Why Agentic Intelligence, Not Workflow Automation: The Architecture Choice Behind groa°
By Polly Barnfield, OBE, CEO of Maybe*
Most "AI agents" in ecommerce are workflows in a new costume. Here is why retention demands something genuinely different, and how we are building it with groa°.
There is a sentence in Chapter 7 of Retention-First Growth® that should be printed on a poster in every ecommerce ops team:
Most AI agents in the market today are better understood as advanced automations. They trigger when specific conditions are met, follow predefined workflows, and rely on thresholds and timing set by humans.
That description fits a great deal of what is currently sold as "agentic AI" in retail. It does not fit what retention actually needs. And the difference is not a marketing distinction. It is an architectural one that determines whether the system can hold under the kind of pressure top-decile performance requires.
This is the choice every client we profile in Scaling Up With Agents has to make. Workflow, or agent. The wrong answer is cheaper to build. The right answer is the only one that scales.
The workflow problem
A workflow-based system works like this. Customer hits a defined trigger. They abandon a cart, cross a 30-day inactivity threshold, lapse past a winback window. The system fires the matching sequence. Done.
This works for narrow, well-defined moments. It breaks for retention because retention is not a sequence of moments. It is a continuously evolving state across five orbits, where the right intervention depends on where the customer is now, how they got there, what their energy profile looks like, and how much margin is still recoverable.
A workflow cannot answer those questions. It can only check whether a trigger has fired. By the time a high-value customer has crossed a static dormancy threshold, the recovery window Emma describes, typically 24 to 48 hours after the first decay signals appear, has already closed. Manual review cycles cannot keep up. Workflow automations cannot detect what they were not pre-configured to look for.
This is the structural failure mode the book calls decision latency. And it is the reason retention-first execution has historically required 70 hours per brand per month of human attention. Humans were the agentic layer.
What agentic intelligence actually means
Agentic intelligence, the genuine kind and not the relabelled-chatbot kind, has four properties that workflows do not.
It observes continuously rather than waiting for triggers. It is reading behavioural signals all the time, against cohort baselines, weighted by confidence.
It reasons about state rather than firing on conditions. The question is never just "did event X happen" but "given everything we know about this customer's orbit, velocity, and energy, what is the right response?"
It takes goal-directed action within guardrails. Goals are set by humans. Preserve Loyalty-orbit retention, protect deliverability, keep VIP churn below threshold. The path to those goals is determined by the agent.
It learns from outcomes without drifting. Each intervention generates response data that refines the next decision, but only within bounded policy parameters that the human team has set.
These four properties are what make an agent an agent. Anything missing one of them is a workflow with better marketing.
How we are building this with groa°
The groa° platform on Maybe* is structured around these four properties from the foundation up.
Continuous observation is handled by a unified identity layer that pulls from Shopify, Klaviyo, and the loyalty and review stack in close to real time, computing each customer's Customer Energy Profile™ as a live state rather than a nightly batch.
Reasoning about state is where the orbit governance logic lives. Each of the five Retention-First Growth® orbits has its own decision policy, encoded from Emma's five years of implementation work. Customers are evaluated against orbit-specific velocity and energy benchmarks, not against a single global churn score.
Goal-directed action sits in the intervention layer. The agents do not just flag at-risk customers. They prescribe the response. Which channel, which content tone, which incentive structure, which suppression rule. All within the guardrails the brand and the groa° team have set.
Learning from outcomes is the policy layer. Response and non-response signals feed back into the system to adjust prioritisation, timing, and content weighting over time, while audit logs preserve the ability to review every change.
The non-negotiable: governance
The thing that distinguishes a serious agentic platform from a hype project is governance. Gartner's own analysis suggests over 40% of agentic AI projects will be cancelled by 2027, usually because someone built autonomy without bounds and watched it drift.
The Maybe* approach to building groa°, and to every client we work with in this series, is governance-first by default. Human judgement is encoded upfront. The orbit logic, the economic thresholds, the brand voice constraints, the deliverability guardrails. Then the agents operate within that envelope. The result is autonomy that scales human expertise rather than replacing it.
Next post: a worked example. We will walk through what happens, end to end, when a Loyalty-orbit customer misses their expected repurchase window, from signal detection through intervention to learning, and show how the architecture described above plays out in a real retention scenario.
Emma's full architectural argument is in Retention-First Growth. Read it here.
If you are about to buy "agentic AI" that is really a workflow in disguise, book a call before you commit.
