What a Fractional AI Data Ops Agent Does

It is Monday, 7:48 AM. Your Head of Data opens the agent inbox before the standup. A 340-word brief sits at the top of the queue. 420 dbt models live across six schemas, 94 downstream metrics tied to a published definition, 42 models flagged red on freshness or test failure in the last 24 hours, the per-domain owner nudge routed to the three analytics engineers with the model, the failing test, and the suspect source table pre-cited, the six executive asks queued against the Thursday metric brief, and the Friday data-quality brief staged against the Q4 reliability number. The 9:00 AM data standup opens with the broken-model list already cleared.

Two quarters ago, the same Monday read differently. The Head of Data opened Snowflake at 7:48 AM, pulled a weekend dbt run log, chased three analytics engineers in Slack to triage overnight failures, hand-answered two Sunday-night Slack asks from the CFO on cash-collection cohorts, and walked into the 9:00 AM standup at 9:21 AM with half the model graph reviewed and no answer on the 42 red models bleeding against the Q4 reliability target. 28 hours of weekend chase bought the function one Monday standup. The agent shipped the exception list, the metric brief, and the Friday reliability draft by 7:48.

What the function was supposed to own

A Data team sits at the junction of five systems. Snowflake or BigQuery on the warehouse side, dbt on the transformation side, Looker or Hex or Mode on the BI side, Fivetran or Airbyte on the ingestion side, and the Slack channel where every ad hoc "can you pull" lands between 9:00 AM and 6:00 PM. The job description reads data modeling, metric definitions, reliability, stakeholder reporting, and the quarterly data strategy against the ARR per analyst line. The reality reads 38 hours a week on ticket chase and broken-pipeline triage, four hours on the Thursday metric brief, and one QBR a quarter where the data-trust question gets answered against a stale catalog.

The analytics stack is a function with a cadence. 420 dbt models, 94 published metrics, 320 Looker explores, 42 dashboards, and 68 ad hoc Slack asks a week at peak. The queue runs on a daily rhythm because the executive reports read a Thursday refresh and the exec dashboard reads a Monday cash cohort. The ticket chase and the broken-model triage sit between the operator and the modeling work, and the chase eats the week.

What the agent runs against on day one

The agent reads against Snowflake or BigQuery for the warehouse state and query history, dbt for the model graph and test results, Looker or Hex for the dashboard and metric object, Fivetran or Airbyte for the connector run log, and the Slack channel for the ad hoc ask flow. The read layer sits inside the systems the Data team already works in. The brief sits inside the operator inbox the Head of Data already checks before coffee. No new tool goes on the stack.

Model-quality pass. The agent reads the 420 dbt models against the last 24 hours of runs, cites the 42 flagged red on freshness or test failure, groups the failures by root-cause source (Fivetran connector lag, upstream schema drift, failed assertion, stale snapshot), flags the 11 downstream dashboards impacted, and routes the Monday nudge to the three analytics engineers with the model, the failing test, and the suspect table pre-cited. The weekend triage that used to burn four hours drops to a 15-minute exception review.

Metric-definition pass. The agent reads the 94 published metrics against the dbt metrics layer and the Looker LookML, cites the seven metrics with inconsistent definitions across two or more consumers, flags the two revenue cuts where the finance and sales reports disagree by $420K on the September close, and routes the reconciliation ask to the data lead and the FP&A lead with the SQL diff and the reconciled definition pre-drafted. The metric drift that used to surface at the Thursday CFO review lands cleared on Monday.

Ad hoc-ask triage pass. The agent reads the 68 Slack asks from the last seven days against the published catalog, cites the 42 that already have a dashboard or a metric match, routes the self-serve link back to the asker with the saved query, flags the 14 that need a new cut and ranks them by requester seniority and expected downstream reuse, and queues the remaining 12 for an analyst block on Wednesday. The ad hoc chase that used to eat 22 hours a week drops to a 90-minute Wednesday block with the pipeline already ranked.

What the agent ships on cadence

Monday 8:00 AM. Data-quality brief. 420 dbt models, 42 flagged red, 11 impacted dashboards, the per-engineer Monday ask routed with the model, test, and suspect table pre-cited, and the three executive asks still open from the previous week queued against the Thursday metric brief.

