Computational Marketing
FeaturesLong read

Who should build marketing's AI agents? The case for marketing engineers over vendor black boxes

Owning the judgment layer, not the plumbing, is what marketing teams should demand.

· 8 min read
Cover illustration for “Who should build marketing's AI agents? The case for marketing engineers over vendor black boxes”
Features · November 19, 2026 · 8 min read · 1,864 words

Marketing’s AI agents should be built around judgment the marketing team owns. That is the case for marketing engineers. For 2027 budgets, the choice is between renting a finished answer and owning the rules, measurement, and control that make the answer useful.

Why 2027 budget season is the moment the build-vs-buy question becomes unavoidable

Agent use in marketing crossed the majority line in 2026. Yet using an agent is not the same as letting it act at scale. Much of the work remains analytical and low-stakes, including reporting and drafting briefs, while decisions that move spending or reach customers still pass through human gates. Adoption has moved faster than delegated authority.

That gap requires governance that determines who can inspect decisions, change the rules, and decide when an agent deserves more freedom. A vendor contract can settle those questions without the marketing team explicitly considering them. If the decision layer stays inside the product, broader adoption can deepen dependence. As 2027 budgets turn experiments into longer commitments, who should actually own the judgment layer in a marketing agent stack?

What "marketing is an engineering discipline" actually means in practice

Treating marketing as an engineering discipline means instrumenting each step, measuring outputs against targets, and using evidence to change the system. Speed alone does not meet that standard. An agent that produces more briefs without a way to judge their quality is guesswork with a faster typewriter. The team must be able to explain what happened and improve what happens next. That requires access to the mechanism.

Three properties make that access useful, namely observability, reversibility, and composability. Observability means the team can see what the system decided. Reversibility means it can undo a decision before its effects build up, while composability means it can swap out one component without rebuilding the whole system. When a black box withholds them, the team must turn to support to address a drop in performance.

Without a target, the team cannot show that a change helped, defend the investment in a budget review, or safely widen an agent’s authority. More activity does not settle those questions. A useful system must connect what the agent does to what the business wants. Otherwise, the team remains dependent on the vendor’s account of performance.

The marketing engineer is the role this discipline produces, a marketer with a builder’s hands rather than a developer placed in marketing or a renamed operations generalist. GrowthPair’s marketing engineer guide distinguishes building new pages, tools, and agents from running the systems already in place. Marketing operations can then take over what has been built. The distinction concerns responsibility. Figma measures its marketing engineer on agents live, hours saved, and revenue driven, so the work is judged by systems in use and results for the business.

Why vendor black boxes underperform owned infrastructure on the metrics that matter to a CFO

A vendor-managed stack can look attractive at procurement because it trades control for speed. Once the cost of the first build has paid for itself, owned infrastructure can do more work without the same per-seat or per-action vendor fees. That changes which costs grow with use and which capabilities the organization keeps after paying the bill.

The strongest counterargument deserves a direct answer. MIT Project NANDA’s findings show that specialized vendors and partnerships have more success getting systems into use than teams that build them in-house. A finance team should weigh that evidence even if it likes the idea of owning the system. Shipping a working system is a real advantage, and an unfinished internal project produces no return. The case for ownership therefore cannot rest on the claim that building always wins.

Projects built in-house fail when teams start before they know what the work takes, how much there is, or how much upkeep it will need. A team can work out those details before it commits. But better project planning cannot make a product show its decisions or let the team swap out components if the product does not allow it. Its limits remain part of the arrangement, even when deployment succeeds.

Teams should own the judgment layer and buy commodity infrastructure. Brand voice rules, funnel qualification logic, and the work of reconciling dashboards with revenue all come from how that business works. Model APIs, retrieval infrastructure, and scheduling plumbing do not need to be invented again. Buying those components is compatible with internal control. The economic test is whether the purchase supplies a replaceable capability or makes the team dependent on someone else’s judgment.

That distinction also changes what a CFO can evaluate. When the team owns the rules and can measure their effects, it has a basis for explaining why spending produces a result. The team needs the vendor to interpret performance because it cannot inspect the system itself. A low initial price does not resolve that loss of leverage.

Who owns marketing's agents: marketing engineers, not vendors or agencies

The judgment layer belongs inside marketing because it depends on brand voice, funnel logic, and competitive context. A marketing engineer joins that context to the ability to build. The role is not simply a new title. It gives one function responsibility for closing the gap between an agent being available and an agent being trusted to act. Neither procurement nor technical deployment alone closes that gap.

