Data Work Gets More Valuable, Not Less

The expectation was that better models would reduce the need for careful data work — the model figures it out, so the cleaning and structuring matters less. The opposite is happening, for a mechanical reason: agents consume data directly and act on it, which makes data quality an operational property rather than an analytical one.

Why the value goes up

Agents act on data without a human sanity check. An analyst looking at a report notices when a number is obviously wrong. An agent given the same number acts on it. The buffer between bad data and consequences was a person, and automation removes it.

Retrieval quality is data quality. An agent's answers are bounded by what it can find. Poorly structured documents, missing metadata, inconsistent identifiers, no clear source of record — these become wrong answers rather than inconvenient searches.

Context is data. The unwritten organizational knowledge that agents most lack is, when captured, a data problem: what gets recorded, in what structure, with what freshness, findable how.

Freshness becomes operational. Stale data in a report is a caveat. Stale data driving an autonomous action is an incident — and freshness metadata, which analysts carry in their heads, has to become an explicit field.

Identifiers matter more. Agents follow entities across systems. Inconsistent identifiers, which humans paper over by knowing that these two records are the same customer, become hard failures.

The specific work that appreciates

Sources of record. Deciding which system is authoritative for each fact, and making that decision legible. Agents encountering three answers need to know which one counts.

Semantic documentation. ⚠️ Not schema documentation — meaning documentation. What the column represents, in what units, with what exclusions, and which business rules apply. A schema dump tells an agent the column is a bigint; it doesn't say the totals include cancelled orders and shouldn't be summed for revenue. That sentence is the difference between a right answer and a plausible one.

Entity resolution. Knowing that these records are the same thing. Previously tolerable to do imperfectly; now it determines whether an agent's traversal works.

Freshness and provenance as first-class fields. Every fact carrying when it was true and where it came from.

Access structure. Curated views, scoped credentials, and query surfaces that make the right data reachable and the wrong data unreachable.

Quality monitoring. Distribution checks, cardinality checks, reconciliation — automated, because nobody is eyeballing the numbers any more.

💡 The reframe for data teams

The traditional customer was an analyst or a dashboard — a human who would notice anomalies, ask questions, and apply context.

The new consumer has none of that. It takes what it's given at face value, can't tell that a number is implausible, and acts. Data work shifts from serving informed consumers to serving literal ones, which raises the bar on everything that used to be handled by the reader's judgment.

→ Practically: the metadata that lived in analysts' heads has to become fields. That's the whole job change in one sentence.

✅ What to prioritize

• Semantic documentation for the tables agents touch. Highest immediate return, and it's writing rather than engineering.
• Freshness fields on everything, surfaced in results.
• Entity identifiers reconciled across the systems agents traverse.
• Curated views rather than raw table access.
• Automated quality checks, because the human sanity check is gone.
• A clear source of record per fact, documented.

The takeaway

Agents consume data literally and act on it, which removes the human buffer that used to catch bad data before it caused anything. That makes semantic documentation, entity resolution, freshness, provenance, and quality monitoring operational requirements rather than analytical niceties. The core shift is that the metadata analysts carried in their heads now has to exist as fields — which is more work than it sounds, and it's the work that determines whether any agent touching your data is right or merely plausible.