Buried inside Adobe's general availability announcement for CX Enterprise Coworker is a phrase worth reading twice. Anjul Bhambhri, Adobe's SVP of engineering for Customer Experience Orchestration, describes the product as "agentic AI that is grounded in brand, customer and channel intelligence." Grounded. Not informed. Not aware of. Grounded, as in constrained by, as in unable to act outside of.

That single verb is doing more work than the rest of the press release. It signals that Adobe is treating brand judgment as an operational layer, something the agent must consult and obey at runtime, not a PDF the team reviews on Fridays.

I keep seeing teams try to bolt an agent onto a marketing stack and then act surprised when the output drifts. My read is simpler. They shipped an agent without shipping a taste contract, and the agent did what agents do without one: optimized against whatever signal happened to be legible.

What Adobe actually shipped

Strip the marketing language off Coworker and you get a specific product claim. It positions itself as a "central intelligence layer" that coordinates AI agents and workflows across analytics, content creation, and journey orchestration for the enterprise applications used by, in Adobe's phrasing, more than 20,000 global brands. It is headless. It runs on Model Context Protocol and Agent-to-Agent frameworks. It talks to Anthropic, OpenAI, Google, Microsoft, and AWS models.

So what is the actual innovation here? It is not the model, and it is not the orchestration. Both are commodities now. The innovation is that Adobe is trying to make brand judgment machine-readable and enforceable at decision time. The Coworker interface is described as "custom for every brand, based on the nuances of their business and overall objectives." That is a polite way of saying the product only works if the brand can articulate its taste with enough precision that an agent can execute against it.

Most brands cannot. That is the interesting problem.

The taste contract, defined

A taste contract is the machine-readable expression of what a brand will and will not do, delivered at the moment an agent is about to act. It is not a brand book. A brand book is a document humans consult before deciding. A taste contract is a set of constraints an agent consults before deploying.

The useful way to think about it has four parts.

Inputs. What the agent is allowed to know about the customer, the moment, the channel, and the campaign objective. This is where consent, data policy, and audience definition live.

Constraints. What the agent must not do. This is the hardest part and the most neglected. Voice attributes, forbidden phrases, competitive comparisons, price commitments, claims that require legal review, tone in a crisis moment. The constraint set is where taste actually lives, because taste is mostly refusal.

Deployment surface. Which channels, formats, and customer states the contract applies to, and where authority to act ends and human review begins. A retention offer to a churn-risk enterprise account is not the same surface as a promoted post.

Update cadence. How often the contract is refreshed and who owns the refresh. This is the piece almost no team has thought about, and it is the one that decides whether the whole apparatus stays useful past month three.

A creative brief has none of these properties. It is a narrative document written for a human collaborator who will apply judgment the brief does not specify. Agents cannot do that. When the agent hits ambiguity, it either refuses or improvises, and both failure modes are expensive at scale.

Two implementations, two philosophies

Adobe's approach is ontological. Encode the brand as a structured intelligence layer, ground the agent in it, and enforce at runtime. Salesforce is pushing a similar shape with Agentforce Marketing, unifying data, content, and conversations so agents can act across the customer lifecycle. Same instinct. Same bet. Taste has to become executable or the agents underneath will not be trusted with real work.

The contrast case I keep thinking about comes from a different industry entirely. The New York Times Book Review recently published a ranking of the 50 best thrillers of the twenty-first century. What makes it interesting is not the list. It is the method. The Book Review built the ranking from ballots submitted by writers, booksellers, editors, Bookstagrammers, and librarians. Five different roles, five different rationales for what makes a thriller work, aggregated into one legible artifact.

That is a taste contract too. But it encodes something Adobe's ontology-first approach can miss: dissent. The ballot preserves the fact that a bookseller and a librarian value different things, and it treats that difference as a feature. The final ranking is defensible not because everyone agreed, but because the disagreement was made visible and structured.

Brand systems tend to do the opposite. They collapse dissent into consensus, ship the consensus as guidance, and then wonder why the outputs feel a little dead. When you then hand that consensus to an agent, you get scale plus flatness, which is the specific texture of most AI-generated marketing right now.

The failure mode nobody is pricing yet

Here is the interpretive tension. Legibility is necessary for agents to work, and legibility is exactly what tends to kill the cultural resonance the brand book was written to protect.

The more precisely you encode taste, the more repeatable the outputs. The more repeatable the outputs, the faster the brand starts to sound like a system that has learned to describe itself. Agents will optimize against any constraint set you give them, including a constraint set that describes a brand at a moment in time that has already passed.

This is where the refresh cadence stops being a footnote and becomes the whole game. A taste contract that does not evolve is a taste contract that lies. It says the brand still believes what it believed six months ago. Sometimes that is true. More often, the market moved, a competitor repositioned, a customer segment matured, and the constraint set is now enforcing a stale identity at machine speed across every channel simultaneously.

This is the same problem I explored in Define 'Good' Before You Ship the Agent, extended into the domain of brand taste. Defining good is the entry ticket. Keeping the definition honest is the actual job.

The CMO becomes a legislator

If you take the taste contract seriously as a product primitive, the CMO role changes shape. Creative direction is still there, but it is no longer the center of gravity. The center of gravity is constitutional design: writing the rules that govern what autonomous marketing systems can do on the brand's behalf, and revising those rules on a cadence that matches how fast the market moves.

That is a different job than the one most CMOs were hired to do. It is closer to how a general counsel thinks about policy than how a creative director thinks about a campaign. It rewards precision, dissent management, and the willingness to say what the brand will refuse to do, which is harder than saying what it aspires to.

The operating shift, in concrete terms, has three moves. Name a single owner for the taste contract with authority to approve exceptions in real time, not through a ticket queue. Set a refresh cadence tied to market signals, not calendar quarters. Preserve dissent inside the contract by encoding role-based rationales the way the Book Review's ballot does, so the system knows which constraints are consensus and which are contested. Contested constraints are the ones most likely to need updating first.

This is the same underlying discipline I have argued for in Hire the Phase, Not the Title. The artifact matters more than the org chart. A taste contract with a named owner and a refresh clock is a phase artifact. A brand book on a shared drive is not.

The closing turn

Adobe's product is a bet that brand judgment can be productized. I think that bet is largely right, and the teams that build early taste contracts will pull ahead of the teams that keep treating brand as tacit knowledge held in senior heads. The tacit approach does not scale to agent throughput. It never will.

But the taste contract is not the finish line. It is a living document that will be wrong sometimes, and the discipline is not writing a better first version. It is building the mechanism that lets you correct the contract faster than the market punishes you for the previous one.

The move in agentic marketing is not slower action. It is faster correction when the contract starts to lie.

Sources