GrowthPair describes the role as shipping landing pages, lead magnets such as calculators, and AI agents without waiting on an engineering team. Those builds sit close to the work they serve. An external vendor or agency cannot fully absorb the same internal context, even when it can deliver the technology. The agency model also has an economic tension in execution work. Adding people and billable hours built that model, while agents are useful precisely because they can reduce the time and cost of delivery.

Meghan Gill argues that these building skills will spread across marketing roles within five years, so there will be no need for a separate title. That is a serious alternative. Today, however, most marketers cannot build these systems, and waiting for the skills to diffuse leaves the team paying vendor margins in the meantime. The team must decide where to place accountability now.

A team assessing that accountability can use three ownership questions. If the agent stopped working tomorrow, who would notice? An answer that points only to the vendor’s support queue puts detection outside the organization. Who would fix it? If correction requires a support contract, the team lacks direct leverage to tune its own system.

Who would check the decisions the agent already influenced? Without access to that audit trail, an internal owner has a title but not the means to investigate. These questions require the marketing team to retain the authority and access needed to run an engineering discipline.

How to choose first agent projects: instrumentable, verifiable, measurable, reversible

For a team committed to that ownership, the first project should be instrumentable, verifiable, measurable, and reversible. Impressive is not a sufficient selection criterion. The team needs to observe decisions at each step and check the output against a known standard before it ships. It also needs a numerical target, whether that is time reclaimed, citations earned, or pipeline influenced. Finally, it must be able to undo a bad action before it reaches a customer or a public surface.

A decision log makes a system instrumentable, but it does not establish that the output is correct. Even an output that has been checked can lack a business goal the team can measure. Likewise, getting a result that helps the business does not make it safe to take an action the team cannot stop. The filter makes the first deployment a controlled way to learn, rather than a test of how much risk the organization will tolerate.

Competitive monitoring is the recommended first deployment for most teams because it combines open-ended judgment with low-risk output. Competitors act unpredictably, so a fixed workflow cannot predefine every relevant change. An agent can check competitor sites, pricing pages, ad libraries, launch announcements, and communities on a schedule, then compile a brief with links. Its judgment has a clear job. Because the brief stays internal, a wrong call costs a conversation rather than a customer, and a marketing engineer can ship the system in days.

Segment-level personalization with pre-approved variants is a second-stage deployment. Here, the agent selects which approved message goes to which segment and when. Fixed variants keep content under human control while the selection logic runs autonomously, making outcomes reviewable and the logic reversible. That does not undo a message already seen. The approval boundary belongs before delivery.

GEO and AEO content pipelines offer a third deployment, with agents researching, structuring, and drafting content for AI answer engines. Humans retain the claims review gate before publication. The useful distinction is not whether a model drafted the page, but whether it contains something verifiable, original, and worth extracting. Production alone is not the engineering asset. A marketing engineer can also own the measurement layer, tracking whether ChatGPT, Claude, Gemini, and Perplexity name and cite the brand, then feeding that signal into content decisions.

For first projects, the exclusions follow from the same filter. An agent should not publish publicly without human approval, hide all its decision logic inside a vendor’s system, or begin without success stated as a number. These are boundaries for learning.

What belongs in a 2027 budget: metered usage and a line-by-line bill

A 2027 budget for owned agents should look like an engineering budget, with metered usage, a line-by-line bill, and costs tied to measurable outputs. Model API calls belong there as costs based on use. So do owned retrieval and storage infrastructure, human review time, and measurement tooling for AI citations and answer-engine visibility. Review needs an explicit budget. Its purpose is to keep judgment inside the system and pay for oversight as part of the work.

The corresponding reductions should be equally visible. Per-action vendor fees should shrink where the marketing engineer now owns the work, as should agency retainers for execution agents now handle. Content spending should go toward work that builds the brand’s own citation authority over time. This is not simply a cheaper purchasing model. When the team can see per-action costs, it has another way to tune the system and judge whether doing more is worth paying for.

The CFO case rests on improving unit economics, accumulating measurement knowledge, and preserving reversibility. Owned infrastructure can lower the marginal cost of another run without adding a vendor’s per-action charge, though the team still pays for metered components. Meanwhile, a team that tracks AI citation rates for 12 months builds knowledge of its own about which types of content earn citations from which engines. That history becomes an asset for future decisions. Keeping actions reviewable before they ship also preserves the team’s ability to say no.

Basic safeguards are the floor. The stronger investment is in a system whose measurement becomes more useful with each deployment and whose rules remain available to change. The 2027 decision should be to fund a marketing engineer who buys replaceable infrastructure and owns the judgment and measurement that compound.

More in Features