral Medium 3.5, Mistral Small 4) [11], Claude, GPT or Gemini in EU data regions (Azure OpenAI EU Data Zone, Vertex AI EU, Bedrock Frankfurt). Put the style guide in a cached system prompt. Return structured JSON (original, suggestion, rule_id, explanation) so the UI can show diffs instead of a silently rewritten text.
• Integration points: Word/Outlook add-ins (Office.js), CMS plug-ins (WordPress, AEM, Contentful), Zendesk/Freshdesk macros, a pre-publish webhook, or an MCP tool for internal agents.
• Choose B when: you need checks inside your own product, CMS or support pipeline, or a check step after other generation use cases.

Path C — Self-host (open source)

• Rule layer: LanguageTool server (LGPL 2.1, Java, runs locally on port 8081, 30+ languages) [3]; Vale (MIT, YAML rules, community style packages, vale-action for CI) [7]; Hunspell dictionaries.
• LLM layer: Mistral Small 4 (Apache 2.0) [11], the Qwen3 family, Gemma 3 or Ministral 3 (8B/14B, Apache 2.0) [11], served with vLLM or Ollama, with structured outputs for the suggestion JSON.
• Hardware: a rule-based server runs on a CPU (HW1). LLM rewriting for a few hundred users fits one 24–32 GB GPU with a quantised ≤ 30B model (HW2). Ops effort is low to medium: update dictionaries and models, run an evaluation set of in-house texts after each change.
• Choose C when: texts are confidential (R&D, legal, health, public authority), or you already run a private LLM server (PERS-005 at home, or a company GPU server).

Path D — Custom training

Rarely worth it for general proofreading. It pays off in two cases. (1) A strong, distinctive brand voice or controlled language at high volume: a LoRA fine-tune of a small open model on several thousand approved before/after pairs or on-brand texts gives more consistent tone than prompting, at lower inference cost. (2) GEC for a low-resource language or domain jargon, using the corrections editors have accepted as training data. Keep the rule and terminology layer deterministic in either case.

State of the art (2026)

• Commodity proofreading: spelling and grammar correction is built into every office suite, browser and OS. Differentiation has moved to governance: company rules, terminology, brand voice, scores and quality gates across many tools via API/MCP [2][5].
• LLMs vs. specialised GEC: an ACL 2024 study of 10 open and commercial LLMs on four GEC benchmarks found that LLMs do not consistently beat supervised GEC systems. Commercial LLMs did best on fluency-edit benchmarks, and several open models beat commercial ones on minimal-edit benchmarks [9]. In practice, LLMs tend to over-correct and to change style when only correction was asked for.
• Checking AI output: new check categories target "AI-sounding" text, unsupported claims and readiness for AI search answers. Markup AI ships dedicated agents for these [2]. Writing tools are becoming agents that operate across apps (Superhuman Go) [6].
• Limits: LLM rewrites can subtly change meaning, figures or legal nuance. Brand-voice compliance is hard to measure objectively. Quality varies a lot between languages: English and German are strong, smaller EU languages are weaker. Explanations for LLM suggestions can be invented after the fact.

Shared resources & combinations

• Same LLM server (llm-general/llm-small, HW2): LANG-001 marketing copy, LANG-002 email drafting, LANG-005 SEO content, LANG-007 social posts, LANG-010 plain language. One self-hosted model plus one shared style-guide/terminology store serves all of them, and LANG-011 acts as the final check step in each.
• Terminology database shared with LANG-009 translation (glossaries) and LANG-022 technical documentation (CCMS).
• Downstream checks: LANG-024 promotional compliance (claims), LANG-021 fact-checking, DOC-026 / CODE-017 accessibility (alt text, readable language).
• Evaluation & observability: a regression set of in-house texts and logs of accepted suggestions, shared with CODE-018 LLMOps.

Effort, cost & KPIs

• Path A: days. Cost per seat per month (e.g. DeepL Write Pro add-on at €6.99 per user/month, billed annually [4]); enterprise governance platforms are quoted individually. The main work is turning the style guide into rules and importing terminology.
• Path B: 2–6 weeks for an add-in or CMS check. Token costs are low because inputs are short.
• Path C: 2–4 weeks. One GPU server; can be shared with other LANG use cases.
• KPIs: errors per 1,000 words before and after, suggestion acceptance rate, terminology compliance %, brand/quality score distribution, review cycles per document, time to publish, translation cost per word (through consistent terminology).

