Fractional CDO

Fractional CDO (Chief Data Officer)

Two of your teams brought two different revenue numbers to the same meeting last quarter, and the discussion became about the numbers instead of the decision. Nobody was wrong. They were counting different things, using definitions never written down, from systems that have never agreed — and the meeting ended with someone assigned to "reconcile it," a task with no completion criteria.

A fractional Chief Data Officer — one of three fractional executive seats Facet places — ends that. Facet places a senior data executive into your leadership team on a fixed monthly fee to own the definitions, lineage, quality, and access rules that decide whether your numbers can be trusted — backed by the data engineering team that builds against those decisions and the Agent Loops that maintain them. The judgment is fractional. The execution is not. We argue that distinction hard because we ran our own company into the wrong side of it first: our agents work from our own governed metrics, and the first time one of those definitions moved without being versioned, what we got back was not a wrong report. It was wrong work.

Why hire a fractional CDO?

Because untrusted data is not an inconvenience. It is a tax you pay in senior hours every month, and a ceiling on everything you intend to automate next.

The reconciliation tax is a real number

Size it before you buy anything. The arithmetic is unflattering and it is the fastest way to get a finance leader's attention.

Count the rebuilds: Three teams each reconstructing the same pipeline or revenue number monthly at roughly 12 hours a cycle (illustrative) is over 400 hours a year (illustrative) — and the output is still three numbers.
Add the time you spend adjudicating: Every leadership discussion that turns into a debate about whose figure is right is a decision deferred a week. That delay costs more than the analyst hours and never gets counted.
Then add the decisions made on the wrong number: This one cannot be cleanly quantified and we will not pretend otherwise — but one mispriced quarter outweighs everything above it.
Compare against the fix: Canonical definitions, one governed pipeline, and automated tests on the metrics that matter retire most of that recurring cost permanently.

First step: Take the metric on your board deck that gets questioned most, and ask two teams to produce it independently. The gap is your baseline.

Your agents will inherit every defect you have

This is why the seat is urgent now rather than eventually, and most companies find out too late.

Bad data does not stay slow anymore: A person reading a wrong number makes one wrong decision. An agent reading a wrong field repeats that error at machine speed across every record it touches.
Agentic work raises the cost of ambiguity: A human analyst silently resolves "which customer table is the real one" a hundred times a week from context. An agent cannot, and will not tell you it guessed.
The quality bar moved with the workload: Data that was adequate for a monthly report is not adequate for a system acting on it hourly.
A pilot that works on sample data and fails on production data is not a model failure: It is a data failure that arrived late, after the budget was spent proving the wrong thing. The blocker was never the model.

Who should not hire a fractional CDO

Companies running on one system: If operations, billing, and customer records live in one platform with a single source of truth, you have a reporting need, not a data leadership need. Hire a good analyst.
Companies that need pipelines built, not decisions made: If the architecture is settled and the work is execution, you want data engineers. We will quote the engineering instead — it is cheaper and usually right.
Companies where data is the product: If you monetize data, that leadership is permanent and internal. Fractional can bridge to that hire but should not be the destination.
Companies unwilling to settle definitions: The hard part here is political, not technical. If the business will not agree what a "customer" is, the engagement produces documentation instead of change.

Our own agents read a governed metric layer, and we built it after we needed it

The argument on this page — that a contested definition stops being a reporting annoyance and becomes an operational defect the moment something automated reads it — is not one we reasoned our way to. We shipped the mistake.

Facet runs its own engineering, content, research, quality review and reporting through Agent Loops under human command, and those loops decide what to work on by reading numbers we generate about our own output. Every piece of our own content is scored on a fixed five-part rubric, zero to one hundred, weighted the same way every time, against a threshold nothing passes without. The definitions of those five parts live in a versioned file rather than in anybody's head, and the score is written into our system of record by a process rather than typed in by whoever remembered.

Each of those design decisions is scar tissue. Definitions were versioned because two scores that used the same word turned out to mean different things. Writing became automatic because a number that lives only in a transcript is a number nothing downstream can act on. And we added a check that fails loudly on a missing score because the failure mode we actually hit was worse than a wrong number — a missing one looked like an unscored item rather than like a broken pipeline, so nothing complained, and the loops kept working from an incomplete picture for longer than we would like to write down.

That is the whole case for this seat, learned at our own expense rather than assembled for a pitch. A person reading a wrong number makes one wrong decision and usually notices. A system reading a wrong number makes the same decision repeatedly, confidently, and silently.

Humans stay in command. Agents inventory, profile, test, and monitor. Your CDO settles what a metric means, and a named Facet engineer owns every change that reaches a production pipeline.
We name the failure modes. Hallucination, scope creep, runaway token cost, and lost context are the reasons governed data matters more in an agentic operation, not less — an agent cannot silently resolve an ambiguity the way an analyst does, and it will not tell you it guessed.
The seat arrives with the builders. Agentic Engineering builds the pipelines the decisions require, AI Managed Services keeps them tested and current, and the Agentic Operating System is what you have once several parts of the business are running on the same governed numbers. A data executive without builders is a very expensive opinion.

