# AI-powered market intelligence workflows: an architecture that scales past the prototype > How to build AI-powered market intelligence workflows that scale: map questions to sources, split knowledge base from library, and ship defined outputs. Source: https://contengi.com/blog/ai-powered-market-intelligence-workflows Published: 2026-10-09 One clever prototype that summarises competitor news is easy to build. Ten of them, each with its own prompt, its own data and its own idea of what good looks like, is a mess. This guide walks through the architecture Stefan Maritz and Diego Cimadevilla Garcia discussed for turning isolated experiments into a system a team can run. A market intelligence workflow answers one defined question and returns the same template every time. The teams that scale this beyond a prototype stop collecting tools and start designing the layers underneath: which questions matter, where the answers live, what stable context the model always needs, and what working material it fetches on demand. The examples below come from Stefan Maritz and Diego Cimadevilla Garcia's conversation on 2 October 2026, so read them as one team's approach and treat the rest as a prompt for your own. ## TL;DR - Start with the questions your team needs answered, map each one to the source that can answer it, and check what your current tools already cover before you build anything. - Keep the knowledge base small and stable, with value propositions, competitors, segments, and brand guidance that every workflow reads on every run. - Put playbooks, transcripts, and meeting notes in a library, and retrieve only the passages each request needs. - Use live queries or a daily snapshot for numbers that move, such as pipeline stages and campaign spend, depending on how current the answer has to be. - Combine the layers into one workflow with a fixed output template so results compare week to week, then put a single assistant in front that calls each workflow. - Give every layer an owner, set confidentiality rules, and add source-quality checks once the basics are stable. ## Start with the questions, then find the sources The conversation begins with the questions the team needs answered, before any tooling. What are competitors saying about pricing this quarter? Which segments are we losing deals in? What did customers raise on calls last month? Each question gets mapped to the data source that can answer it. Mapping has a second half: assessing what you already own. Some questions the existing tools answer well enough. Others need a direct first-party API, because a dashboard export strips out the detail you need. Doing this audit first stops you building a pipeline for a question your current tool already handles. ## Keep a lean, curated knowledge base for stable context The [knowledge base](https://contengi.com/features/knowledge-base) holds the things that rarely change: value propositions, customer and competitor information, segments, brand strategy and writing guidance. Every workflow reads from it, so it earns its place by being small and accurate. The model gets the same ground truth on every run, whatever question you ask it. Every extra document you add dilutes the signal and costs tokens on every request. If a document is only useful for one project, it belongs somewhere else. ## Use a library for working materials Playbooks, transcripts, campaign notes, webinars, podcasts and meeting notes are working materials, and they live in a library. The library grows constantly and changes shape, which is exactly why it stays separate from the knowledge base. A [content library](https://contengi.com/features/library) gives you one place to keep all of it searchable. The conversation stresses selective curation and retrieval. You choose what goes in, and each request retrieves the relevant passages rather than loading whole documents. A 90-minute webinar transcript might contain two paragraphs relevant to a pricing question, and those two paragraphs are what the workflow should see. The same principle sits behind good [context engineering](https://contengi.com/blog/context-engineering), where the aim is the smallest set of tokens that carries the answer. ## Query live data when current values matter Some questions depend on numbers that change daily, such as pipeline stages, deal values or campaign spend. For those, the workflow either queries the API directly or reads from a snapshot. The Salesforce daily snapshot was the example given in the conversation: a copy of the relevant CRM data refreshed each day, so workflows read a stable, recent view instead of hitting the live system on every request. Whether a snapshot or live query fits depends on how quickly the values move and how current the answer needs to be. A weekly competitor pricing check has very different needs from a daily pipeline review. ## Combine the layers into a defined workflow A workflow is where the layers meet. It pulls stable context from the knowledge base, retrieves relevant material from the library, fetches live data from an API or snapshot, runs the analysis, and writes the result into a defined output template. Teams skip the output template, yet it is what makes outputs comparable week to week. A competitor brief with the same headings every time can be scanned in two minutes and compared against last month. A free-form summary cannot. [Anthropic's guidance on effective agents](https://www.anthropic.com/engineering/building-effective-agents) favours simple, composable patterns, which suits a workflow built from separate layers. ## One assistant, many callable workflows The conversation suggests a single conversational assistant that can call specialist workflows, in place of many disconnected agents. Someone asks a question in plain language, and the assistant routes it to the competitor brief, the pipeline review or the call-insights workflow. The team learns one interface, and every workflow draws on the same knowledge base and library. This is where open standards help. The [Model Context Protocol](https://modelcontextprotocol.io/docs/getting-started/intro) gives AI applications a standard way to connect to data sources, tools and workflows, so adding a new source does not mean rebuilding the assistant. A [marketing assistant](https://contengi.com/features/marketing-assistant) built on that pattern can grow one workflow at a time. ## Assign ownership, set confidentiality rules, then add quality checks Someone has to own each layer. The knowledge base needs an owner who keeps it accurate, the library needs a curator, and each data connection needs a person who knows what it exposes. Confidentiality rules decide which sources a given workflow may read and who may see the output. Backbase's governance setup suits one organisation, so use it as a prompt for your own rules based on your data, regulators and team size. Add source-quality checks once your questions, knowledge base and output templates are stable, because only then have you got something to check sources against. Teams that start with quality scoring before the basics exist end up scoring a mess. ## A practical order of work Start with the question list and source map. Build the knowledge base next, keeping it short enough that you can read it in one sitting. Add the library with a clear rule for what gets in, then connect one live source and build one workflow with one output template. When that workflow produces something the team uses, add the second. Retrieval also keeps costs sane. Selective retrieval means each request carries only the passages it needs, and stable context placed at the start of a request can be reused cheaply, as the [prompt caching documentation](https://docs.claude.com/en/docs/build-with-claude/prompt-caching) describes for the Claude API. ## Frequently asked questions ### What is an AI-powered market intelligence workflow? It is a repeatable process that takes a defined question, gathers context from a knowledge base, retrieves relevant working material from a library, adds live or snapshot data where current values matter, and produces an analysis in a fixed template. The repeatability is what separates it from a one-off prompt. ### What is the difference between a knowledge base and a library? The knowledge base holds stable context such as value propositions, segments, competitor information and writing guidance, and stays small and curated. The library holds working materials such as transcripts, playbooks, campaign notes and webinars, and the workflow retrieves only the relevant parts per request. ### Should I use live API queries or snapshots for dynamic data? It depends on how fast the data changes and how current the answer must be. In the conversation, a daily Salesforce snapshot was one team's example, and a live query suits cases where a stale value would give a wrong answer. ### Do I need many agents or one assistant? The conversation suggests a single conversational assistant that calls specialist workflows. Your team learns one interface, and every workflow shares the same knowledge base and library, which avoids a pile of disconnected agents with different context. ### When should I add source-quality checks? After the foundations are in place. Once your questions, knowledge base, library and output templates are stable, you have something concrete to check sources against, and the checks catch real problems.