Every market intelligence vendor now claims to be "AI-powered." The term has become so broad that it means almost nothing. At Delphi, we use machine learning and large language models extensively in our data pipeline. But we've also learned exactly where AI delivers real value and where it falls apart — and we think being honest about those boundaries is more useful than another pitch about how AI will transform everything.
Where AI works: signal processing at scale
The most impactful application of AI in our pipeline is processing raw signals at scale. Every day, thousands of new documents enter our system: press releases, regulatory filings, patent applications, company announcements, trade publications, government tenders. A human team reading all of these would need hundreds of analysts working around the clock.
Language models excel at extracting structured facts from unstructured text. "Company X announced a 200 MW electrolyzer project in Region Y, using Technology Z, with an expected commissioning date of Q3 2027" — extracting these entities and attributes from a press release is something a well-prompted model does reliably and quickly.
Named entity recognition, relationship extraction, and document classification are all areas where AI has moved from experimental to production-grade. We use these capabilities every day, and they're genuinely transformative for throughput.
Where AI struggles: entity resolution
Here's where it gets harder. The same company might appear as "thyssenkrupp nucera," "tkNucera," "Thyssenkrupp Uhde Chlorine Engineers" (the former name), "TKUCE," or a dozen other variants across different sources and languages. The same project might be described with different names in a company press release, a government subsidy list, and a local newspaper article.
Entity resolution — determining that these different mentions refer to the same real-world entity — is one of the hardest problems in data engineering. AI helps (embedding-based similarity matching, for instance), but it can't fully solve the problem without human-curated reference data, domain-specific rules, and continuous quality checks.
This is where our knowledge graph architecture matters. Once an entity is resolved and assigned a canonical ID, every future mention of that entity can be linked automatically. But building and maintaining the canonical reference — the ground truth — requires human expertise that no model currently replaces.
Where AI fails: verification and provenance
The most dangerous application of AI in market intelligence is also the most tempting: asking a model to answer market questions directly, without structured data underneath.
"How many hydrogen projects are under development in Europe?" If you ask a language model this question, you'll get a fluent answer with a specific number. That number will be wrong — or rather, it will be plausible but unverifiable. The model is synthesizing training data, not querying a structured database. There's no source trail, no timestamp, no way to know which projects are included or excluded.
For casual research, this is fine. For decisions involving capital allocation, procurement strategy, or policy design, it's dangerous. The fluency of the answer obscures the absence of rigor underneath.
This is why we use AI as a synthesis layer on top of structured data, not as a replacement for it. When Aletheia answers a question, the facts come from our knowledge graph (structured, sourced, timestamped) and the language model's job is narration, not fact-finding.
The honest picture
AI has compressed certain parts of our pipeline by orders of magnitude. What would have taken a team of 50 analysts to process manually, we can handle with a fraction of that. But the hard parts of market intelligence — deciding what to track, building domain-specific ontologies, resolving ambiguous entities, verifying data quality, maintaining provenance — remain deeply human tasks that AI assists but doesn't automate.
We think the companies that will build the most valuable market intelligence products are the ones that understand this distinction clearly: use AI where it's strong, invest in human expertise where it's essential, and never pretend that one replaces the other.