Summary: Firecrawl published a guide to enrichment for AI agents on October 23, 2026, arguing that quality comes from routing each missing field to the source that knows it best. The piece includes fresh timing tests, a provider-accuracy benchmark, and details on Firecrawl Alexandria’s licensed data layer. For builders, the practical takeaway is architectural: enrichment is a routing and provenance problem, not a single API call.
Why stale records hurt agents
An AI agent acts on whatever a record says. When a field is missing or stale, the model either guesses or carries the error forward. Validity’s State of CRM Data Management in 2026 report, released August 25, 2026, puts numbers on the risk: 92% of surveyed SVP and VP marketers said they acted on an AI recommendation they later suspected was wrong because of bad data, and only 21% called their CRM data “very well prepared” for AI. Meanwhile, two in three organizations gave more marketing decisions to autonomous agents over the past year.
More decisions, patchier records. That gap is why enrichment matters now, and why each appended field should carry a source and timestamp an agent — or a reviewer — can check.
Route fields by how fast they change
Firecrawl groups enrichment targets into seven types: company, people, product, financial, risk, technical, and event data. The routing logic is simple:
- Stable reference fields (HQ, founding year, industry) go to a provider database.
- Facts a company states about itself (price, specs, policies) come from its own pages via structured extraction.
- Events (funding, launches, open roles) need agentic web research, because a week-old fact hasn’t reached most databases yet.
In Firecrawl’s timing tests on five companies on October 23, 2026, provider lookups through Alexandria (using People Data Labs) took a median 1.8 seconds at 50 credits, structured extraction via /scrape took 1.5 seconds at 5 credits per page, and an /agent research run took about 42 seconds. Speed, cost, and freshness trade off predictably — pick per field, not per project.
Provider accuracy is uneven where it matters most
Company data is the most measured enrichment category, and the benchmark results are worth internalizing before you trust any provider blindly. In Firecrawl’s test of 200 US companies, Apollo led four common fields at 89.3% and People Data Labs placed second at 80.6%. But employee count averaged just 57% across six providers, and revenue answers for public companies split sharply — point estimates like Apollo’s (82.8% within 5% of the filing) beat range-based answers.
I covered related ground in picking a company enrichment API when public scorecards exist; the same rule applies here. Decide which source your agent trusts for each field, check headcount against a second provider before it drives routing or scoring, and store where every value came from.
One interface across five tool categories
Firecrawl lays out five enrichment tool types: provider databases, waterfall platforms, CRM-native enrichment, web extraction APIs, and agentic research APIs. Agents tend to need a mix, reachable through one interface instead of five integrations. Firecrawl Alexandria covers licensed provider data — company, people, financial, product, public records, and developer data — through the same API key, MCP server, and CLI as /scrape. Providers are paid per pull in credits, with no per-vendor minimum, though some require a one-time org admin approval of terms.
The claim that this quality gap is real rests on a September 2026 Firecrawl test: across 845 tasks with the same model and prompts, agents using Alexandria scored 21% higher on blind-judged answer quality than agents with built-in web tools. That’s Firecrawl’s own evaluation of its own product, so treat the margin as directional — but the setup (same model, same prompts, blind judging) is at least a reasonable comparison. If you’re wiring tools into Claude Code or Codex, my earlier MCP server roundup covers how enrichment fits into a broader agent toolbelt.
Practical next step
Start with your match keys, not your tools. A stable domain, SKU, ticker, or package URL makes every downstream lookup land; a bare company name collides. Then map each missing field to its fastest-changing source, define the JSON schema up front, and store source plus retrieval time beside every value. A record like “2,510 employees, People Data Labs, retrieved October 23, 2026” is one an agent can cite and a human can audit — that provenance is what makes enrichment safe to automate against.