The job posting said AI Content Specialist. The work turned out to be prompt writing with a content calendar attached. Six months later the team had far more drafts, a brand voice that drifted every other week, and no way to answer the question the CFO kept asking: which of this actually moved pipeline?
That is the most common hiring mistake in marketing right now. Not the wrong person. The wrong job description.
Content engineering is a systems discipline. It has more in common with platform work than with copywriting. Hire for it as if it were a writing role with a chatbot bolted on and you get a person who can generate a decent draft on demand and nothing else. No repeatability. No measurement. No way to tell whether last month was better than this month.
The role exists. It gets hired by accident, under a title that hides what it does.
Read those four numbers together. Adoption is near total. Productivity gains are real and self-reported. Trust is still broken for a third of the people using it. That gap is an ownership problem, not a model problem. Everyone is generating. Almost nobody is accountable for the output, which is the job the AI content engineer is supposed to do.
Six layers, one owner
Content engineering is a loop: source, knowledge, generation, evaluation, distribution, measurement. Each layer produces an artifact. Each artifact can be inspected, versioned, and improved. When no one owns a layer, the layer does not disappear. It just fails quietly and gets blamed on the model.
Here is the map. If you are hiring, this is the scope of the job. If you are the hire, this is what you should be measured on. Writing shows up once, inside generation, and even there the artifact is the spec, not the draft.

That is the part most job descriptions get backwards. The output of this role is a pipeline that reliably produces content of a known quality, plus a record of what that quality cost.
The knowledge layer gets underestimated because it looks like admin work. It is a living index of approved claims, current offers, and usable proof, and it has to be versioned the way code is. Most teams keep it in a shared drive and wonder why the model keeps quoting a product description from fourteen months ago. Keep that index in a structured doc store where every entry has an owner and a date, something like Notion or a repo-backed markdown folder. That is the difference between retrieval you can trust and retrieval you have to re-check by hand.
The evaluation layer is where the role earns its salary. The scarce skill is deciding what good looks like before the work exists, then encoding that decision so it runs without you. Software teams call this testing, and it is the piece most content teams still skip. For the concrete version of that gate, the five-gate quality pipeline walks through it step by step.
Where the role should sit
This decision gets made casually and costs a year. The role needs three things at once: authority over content standards, access to the data and systems that hold the knowledge, and the ability to instrument results. No single team owns all three, so pick the home that protects the hardest one to borrow.
| Reporting line | What you get | Failure mode |
|---|---|---|
| Content team | Fast adoption, strong editorial standards | No data access. Evaluation turns into opinion, and the knowledge index rots. |
| Content, dotted line to data | Editorial judgment with real instrumentation | Requires a leader who will defend the dotted line when deadlines hit. |
| Growth or marketing ops | Instrumentation and channel plumbing | Volume becomes the scoreboard. Quality gets traded for throughput. |
| RevOps or data | Clean measurement and governance | Slow. Content gets treated as a reporting problem instead of a product. |
The workable answer for most mid-size teams is the highlighted row: inside content, with a dotted line to whoever owns the data warehouse and the marketing stack. That keeps editorial judgment where it belongs while giving the role the keys it needs to build evals and read results.
What does not work is parking the role under a content manager who is measured on publishing volume. The moment the quarter gets tight, the eval layer is the first thing cut, because it is the only part of the job that produces nothing visible this week. Six weeks later you are back to reading every draft by hand and calling that quality control.
Five signals that separate a builder from an operator
Resumes in this field are useless. Everyone lists the same tools, and tool fluency decays in months. Interview for artifacts and decisions instead. These five questions do most of the work.
-
1Ask for a system, not a portfolioRequest one artifact they built that other people use: a brief template, a retrieval setup, a scoring rubric, a publishing job. Portfolios show output. Systems show whether they can make other people’s output better.
-
2Make them design an eval on the spotHand them a real brief from last quarter and ask what a passing output would have to contain and how they would score twenty samples of it. Strong candidates write criteria before they write prompts. Weak ones jump straight to the prompt.
-
3Probe retrieval hygieneAsk what goes into the knowledge index, what gets excluded, and how a claim gets retired when the product changes. The answer tells you whether they have run a system that stayed accurate for more than a quarter.
-
4Ask about cost and latency per assetAnyone who has operated at volume knows what a piece costs to produce and how long it takes end to end. Anyone who has only demoed does not. This question separates the two in about ninety seconds.
-
5Ask what they would shut offEvery real builder has killed something: a prompt that kept failing, a channel that never converted, a source that poisoned the index. If nothing has ever been turned off, nothing was ever measured.
One more signal worth watching in the room: how they talk about the people downstream. Builders describe the editor, the sales team, and the buyer as users of the system, with specific complaints. Operators describe them as audiences. That difference shows up in the work within a month.
Three deliverables, in order
Give the new hire a quarter and three artifacts. Not a content calendar.
Weeks 1 to 3: a brief spec. One page that defines what a brief must contain before anything gets written: audience, job to be done, the claim being made, the proof behind it, the channel shape, and the success measure. Everything downstream inherits this document, so it comes first. The feedback loop that keeps the system honest only works if the input is consistent enough to compare against.
Weeks 4 to 7: a 30-item eval set. Real briefs, real outputs, human-scored against a rubric. Thirty is enough to catch systematic failure and small enough to finish. This is the artifact that makes quality a number instead of an argument, and it is the one most teams never build.
Weeks 8 to 12: one measured loop. Pick a single channel and instrument it end to end: source to draft to publish to citation or click to pipeline. One loop that closes beats six that half-work. Once the loop is measurable, expanding it is a resourcing decision instead of a leap of faith.
Questions I get about this role
Do you need an engineer to do this job?
No. You need someone comfortable with structured data, APIs, and instrumentation, which is a different bar. The last three people I have seen do this well came from content strategy, marketing ops, and technical SEO. None of them wrote production code. All of them could read a data model and argue with it.
Can a content marketer grow into it?
Yes, and that is usually the better hire. The editorial judgment is harder to teach than the tooling. A strong content marketer who is curious about systems will beat a strong engineer who has never had to defend a claim to an editor. The transition takes about six months of deliberate practice on evals and retrieval.
What if we cannot afford a dedicated hire?
Then do not hire a junior version of the role. Assign the six layers to existing people, name one owner for the loop, and start with the eval set. A shared owner with a real rubric beats a dedicated hire with no mandate. The role can be fractional. The accountability cannot.