Which of the three seats you actually need

Facet fills three fractional executive seats. They are distinct mandates, and the sequencing matters more here than anywhere else on the ladder.

The CDO decides what the numbers mean and who gets to say so: Definitions, lineage, quality, access. This seat.
The CIO decides what you run and what you stop paying for: The application portfolio, vendors, spend, security posture, and the identity and access rules that agents operate inside.
The Chief AI Officer decides where agents go first and what gets canceled: An allocation seat. It is usually the one companies ask for first and the one that stalls hardest when the data underneath is contested — agents inherit whatever quality your data already has, so this seat frequently has to go first.

What a fractional CDO does

The mandate is the decision layer that outlives your tools. Platforms get replaced every few years; definitions, lineage, and ownership survive every migration — which is exactly why they deserve an executive owner rather than a project.

The metric layer: Canonical definitions for the numbers the business runs on — revenue, pipeline, retention, utilization, cost to serve — written down, versioned, and implemented once rather than re-derived per report.
Lineage and quality: Where each number comes from, what transforms it, who owns it, and automated tests that fail loudly when it drifts. This converts trust from a feeling into a status.
Access, privacy, and retention: Who sees what, how long you keep it, and what your regulatory footprint actually requires — HIPAA in healthcare and aesthetics, contractual data terms in B2B, privacy obligations by jurisdiction.
The platform roadmap: What the warehouse, transformation layer, and reporting stack need to become over 18 months, sequenced so each step pays for itself instead of arriving as one large project.

What the loop actually does

Data governance has had a bad reputation for a specific and defensible reason: the documentation and stewardship labor used to cost more than the result appeared to be worth, so the catalog got built once, went stale, and taught everyone that governance is a binder. That arithmetic is what changed. The work underneath it is high-volume, mechanically checkable, and almost always fine — which is why it was miserable for people and is well suited to an Agent Loop.

Source inventory and column profiling, continuously: What tables and feeds exist, what is actually in them, what has changed since last week. A hand-assembled catalog is out of date the month it is published; a maintained one is a live system, which is the only kind anybody opens twice.
Draft definitions and lineage maps: Where a number comes from, what transforms it, and where it forks — traced through the pipelines rather than reconstructed from memory in a workshop. Your executive edits a draft instead of convening a committee to produce one.
Quality tests written at the rate the data changes: Freshness, volume, uniqueness, referential checks, and anomaly detection generated and maintained against your actual pipelines, instead of losing to feature work every sprint. Tests that do not keep up with the schema are tests that stop meaning anything.
Drift and reconciliation monitoring: The same metric computed two ways and compared, so a divergence surfaces as an alert the day it starts rather than as an argument in a board meeting a month later.
Access and retention evidence: Who can see what, what is being kept, and how that compares against the position your executive actually wrote down — gathered continuously, because this is the register that decays quietly and is needed suddenly.

What a person still decides

None of that settles anything. Profiling can tell you that two tables both claim to hold customers; it cannot tell you which one the company means.

What a contested word means, and whose definition loses. Finance counts revenue at invoice, sales counts it at signature, and both are defensible. Somebody has to pick one, tell the other department their number is changing, and hold that line the following quarter when it makes a dashboard look worse.
Who owns a domain, when two departments both think they do. Ownership disputes over customer, product, or pipeline data are authority problems wearing an analytics costume. They are resolved by a person with standing, or they are not resolved.
What the business should stop collecting. Every field retained is a privacy obligation, a storage cost, and something a future agent might read and act on. Deciding to stop is a judgment call about risk appetite that no profiling pass will make for you.
Which numbers are allowed to drive something automated. A metric good enough for a monthly review is not automatically good enough to let a system act on hourly. Somebody has to say which ones have cleared that bar, and accept what happens if they were wrong.
What to tell leadership when the trace is embarrassing. The end-to-end trace of a board metric frequently ends in a spreadsheet on one person's machine. Reporting that plainly, to the people who have been presenting the number, is the least technical and most valuable thing the seat does.

How it works

The first 90 days

Weeks 1–4, trace one metric end to end: We take the number your board argues about and follow it from source system to slide — every join, every manual step, every place it forks. The finding is usually specific, embarrassing, and immediately fixable.
Weeks 5–8, definitions and the trust baseline: Canonical definitions agreed with the departments that own them, a documented lineage map, and automated tests on the top metrics so drift surfaces the day it starts.
Weeks 9–12, roadmap and first build: A sequenced platform roadmap plus one governed pipeline shipped — built on the modern data stack, with the transformation layer documented and tested rather than trapped in saved queries.
Pricing throughout: A fixed monthly fee for the executive seat, quoted before we start, never hourly. Engineering builds are scoped separately so each is weighed on its own return.

