AI does not help marketing by producing more: it helps by producing the same with the brand on it, at a cost you can look at. We put the brand guide inside the system and measure how much content ships without a human rewriting it.
3–6 weeks to the first flow in production · 10 years of engineering, not of agency work · 48 hours to a proposal with fixed scope and price
Built on
Ninety seconds: who we are, how we work and what you get at the end.
WhatsApp, the website or email land in the same system: it understands the question, searches your documents, answers with a citation and hands over to a person when unsure.
Example: you need to email customers who stopped buying. The system drafts it in your brand voice.
01 Channels
Searches by meaning and by exact words, and filters by the tag before choosing.
answers with a source · 91 of 100
Swipe to follow the flow →
01
You personalise for three segments because that is what the team can handle, not what the data supports. The limit is operational, not analytical.
02
Follow-up depends on somebody remembering, and the sequence that does exist was written for the average of the list.
03
By the time the report shows which campaign worked, the month’s budget went where it was planned.
04
Between writing, review, design and build, the agility is gone before it reaches the channel.
05
It is the real reason most content pilots get abandoned: producing is cheap, correcting is not.
01
The technical part
Asking a model to "write in the brand’s tone" produces what the model believes that brand is. What works is different: the guide becomes retrievable context — tone, required and banned vocabulary, structure per format, and above all a set of already-approved assets that serve as examples — and the system retrieves the closest examples before writing.
The second decision is where you cut. A content system that publishes without review fills a site with assets nobody signed, and that is exactly what Google’s helpful-content updates penalise. We design with mandatory human review at the output, and the success metric is how much reaches that review already usable.
And segmentation: it pays when it runs over connected first-party data — CRM, on-site behaviour, purchase history — not over sector assumptions. That usually means integrating first, which is why the right marketing project often starts as an integration project.
Out of every hundred assets generated, how many clear review without somebody writing them again.
How it gets fixed
The set of approved examples gets widened and split by format: what works for an email does not work for a product page.
On a sample reviewed by the brand team: tone, banned vocabulary, structure, length.
How it gets fixed
The guide goes in as retrievable context, and banned words get verified after generation rather than requested in the prompt.
What each asset costs in model calls, with task-based routing in place.
How it gets fixed
A smaller model for the simple tasks, context trimming, and caching what repeats across assets.
When an asset mentions a price, a date or a feature, it comes from the system rather than the model.
How it gets fixed
Retrieval against the real source, and a ban on asserting figures that did not come from it.
If the team publishes four assets a month and the bottleneck is strategy, a generation system fixes nothing — it multiplies the problem.
And if the goal is filling a blog with generated articles, we are not the vendor. That is the practice Google has been penalising for two years, and the site that applies it loses the traffic it already had. What we do stand behind is producing with the brand on it and human review at the output.
Lead qualification and follow-up, measured against the real CRM.
Applied where there is volume and stable rules, not where there is expectation.
LangGraph orchestration, durable state and human approval on steps with consequences.
MCP servers against the systems already in operation.
Key takeaways
Marketing’s problem with AI is not a shortage of volume: it is that the volume produced does not sound like the brand, somebody ends up rewriting it, and nothing was saved.
The brand guide has to live inside the system — tone, banned words, structure, approved examples — and adherence gets measured on a sample rather than assumed.
The metric that matters is what share of assets ships without a human rewrite, with cost per asset beside it. Without those two numbers, "AI helps us with content" is an impression.
Segmentation pays when it runs over connected first-party data — CRM, on-site behaviour, purchase history — not over what the model believes about a sector.
An AI system is worth what its sources are worth. These are the standard connectors; anything with an API or a database connects the same way, and what has no API is handled by file.
SAP
Enterprise ERP
Oracle
ERP and database
NetSuite
Cloud ERP
Salesforce
CRM and service
HubSpot
CRM and marketing
PostgreSQL
Database and pgvector
Microsoft SQL
Database
Snowflake
Data warehouse
BigQuery
Google data warehouse
Databricks
Data platform
Redshift
AWS data warehouse
Synapse
Azure data warehouse
Supabase
Managed Postgres
Workday
Payroll and HR
QuickBooks
Accounting
Sage
Accounting and ERP
Xero
Cloud accounting
Shopify
Catalogue and orders
WooCommerce
Catalogue and orders
Magento
Catalogue and orders
Stripe
Payments and subscriptions
Google Drive
Documents and folders
CSV y Excel
Flat files
Nothing in this category
We built and operated the assistant for a loyalty platform serving more than two million active users across ten countries.
We built the full pipeline: document ingestion and normalization, chunking, embedding generation, vector store on Azure AI Search and Pinecone, and retrieval with grounded generation on LangChain.
We held it above 99% availability for three years.
RAG systems →Not for being AI-generated: Google said so explicitly. What gets penalised is content made to rank rather than for a reader, which covers most content generated without judgement. The practical difference is mandatory human review at the output, factual claims coming from a real source rather than the model, and somebody putting their name on it. That is why we design the system to reach review with something usable, not to skip it.
By turning the guide into retrievable context instead of an instruction. In go the tone, the required and banned vocabulary, the structure per format, and a set of already-approved assets; before writing, the system retrieves the examples closest to the brief. After generation we verify the banned words, which is a cheap check that catches most of what sounds wrong. And we measure it: adherence on a sample reviewed by your brand team.
Yes, and it usually decides whether the project is worth doing. Useful segmentation runs over first-party data — lifecycle stage, on-site behaviour, purchase history — and that data lives in the CRM. We integrate over API or with MCP servers against whatever exists, and we work with what sits around it too: the email platform, the CMS and the ads tooling.
It depends on volume and routing, and it is a figure we insist on putting on the table before building. Cost per asset is measured from day one and controlled with three things: a smaller model for the simple tasks, trimming the context sent, and caching what repeats across assets. A content system without cost control becomes a conversation with finance in month three.
No, and anyone pitching it that way is offering you a problem. It takes away the mechanical part — twenty variants of the same ad, adapting one asset to five formats, the first draft of a product description — and leaves what needs judgement: what to say, to whom and why. A small team with a good system produces like a bigger one; a system without a team produces noise.
Three to six weeks for a first flow in production, with something usable from week one. The timeline is driven more by the state of your inputs than by the build: if the brand guide is written and there are approved assets to serve as examples, it moves fast; if the guide is "we know it when we see it", the first job is writing it, and we do that with the team.
It gets fixed, it gets built, or it is not worth it. And if the diagnostic does not reach three actionable findings, it is not charged.
We use cookies to improve the user experience. Privacy