People & change

Involve management for the decision which tone and rules apply, and the department lead (marketing, communication) as owner of the style guide. End users take part from the start, since they decide daily whether suggestions help. IT handles accounts, plugins and hosting; the DPO checks AVV and what text may be sent to the provider.

Works council: yes, in practice. Writing assistants see everything employees type, and quality scores per document or per author are objectively suited to monitor performance, which triggers §87(1) no. 6 BetrVG (BAG: suitability suffices). Agree on a works agreement that excludes individual evaluation, limits logging and stores no author-level scores (GUIDE-004). Inform the council early under §90(1) no. 3; it may call an expert under §80(3). In the public sector, involve the Personalrat.

AI literacy: support staff with short sessions on how suggestions are produced, why a rewrite can change meaning, and what must not be pasted in (GUIDE-006). Communicate that suggestions are optional and that nobody is measured on acceptance (GUIDE-008). Choose a pilot group of volunteers from marketing and support, and keep humans in control: every suggestion is accepted or rejected by the author, and nothing is auto-published. Concerns about surveillance and loss of individual voice are best met with transparency and opt-out rules.

Risks, regulation & compliance

• EU AI Act: minimal risk. No Annex III category applies. Art. 50(2) expressly exempts AI systems that "perform an assistive function for standard editing" or do not substantially change the input or its meaning from the machine-readable marking duty [12]. Substantial rewrites published to inform the public on matters of public interest need disclosure under Art. 50(4), unless a human reviewed them and a person holds editorial responsibility [12]. Art. 50 applies from 2 Aug 2026. Art. 4 AI literacy applies to staff.
• GDPR: emails, HR letters and support answers contain personal data, sometimes special categories. Browser extensions that read every text field are a big risk for data leaving the company. Use business plans with a DPA, EU processing and a no-training clause, or self-host. Block consumer extensions by policy (MDM/browser management).
• Sector rules: regulated texts (pharma advertising, financial promotions, contracts) need the claim and legal checks of LANG-024. A style rewrite must never replace legal review. In the public sector, accessibility and plain-language duties (BITV 2.0, EAA) apply to published text.
• Risks & mitigations: meaning drift → diff view and correct-only mode for sensitive text; homogeneous "AI tone" → brand examples and an AI-voice check; over-reliance → keep human editorial responsibility; prompt injection from pasted content → guardrails and no tool access in the rewriting step.

Getting started

1. Collect what exists: style guide, terminology list, 20–50 approved texts that show the brand voice.
2. Turn the style guide into checkable rules (do/don't, spellings, forbidden terms) and import the terminology. This step matters more than which tool you pick.
3. Pilot 2–3 Path A tools (e.g. LanguageTool Team, DeepL Write Pro, Markup AI or Writer for enterprises) with 100 real texts. Measure acceptance rate and false positives.
4. Set data rules: business accounts only, EU processing, no training. Self-host (Path C) for confidential text.
5. Add a pre-publish check to the CMS or CI (Vale / API) and review acceptance statistics monthly.

Sources

1. DeepL — API reference: Improve text (DeepL Write API) — https://developers.deepl.com/docs/api-reference/improve-text
2. Markup AI — "Acrolinx is now Markup AI" (Content Guardian Agents, API/MCP), Sep 2025 — https://markup.ai/blog/acrolinx-is-now-markup-ai ; https://www.acrolinx.com/
3. LanguageTool — Business features and open-source server (LGPL) — https://languagetool.org/business ; https://github.com/languagetool-org/languagetool ; https://dev.languagetool.org/http-server
4. DeepL — DeepL Write Pro (pricing, data security) — https://www.deepl.com/en/pro-write
5. Writer — enterprise AI platform, customer results (KPMG, Vanguard) — https://writer.com/
6. Grammarly / Superhuman — "Introducing the new Superhuman" (Oct 2025) — https://www.grammarly.com/blog/company/introducing-new-superhuman/
7. Vale — prose linter documentation — https://docs.vale.sh/
8. Grammarly — Brand Tones for business — https://www.grammarly.com/business/brand-tones
9. Davis et al., "Prompting open-source and commercial language models for grammatical error correction of English learner text", ACL Findings 2024 — https://arxiv.org/abs/2401.07702
10. Congree Language Technologies — https://www.congree.com/en/
11. Mistral AI — Models overview — https://docs.mistral.ai/getting-started/models/
12. EU AI Act Service Desk — Article 50 (transparency obligations) — https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50