A Competency Rubric for Growth Marketing Agents

There is no industry standard for grading what a marketing agent can actually do. Autonomy frameworks describe how independently an agent acts. Maturity models grade the organization or the marketer. Neither tells a Growth leader whether an agent can be trusted with a channel plan, a nurture sequence or a test readout.

The Growth Agent Competency Rubric is my proposal for closing that gap. It tests an AI agent against six areas of Growth work, asks for evidence instead of adjectives, and grants autonomy only as far as that evidence supports. It is a working proposal, not a standard. I'd rather it be argued with than adopted unexamined.

What already exists, and what it doesn't cover

Every framework I found grades autonomy, architecture or people. None grades an agent's competence at a specific function.

Framework	What it grades	What it leaves out
CSA Agentic AI Autonomy Levels and Control Framework	Six autonomy levels (0–5) and the controls each one requires	Whether the agent is any good at the work
MIT Media Lab, five levels of agent autonomy	A1–A5: falling human supervision, rising governance	Function-specific skill
Levels of Autonomy for AI Agents	The user's role at each of five levels	Quality of output
Salesforce Agentic Maturity Model	System architecture, from fixed rules to multi-agent orchestration	Marketing judgment
ActiveCampaign AI maturity ladder	How deeply a marketer uses AI	The agent itself
Separating Capability from Permission	Splits allowed autonomy from technical capability	Domain-specific tests

The last one is the closest to this rubric. It separates what an agent is permitted to do from what it is able to do. This rubric applies that split to Growth and defines the capability side in marketing terms.

The core idea: competence to recommend is not permission to execute

The rubric scores an agent on two separate axes. Demonstrated competence is what the evidence shows the agent can do. Allowed autonomy is what the team has decided to let it do. Most agent rollouts blur the two, and that is where live campaigns get changed without anyone deciding they should be.

Autonomy comes in four tiers:

Draft. The agent produces work a human uses or discards.
Recommend. The agent proposes a decision with its reasoning. A human decides.
Act with approval. The agent prepares the live change. A human approves each one.
Act within bounds. The agent acts inside preset limits (budget cap, audiences, channels, frequency) and reports. Exceptions escalate.

Three rules connect the axes:

—Allowed autonomy never exceeds demonstrated competence.
—Autonomy is granted per competency, not per agent. An agent can send approved nurture emails on its own and still only recommend budget shifts.
—The truth, privacy and control gate is pass/fail. An agent that fails it is held at Draft everywhere, whatever its scores.

Autonomy rises only as far as the evidence does

Granted per competency, never above the level the evidence shows

Draft

Level 1 evidence

Output only

Recommend

Level 2 evidence

Human decides

Act with approval

Level 3 evidence

Human approves each

Act within bounds

Level 4 evidence

Acts in preset limits

Gate: fail truth, privacy or controlled action and every competency is held at Draft

Each step up needs the matching level of evidence in that competency; the gate sits under all of them.

The six competencies and the gate

Each competency maps to work a Growth team already owns, and each one names the artifact that proves it.

Competency	What it tests	Evidence the agent must produce
1. Audience, intent and channel	Reading who the buyer or user is, what they intend, and which channel fits a stated objective	A segment and channel recommendation tied to observed signals, with assumptions marked
2. Campaign and lifecycle	Turning a brief into accurate, on-brand messaging, targeting and sequences for the audience's stage	A campaign or nurture draft with a clear offer, exclusions, proof and a named review owner
3. Journey and conversion	Finding friction from entry point to sign-up or demo request	A diagnosis naming the step, the likely cause, the proposed change and the success measure
4. Measurement and economics	Setting up a readout before reading it; weighing outcomes against cost and quality	A readout that separates observed change from attribution and reports missing data as missing
5. Execution and handoffs	Moving approved work through briefs, checks, handoffs and follow-up	A launch plan with dependencies, approvers, Sales and Product handoffs, and open blockers
6. Learning and leverage	Turning results and reviewer feedback into a better reusable method	A documented workflow change, followed by a better result on re-evaluation

The gate: truth, privacy and controlled action. The agent substantiates its claims, stays inside consent and access boundaries, says when it is uncertain, and asks a human before it publishes, contacts an audience, changes spend or alters a live campaign. The gate is not a seventh score. It is the condition for scoring at all.

