The word is being abused
In 2026, almost every CMS, MarTech platform, and customer engagement tool has shipped something labeled "agentic." HubSpot has agents. Salesforce has agents. Adobe has agents. Sitecore has agents. Optimizely has agents. The category is now broad enough to be approximately meaningless.
Our team has spent the last 18 months evaluating these capabilities at client engagements. The pattern is that roughly 80 percent of the features marketed as "agentic" are not agentic in any operationally meaningful sense. They are scripted automations — sometimes useful, sometimes not — with an AI-generated UI wrapper that makes them look agent-like. The remaining 20 percent are genuinely agentic, and they have operational characteristics that buyers are usually not warned about.
This piece is about how to tell which is which, and what to do about each.
The four questions
The test we use is four questions. A feature has to answer "yes" to all four to be agentic in a sense that matters operationally.
1. Can the feature take an action without a human explicitly approving each instance? If the answer is no, it is automation with AI suggestions. Useful, but not agentic. A genuine agent decides and acts within its authority envelope.
2. Can the feature observe its own action and adjust the next action based on what it observed? If the answer is no, it is a scripted automation that happens to use a language model. Genuine agents close the loop — observation feeds the next decision.
3. Can the feature explain why it took a specific action in a way that lets you audit and reproduce the decision? If the answer is no, the feature is a black box wrapped in confidence. You cannot put it in production responsibly.
4. Can the feature be undone, and can the team understand what it would take to prevent the action next time? If the answer is no, the feature is fragile in production — every misstep requires a manual recovery and every recovery requires hypothesis-testing about what the agent was doing.
Most "agentic" features sold to enterprise CMS teams in 2026 fail at least one of these four. The most common failure is question three — the feature cannot explain itself, which means the brand cannot audit its own digital surface, which means the brand cannot ship the feature into production without unacceptable governance risk.
What genuine agentic features actually require
For the 20 percent of features that pass all four questions, the next conversation is about whether the brand is ready to operate them. Genuine agentic features have specific operational requirements that the vendor sales conversation rarely surfaces.
A monitoring substrate. Genuine agents take actions. Actions in production have consequences. You need to see, in something close to real time, what the agent did and what changed as a result. Most enterprise CMS environments do not have monitoring of this kind out of the box. Building it is a 4-8 week effort that does not get included in the vendor TCO.
An authority envelope. What can the agent do without human approval, and what requires approval? The envelope has to be defined explicitly and reviewed regularly. Brands that deploy agents without an explicit envelope tend to discover the agent's authority by accident, usually during an incident.
A rollback path. Every agent action should be reversible without engineering intervention. If a content-modification agent edits a page and the edit is wrong, the team should be able to roll it back from the CMS UI, not from a database query. Most deployed agentic features do not meet this bar in 2026.
A model evaluation cadence. The model behind the agent will drift. Vendor updates, fine-tuning changes, prompt revisions. Without an evaluation cadence, the team cannot tell when the agent's behavior has changed in ways that matter. The cadence does not have to be elaborate — a quarterly sample-and-review on a fixed eval set works fine — but it has to exist.
Brands that deploy agentic features without these four things in place end up with the worst of both worlds: the features are genuinely making decisions in production, but the team cannot see them, control them, undo them, or maintain them. The vendor will not warn you about this. The vendor sales conversation is incented to close the deal, not to explain the operational substrate the deal implies.
What to do about the other 80 percent
The features that fail one or more of the four questions are not necessarily worthless. Many of them are useful — they are just not what the marketing says they are. The right framing is to treat them as scripted automations with AI assistance, and to deploy them as you would any scripted automation.
A "writes blog post drafts" feature is useful even though it is not agentic. A "suggests product tagging" feature is useful. A "drafts customer service responses" feature is useful. None of these decides anything in production without human approval. All of them speed up specific workflows. Buy them on those grounds, deploy them with the same governance as any other tool, and call them what they are.
The harm in the agentic framing is that it confuses the buying decision. A brand that thinks it is buying an agent will budget and govern accordingly. When the feature turns out to be a scripted automation, the brand has overpaid and over-governed. A brand that thinks it is buying a scripted automation will buy it for what it is worth and deploy it appropriately.
The test is not whether the feature uses AI. The test is whether the feature decides.
What we are telling clients in 2026
Across our active engagements, the recommendation our team is making this year is consistent enough to summarize.
First: do not deploy agentic features in production until the four-question test passes. Pilots and sandboxes are fine. Production deployment requires the operational substrate to be in place.
Second: budget the operational substrate before budgeting the agentic capabilities. Monitoring, authority envelope, rollback path, evaluation cadence. These are the bottleneck on AI value capture in 2026, not the model capability itself.
Third: match the deployment ambition to the brand's tolerance for incident risk. Brands that have not had a digital incident in two years are usually under-prepared for what happens when an agent takes a wrong action. Pilot first in low-stakes surfaces. Earn the trust to deploy on higher-stakes surfaces.
There is real value in agentic capabilities. The brands that capture it in 2026 will be the ones whose engineering and content teams treat the work seriously — and whose CMOs do not let the vendor sales cycle compress the operational work that the substrate requires. If you are evaluating an agentic capability right now and want a second perspective, we are happy to walk it with you.