What we report

Trust status on the metrics that matter: Which board-level numbers are governed, tested, and reconciled, and which are still contested — a list that should get shorter every month.
Reconciliation hours retired: Recurring analyst and leadership time recovered against the week-one baseline, in hours and in dollars.
Quality incidents and time to detection: How often something breaks and how long it takes to notice. Falling detection time is the clearest proof the tests are real.
Automated thresholds on freshness and volume: A stalled feed or a volume anomaly routes to an owner within the hour, rather than being discovered by whoever opens the dashboard on Monday.

Which of your numbers is an agent allowed to act on, and who decided that? Book the 4-Week Data Trust Audit — fixed fee, run against one board-level metric in your own systems rather than a maturity questionnaire. You leave with the end-to-end trace, canonical definitions agreed with the teams that own them, the reconciliation hours you are currently paying, and an honest answer about whether this is a leadership problem or a reporting one.

Fractional CDO FAQ

What does a fractional Chief Data Officer do? A fractional Chief Data Officer owns the decisions that determine whether your data can be trusted and used — definitions, lineage, quality standards, access and retention rules, and the platform roadmap — part-time and ongoing, for a fixed monthly fee.

They arbitrate, which is the part nobody else can do: Getting finance, sales, and operations to accept one definition of a shared metric is an authority problem, not an analytics problem.
They set the standard the builders build to: Engineers need a specification. The seat produces it and defends it.
They sequence the platform: What gets built, in what order, so each step pays for itself instead of arriving as one project you must justify at once.

Can data leadership really be part-time? The governance and architecture decisions can be, because they are periodic and high-consequence. Running pipelines cannot, which is why a fractional seat only works with a delivery team behind it.

The decision work is genuinely intermittent: Definitions get settled, standards get set, roadmaps get sequenced. Dense weeks, not a permanent daily load.
The execution work is continuous, and it is a team: Pipelines, modeling, testing, and monitoring are ongoing engineering. Facet supplies that team so decisions become running systems.
A fractional executive without delivery is a document generator: You get a catalog, a stewardship model, and a maturity assessment nobody builds against. That is this market's common failure, and it is worth naming before you buy.

Do we need a fractional CDO or just a data engineer? If the architecture and definitions are already settled and the work is building, hire the engineer — it costs less and it is the right answer more often than we are supposed to admit.

Engineers need a specification: If two departments disagree about what a customer is, an engineer implements one of the two and you have the same argument with more infrastructure.
The tell is whether your numbers are contested: Contested numbers are a leadership problem. Slow numbers are an engineering problem.
Many engagements are both, sequenced: A short executive engagement to settle the decision layer, then engineering to build it, then a light cadence to keep it honest.

What does this produce that we can point at? Governed metrics with tested pipelines behind them, a documented lineage map, written access and retention positions, and a sequenced platform roadmap — plus the reconciliation hours you stop paying every month.

The artifacts are operational, not archival: Definitions are implemented in the transformation layer, not filed in a wiki, and tests fail loudly on a schedule.
The trust list is the headline: A short, honest register of which numbers are governed and which are not, reviewed monthly with leadership.
You keep all of it: Models, tests, documentation, and roadmap live in your repositories and your warehouse. The engagement ends; the capability does not.

Settle the Numbers, Then Automate

Every automation, every agent, and every AI initiative you are planning runs on top of your data. If the numbers underneath are contested, you are not building leverage — you are scaling an argument.

Start with a 4-Week Data Trust Audit: a fixed-fee engagement that traces one board-level metric from source system to slide, documents every fork and manual step along the way, delivers canonical definitions agreed with the teams that own them, and quantifies the reconciliation hours your company currently spends rebuilding numbers it already has. You keep the lineage map, the definitions, and the costing whether or not anything else happens.

The trace also settles whether you need this seat at all, and it regularly says you do not. If the number resolves cleanly to one system with one owner, what you have is a reporting job and the right hire is a good analyst — not an executive, not us, and not a platform project. We will write that down and hand you the trace. It costs us the engagement, which is precisely why the answer is worth something when it comes back the other way.

Facet has built analytics and data infrastructure since 2014 — modern data stack, documented transformation layers, and reporting your team can defend — recognized by CIOReview as a Top 10 Digital Transformation Services Company.

One thing to weigh before you decide. A definition is close to the cheapest artifact on this page to fix: it is a conversation, a document, and a few tests. It is also the most expensive thing to leave wrong, and the gap between those two is not fixed — it is set by how many things are reading the number. One analyst reading it wrong produces one bad slide. A dashboard, two departments, and a loop reading it wrong produce a company that is confidently coordinated around the same error. The mechanism is the number of consumers, not the passage of time, which is why the cost of settling this rises every time you point something new at your data and never falls on its own.

Book the 4-Week Data Trust Audit, and find out what one number is really costing you.