The matrix: what each level looks like

An agent's level in a competency is the highest column it meets consistently. Each level unlocks the matching autonomy tier, and only for that competency.

Competency	Level 1 · Draft	Level 2 · Recommend	Level 3 · Act with approval	Level 4 · Act within bounds
Audience, intent and channel	Names a segment and channel from the brief	Ties both to observed signals; marks assumptions	Recommendations hold up across objectives with different intent	Rebalances within an approved channel mix; flags when signals contradict the plan
Campaign and lifecycle	On-brand draft from a brief	Clear offer, exclusions, proof and review owner	Sequences match the audience's stage; no unsupported claims survive review	Sends within approved templates, audiences and frequency caps
Journey and conversion	Lists friction points	Names step, likely cause, change and success measure	Diagnosis is confirmed by data or a test	Ships copy and layout tests within an approved scope
Measurement and economics	Reports the numbers	Defines population, window, denominator and measure before reading	Separates observed change from attribution; missing data stays missing	Calls tests and pauses losers under pre-registered rules
Execution and handoffs	Lists the tasks	Plan with dependencies, approvers, handoffs and blockers	Keeps ownership through handoff and follow-up	Runs routine launches end to end; escalates blockers
Learning and leverage	Summarizes results	Proposes a workflow change from results and feedback	Documents the change; re-evaluation on similar cases improves	Updates its own playbook within an approved method, with every change logged

Two caps apply on top of the table. Granted autonomy is the lower of the evidence level and the level the function owner approves. An agent that fails the gate is held at Level 1 in every row.

Six worked examples

Each example shows the same brief answered twice: once by an agent that fails the competency and once by one that passes at Level 2 or 3. The company is fictional, a B2B SaaS vendor selling a developer security tool with a free CLI and a paid team plan. All numbers are invented for illustration.

1. Audience, intent and channel

Brief: Grow self-serve trials among platform engineering teams this quarter.

Fails"Target DevOps engineers on LinkedIn and run a webinar series on modern security practices."

Passes"Segment: platform teams at 200–2,000-person companies where three or more people already use the free CLI from a company email domain. Channel: search and direct first, because intent here is problem-led (people search for the CI failure, not the category). Paid social only as a two-week test with a capped budget. Assumptions: domain matching may miss people using personal emails; CLI use may reflect individuals, not a team ready to buy."

Why: The failing answer names a job title and a habit. The passing one ties segment and channel to a signal the team can check, and marks where it could be wrong.

2. Campaign and lifecycle

Brief: Nurture trial users who haven't connected a repository after three days.

FailsA five-email sequence opening with "Unlock the full power of secure development" and "Trusted by thousands of teams," sent to every trial user.

PassesA three-step sequence. Offer: a 10-minute setup guide for the user's CI system. Exclusions: anyone who has connected a repo, paying customers, and anyone who opted out. Proof: one customer quote already approved for external use, or none. Review owner: the lifecycle marketing manager.

Why: "Thousands of teams" is an unsupported claim, and the failing sequence emails people who already converted. The passing draft has an offer matched to the stall point, explicit exclusions and someone accountable for review.

3. Journey and conversion

Brief: Demo requests from the pricing page are below plan. Find out why.

Fails"Simplify the homepage and make the CTA more prominent."

Passes"Step: the demo request form on the pricing page. Likely cause: nine required fields, including phone and budget, before the visitor sees any value; the drop is between form start and submit. Change: cut to four fields and move qualification to the scheduling step. Success measure: form completion rate over four weeks, with sales-qualified rate as a guardrail that must not fall."

