- ×5
- articles processed every day
- ×10
- faster to qualify an opportunity
- −58%
- false positives
1. The starting point
L-Acoustics follows the news in its markets to spot, as early as possible, the projects that can become sales opportunities. This monitoring relied on people: reading articles, sorting what matters, qualifying each project, then alerting the right team.
Your teams spend their days reading, sorting, qualifying. Meanwhile opportunities go cold and competitors close. The paradox: the more you invest in monitoring, the less it pays off.
The assessment shared at the start was harsh: with manual monitoring, more than 70% of opportunities went unnoticed.
2. Three friction points
Before talking technology, we sorted the teams’ difficulties into three families. Each calls for a different answer.
- Timing
- An opportunity detected too late is lost to the competition. The value of information drops over time.
- Noise
- Lots of noise, little signal. Teams waste time sorting information that does not concern them.
- Volume
- Hundreds of articles a week. Impossible to read everything, hence blind spots.
3. The promise: from article to qualified opportunity
- An agentic framework
- AgentsAgentA program that uses an AI model to carry out a task: it reads information, calls tools and suggests or performs actions.See the glossary that extract the business-relevant information from each article: the project, the place, the players, the deadline, the budget.
- Scoring
- An objective, explainable qualification: each score rests on readable criteria, not a black box.
- Actionable for the field
- The opportunity goes straight to the sales team concerned, in the tools it already uses.
4. A four-step pipeline
- 01CollectFeedly
- 02IndexAgentAI Search
- 03AnalyseAgents
- 04DistributeWarehouseExcel
- Collect: sources are followed in Feedly, organised by market vertical.
- Index: an agent extracts the content of each article and indexes it in Azure AI SearchAzure AI SearchMicrosoft Azure’s search engine, which indexes documents so they can be found by keyword or by meaning.See the glossary , so it can be found and compared.
- Analyse: agents enrich each article and assess its commercial potential.
- Distribute: the opportunities kept are consolidated in the data warehouseData warehouseA database organised for analysis: cleaned, linked data ready for dashboards.See the glossary , then sent to the teams for approval.
On the processing side, the agenticAgentic AIAI that does more than answer: it chains steps together and acts in tools to reach a goal.See the glossary frameworkFrameworkA development toolkit that provides an application’s structure; for agents: LangGraph, Strands, CrewAI…See the glossary runs in Azure FunctionsAzure FunctionsA Microsoft Azure service that runs code on demand, without managing servers.See the glossary , in two main enrichment stages: extract and index first, then enrich and analyse. Between stages, data travels as JSONJSONA simple text format for exchanging structured data between programs.See the glossary metadata, which makes each stage testable and replaceable.
5. The data architecture on Microsoft Fabric
Opportunities are processed, stored and distributed across three complementary layers.
| Lakehouse | Warehouse | Excel |
|---|---|---|
| DeduplicationDeduplicationSpotting and removing duplicates, for example two articles about the same project.See the glossary : duplicates are detected by semantic similaritySemantic similarityA measure of how close two texts are in meaning, even when they use different words.See the glossary ; above 0.90 they are excluded automatically. | Consolidated views: deduplicated opportunities, statistics by area, team coverage. | Weekly export: a file with the week’s new opportunities to approve. |
| Reference tables: sales teams by geographic area and by vertical. | Business joins: each opportunity is linked to the sales lead for its area and vertical. | Two-way sync: the sales teams’ decisions (approved or rejected) come back every day. |
| History: every raw opportunity is kept, for traceabilityAudit logThe record of who did what, when and with which data. It lets you check and explain every action.See the glossary and retrospective analysis. | BI feed: Power BIPower BIMicrosoft’s dashboard and data visualisation tool.See the glossary dashboards and tracking of sales indicators. | Notifications: an email goes to the lead concerned as soon as an approved opportunity concerns them. |
Three layers: prepare the data (Lakehouse), link it to the business (Warehouse), put it in the teams’ hands (Excel).
Choosing Excel is not a detail. Teams approve where they already work, and their decisions flow back into the platform: that feedback is what lets the rules improve.
6. Results
| Indicator | Before | After |
|---|---|---|
| Articles processed per day | 10 to 20, read by hand | 50 to 100, analysed in full |
| Qualifying an opportunity | 30 to 60 minutes | immediate, with a reasoned score |
| Share of noise | 60% | 25% |
Scoring assesses commercial potential from the project phase, timing, budget and competition.
The drop in noise does not come from a more powerful model, but from explicit business rules: a project already finished, a deadline too close or uncertain funding no longer reach the teams. Only real opportunities land on their desk.
7. What we take from it
- Start from the field’s frictions — timing, noise, volume — rather than from technology.
- Write the business rules down: they do more for quality than the choice of model.
- Make the score explainable, so sales teams trust it.
- Deliver in existing tools, and feed decisions back to improve the system.
- Keep human approval: the agent qualifies, the team decides.
Further reading