Content received via webhook (social scheduler): LinkedIn post scheduled for Thursday 15:00. Review: 2 deviations from the guide.
AGENT · BRAND VOICE MONITOR
The brand voice covered before publication.
Brand Voice Monitor reads every piece of content about to leave the company — emails, social posts, brochures, press releases — and compares it against the brand voice guide. It flags tone deviations, words outside the corporate vocabulary, formulations the brand avoids. It proposes concrete alternatives before publication.
Brand Voice Monitor at work.
⚠️ 'revolutionary': the brand avoids unsubstantiated superlatives. Alternative: 'a different approach from standard tools'. ⚠️ 'find out more': the brand replaces generic CTAs. Alternative: 'see the full case study'.
I'll take the first one, for the CTA I'll use 'read the case study'. Going ahead with publication.
Decisions recorded. Event traced in audit. Good publishing.
Why it exists.
Brand voice is one of the hardest assets to keep consistent in a mid-large company. Companies invest in style guides and workshops with the marketing team; then in practice the content going out every day comes from dozens of different people — internal team, external agency, freelancers, sales posting on LinkedIn, customer success replying by email. Consistency gets lost in the volume.
How it reviews before publication.
Brand Voice Monitor works as a silent reviewer. It activates pre-publication: reads the content, compares it against the declared guide, recognises deviations, proposes alternatives. The deviations flagged are of three kinds: tone (more formal or informal than the brand register), vocabulary (words the brand avoids or that are not in the corporate vocabulary), structure (headline too long, opening pattern inconsistent with the guide).
The decision stays with the content author.
The human reviewer decides what to apply. The agent does not publish on its own, does not correct silently, does not modify creatives without approval.
Three roles, three different perspectives on the value.
Brand manager
Sees, in a structured way, the consistency of the published content flow — not by sampling. Recurring deviation patterns emerge from the volume and become the subject of guide updates or targeted workshops.
Cross-channel · 7d 1,420 mentions
Site chat NPS 7.2
Reviews top theme: shipping
Social · IG sentiment +0.4
Brief to the brand owner
Head of marketing
Reclaims time from manual review. Automatic coverage handles most standard cases; the manager's capacity concentrates on strategic content and complex cases.
Promo #2024-S03 Consumer Code
Health claim verify DM 219
Green claim COMPLIANT
Price transparency COMPLIANT
1 suggested change · pre-publish
Content author
Sales, customer success, freelance, agency: receives the review right next to the text they wrote. The brand voice learning curve accelerates because the feedback arrives on the concrete case, not in a workshop weeks later.
Proposal no. 2024-081 In review
Missing disclosure
MiFID II art. · regulated financial instrument
Alt. 1 …in compliance with MiFID II and applicable supervisory provisions.
Alt. 2 …with full disclosure attached to the offer document.
Audit trace recorded · 14:31
A LinkedIn post goes through review before entering the scheduler.
The agent is integrated with the social scheduler via webhook.
For a regulated mid-size B2B SaaS company, the marketing team publishes five to eight pieces of content on LinkedIn every week. The brand voice guide declares a professional but direct register, no industry buzzwords, no superlative claims, concrete examples always before principles. The agent is integrated with the corporate social scheduler via webhook, configured during delivery.
Two deviations recognised, two alternatives ready.
In seconds the agent runs the analysis. It recognises two deviations: the opening uses 'revolutionary' (the brand avoids superlatives); the closing has a generic CTA 'find out more' that the brand always replaces with a concrete formulation. The agent proposes two ready-to-use alternatives.
The content manager applies the alternatives and publishes.
The content manager sees the outcome on the marketing Slack channel. They apply the alternatives and confirm publication. The event — content reviewed, rules triggered, alternatives chosen — stays in the audit registry, readable by the brand manager with a standard SQL client.
Configuration and technical resources.
The Brand Voice Monitor rules are declarative. The customer's brand or marketing team defines in a readable format the brand vocabulary (words to prefer, words to avoid), the register patterns (formal, informal, technical, accessible per content category), the structural patterns (maximum opening length, CTA format, headline structure). The rules live in the customer's repository, versioned, updatable without Exelab team involvement.
The integration pattern with publication tools — CMS, email marketing, social scheduler — is delivered during the project through the customer's systems' APIs. The three most common activation modes are: pre-publication via webhook, manual review via channel, periodic batch for monthly pattern analysis.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, topic-guardrail, message-length-limit
- Native channels
- Telegram, Slack, WhatsApp, OpenAI-compatible HTTP
- CMS integration
- dedicated adapter delivered during the project (Contentful, Adobe Experience Manager, Sitecore, Strapi)
- Social scheduler integration
- dedicated adapter delivered during the project (Sprinklr, Hootsuite, Buffer, Khoros)
- Email marketing integration
- HubSpot native; Adobe Marketo Engage, Salesforce Marketing Cloud via delivery
- Memory
- persistent per instance, pgvector + PostgreSQL FTS on the brand guide
- Registry
- immutable, queryable with a standard SQL client
Frequently asked questions about the agent.
No. The agent reads the content, recognises deviations from the guide, proposes alternatives. The decision on what to apply stays with the content author or brand manager. The agent does not publish, does not correct silently, does not modify creatives without approval.
The runtime's native channels are Telegram, Slack, WhatsApp, and OpenAI-compatible HTTP. Integration with CMS (Contentful, Adobe Experience Manager, Sitecore), email marketing and automation (HubSpot native; Marketo, Salesforce Marketing Cloud via delivery), social schedulers (Sprinklr, Hootsuite, Buffer via delivery) is delivered during the project by the Exelab team.
The rules are declarative in a structured format (YAML or markdown) that the brand team reads and modifies without writing code. The typical pattern includes: a flag-word list (preferred and avoided), structural patterns (maximum opening length, CTA format), target tone per content category (LinkedIn post vs customer email vs press release).
Yes. The persistent memory per instance saves the customer's recurring patterns (words that come up often, alternatives that get accepted). The declarative rules stay editable: when new patterns emerge, the brand team modifies the guide, promotes it to production, and the agent runs the updated rules from that point on.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how Brand Voice Monitor would configure to the customer's case. Which brand voice guide, which publication tools, which review flow.