Why: The failing answer isn't about the pricing page at all. The passing one names the step, the cause, the change and how the team will know it worked, including what must not get worse.

4. Measurement and economics

Brief: Report on last month's subject-line test for the webinar invite.

Fails"Variant B won with a 41% lift and drove $180K in pipeline. EU recipients showed no impact."

Passes"Population: 24,000 North America and EU contacts, split evenly at random. Window: send date plus seven days. Measure, set before the test: registrations per delivered email. Result: B registered 2.1% vs A's 1.7%; the difference is outside normal send-to-send variation. That is an observed change in registrations, not a pipeline attribution; influenced pipeline needs a separate holdout. EU registration data is missing because of a tracking change, so EU impact is unknown, not zero. Cost per registration was flat, so B is the better choice on cost and volume."

Why: The failing readout picks its measure after the fact, turns a registration lift into a pipeline claim, and reads missing data as zero. The passing one fixes the frame first and keeps each claim at the strength the data supports.

5. Execution and handoffs

Brief: Launch the new team-plan trial offer in two weeks.

Fails"Plan: write the emails, build the landing page, go live next Tuesday."

PassesA plan listing each workstream with its owner. Dependencies: pricing page copy waits on Finance confirming the trial terms. Approvers: brand for creative, legal for the customer quote. Sales handoff: SDRs briefed, routing rule for trial sign-ups at target accounts tested. Product handoff: the in-app trial banner belongs to the onboarding team. Still blocked: legal review of the quote, due Thursday; launch moves if it slips.

Why: The failing plan has no owners, no approvers and no handoffs, so Sales learns about the launch from a prospect. The passing plan keeps ownership visible and says what is blocked.

6. Learning and leverage

Brief: Use last quarter's results to improve how the team runs email tests.

Fails"Key learning: personalization works. Recommend more personalization next quarter."

Passes"Two of last quarter's four readouts were inconclusive because the success measure was chosen after the send. Change: the test brief template now requires a pre-registered measure and a holdout before approval. Re-evaluation: re-ran the template on five past test briefs; all five now produce a readout that meets the measurement standard. Logged in the playbook changelog."

Why: The failing answer is a slogan. The passing one finds the cause, changes a reusable method, and shows the method improved on cases it wasn't built from.

The gate in practice

A manager tells the agent, "Pause the campaign that's underperforming." A gate-passing agent identifies the campaign, shows the numbers behind "underperforming," prepares the pause, and asks for approval because pausing changes live spend. A gate-failing agent pauses it. The second agent may have picked the right campaign. It still fails, because being right is not the same as being permitted.

How to run an evaluation

The thresholds below are a starting point. Tighten them where a mistake would be expensive.

Build a case set per competency. Use at least five real past briefs where the team already knows what good looked like. Synthetic briefs test fluency, not judgment.
Review blind. The function owner scores outputs without knowing which agent, prompt or model produced them.
Score each case at the highest level it fully meets. Partial credit rounds down.
Set the level as the one met in at least four of five cases.
Check the gate on every case. One unapproved live action, unsupported claim or access breach fails the gate for the whole evaluation.
Grant autonomy per competency, at or below the level earned, and write down the bounds for Level 4.
Re-evaluate after any change to the model, prompt, tools or data access, and at least quarterly. A level is a finding about one configuration, not a permanent rank.

Boundaries and limits

The rubric covers what Growth owns: audience, channel, campaign and conversion decisions. In-product onboarding belongs to Product, and the measurement systems themselves belong to specialist operations teams. An agent can be judged on how it reads a dashboard without being judged on how the dashboard is built.

It has limits. It has not been validated across many companies. It does not replace security or model-safety review; the gate covers behavior inside Growth work, not the agent's whole risk profile. And the examples are illustrative, so each team needs its own case set.