Tuesday 8:00 AM. Ingestion-reliability brief. The 68 Fivetran and Airbyte connectors against the last seven days of run logs, the four with a 24-hour lag flagged, the two with a schema-drift warning routed to the owning engineer, and the one source where volume dropped 40 percent against the trailing 14-day baseline flagged for upstream review.

Wednesday 8:00 AM. Ad hoc-ask reconciliation. The 68 Slack asks from the last seven days, 42 routed to self-serve, 14 ranked for the analyst block, 12 queued for next week, and the two recurring asks promoted to a dashboard-build ticket.

Thursday 8:00 AM. Executive metric brief. The six executive asks against the published metric layer, the three questions pre-answered with the SQL, the dashboard link, and the trailing 90-day context, the two outliers flagged against the plan, and the Friday ask routed before the exec dashboard refresh.

Friday 8:00 AM. Weekly reliability brief. The 420 models against the test-pass rate, the 94 metrics against the definition-sync rate, the 42 red models cleared or escalated, the SLA on the Thursday brief delivery, and the four strategy asks ranked by expected trust movement. The Head of Data reads the brief at 8:08 AM and the executive dashboard refresh at 10:00 AM opens with the numbers already cited.

What the operator does with the hours back

The 26 to 34 hours a week the agent clears land on four decisions. The metric-definition cleanup against the seven conflicting cuts the agent surfaced. The ingestion-contract review against the two sources carrying 61 percent of the weekly connector lag. The dashboard-retirement pass against the 94 dashboards with fewer than two views in the trailing 60 days. The data-strategy ask against the two executive questions the brief flagged as recurring. The function reads a Monday brief and ships four decisions by the Friday reliability brief. The queue runs on cadence.

The Head of Data stops being the ticket-triage person and starts being the metric-architecture person. The CFO stops Slack-asking for the cash cohort on Sunday night. The number is already in the Thursday brief. One operator owns the Monday review per domain, the agent clears the queue, and the Friday reliability brief opens with the exception list cleared, the ad hoc asks resolved, and the metric drift closed.

The unit economics on the function

A $190K to $240K Head of Data reads $260K to $340K loaded against the US entity. Three analytics engineers at $160K base read $220K to $280K loaded each, call it $660K to $840K on the engineering seat line a year. 38 hours a week on ticket chase, broken-model triage, and ad hoc pulls at a loaded hour of $85 to $140 costs the function $155K to $280K a year in senior time on queue work. A 14-day sprint to stand up the Data Ops agent runs in the low to mid five figures. Ongoing cost lands at $3K to $8K a month on API spend, Snowflake and Looker read access, and the operator who owns the Monday review.

Pull the same math against two incremental analytics-engineer hires. The $160K base reads $220K to $280K loaded, 120-day ramp, 180-day replacement risk, and five quarters before the model-coverage contribution clears the loaded cost. The sprint-versus-hire math lands three to six times cheaper on year one for the triage and reconciliation work. The current engineering team stays on the plan and lands on the modeling and architecture lane. The two incremental hires get replaced with a sprint. The $460K line clears to the metric-definition cleanup above the queue.

What the sprint ships in 14 days

Day one, the agent reads against Snowflake or BigQuery, dbt, Looker or Hex, Fivetran or Airbyte, and the Slack ad hoc channel. Day seven, the first Monday reliability brief lands in the Head of Data inbox with the 42 red models routed to the three analytics engineers. Day fourteen, the Thursday executive brief lands at 8:00 AM with the six questions pre-answered and the Friday reliability brief sitting on cadence. Day fifteen, the function sits on the same cadence it will run every week for the next two years. One operator owns the Monday review per domain. The agent clears the queue.

If your Monday reads 42 red models and the CFO is still Slack-asking for the cash cohort on Sunday night, the version where the brief lands at 8:00 AM and the Head of Data owns the metric architecture is fourteen days away. Snowflake still stores the warehouse, dbt still runs the models, and Looker still serves the dashboards. The agent closes the queue the stack was already supposed to close. You can keep running the Monday standup on a weekend dbt log, or you can scope a sprint and have the Thursday metric brief landing before the next executive review.