Ilyas Kurklu, Co-founder and CEO of Replenit – Interview Series

Ilyas Kurklu, Co-founder and CEO of Replenit, is an entrepreneur and retail technology executive focused on applying artificial intelligence to customer decision-making and engagement. At Replenit, he leads strategy, product development, and partnerships, building on more than a decade of experience helping scale enterprise retail engagement businesses from approximately $10 million to $2 billion in revenue while working across more than 300 retailers. His career also includes advisory and business development roles, as well as six years as a team lead and weather forecaster at Türkiye’s General Directorate of Meteorology, where he worked with real-time data, radar, satellite imagery, and forecasting systems.

Replenit is an AI decisioning platform designed for retail and e-commerce, providing a layer between a retailer’s data infrastructure and its existing customer engagement systems. Its AI Decision Engine combines customer, product, behavioral, and other signals to determine the next best action for individual customers in real time, while its Maestro AI CRM Manager can autonomously manage workflows including replenishment, cross-selling, engagement, churn prevention, and win-back. Rather than requiring retailers to replace their existing technology stack, Replenit is designed to integrate with current commerce, marketing, and data platforms and turn its decisions into execution-ready communications and actions. The company says its technology is being used by retailers including L’Occitane, Mumzworld, e-bebek, Beko, and Faith In Nature.

You spent years helping scale enterprise retail technology before founding Replenit with five co-founders. What did you repeatedly see retailers struggling with that convinced you there was a missing decision-making layer in the technology stack, and how did that insight lead to Replenit?

Working with hundreds of retailers across different markets, we kept hearing and seeing the same problem: retailers had become very good at collecting data and very good at execution, but the intelligence in between was still missing. They had CDPs, warehouses, CRM platforms and marketing automation systems, yet the decision — what should actually happen next for this individual customer? — was still guesswork based on static rules, broad segments, campaign calendars and human judgment. The stack and the team were both effectively workarounds for the fact that nobody could reason about every customer individually at scale. Replenit is that missing reasoning and decision layer. We’ve built an engine that can interpret customer and product context at the individual level, decide the appropriate next commercial action, and then use the retailer’s existing systems to execute it.

Replenit argues that retailers already have plenty of customer data and plenty of tools for executing campaigns, but lack intelligence that can decide what should happen next. Why haven’t Customer Data Platforms (CDPs), Customer Relationship Management (CRM) platforms, and marketing automation systems been able to solve that problem themselves?

Those systems were built for different jobs. A CDP can unify customer data, but unifying data is not the same as deciding what to do with it. CRM and marketing automation platforms are very good execution layers. They can send, orchestrate and optimize communications once somebody has defined the specific audience, workflow or campaign.

Even as those platforms add predictive scores, generative AI and send-time optimization, the fundamental unit of work is often still the campaign, journey, audience or segment. The AI may optimize something a marketer has already chosen to do. It does not necessarily decide whether anything should happen at all, which lifecycle objective matters most for that particular customer, which product is relevant, why now, in what tone, and toward which commercial outcome. That is the distinction we make between a score and a decision. Prediction can rank likelihood. A decision requires reasoning across context and alternatives and then committing to an action. Replenit sits above the existing stack to provide that layer rather than asking retailers to replace the systems they already have.

A core part of Replenit’s approach is applying Theory of Mind to infer a customer’s intent, context, and likely future behavior. How do you translate a concept from cognitive science into an AI system that can make reliable commercial decisions, and how do you validate whether the system has correctly interpreted a customer’s intent?

We translate Theory of Mind into a reasoning layer that looks beyond what a customer did and asks why that behaviour may have happened, what it signals about intent, and what it means in context. Rather than relying only on historical patterns or product similarity, the system reasons at the level of each individual customer-product relationship and uses that context to determine what should happen next.

That reasoning can include understanding the likely motivation behind a purchase, how products relate to one another, what need the customer is trying to fulfil, and whether the timing is right for a particular action. Replenit then moves beyond simply suggesting an option: it makes a decision and prepares that decision for execution through the retailer’s existing systems.

The validation happens through outcomes. Once a decision is made, Replenit tracks what actually happened and feeds that result back into future decisions for that customer. It also continuously evaluates which of its skills and decision-making approaches are delivering results, scaling those that perform well and deprioritising those that do not. In that sense, the system gets better at knowing the right move for the right customer at the right time because it is learning from real outcomes, rather than simply running the same logic repeatedly.

Most personalization systems are ultimately prediction engines, estimating what someone is likely to click or buy. Replenit emphasizes moving from prediction to decision-making. From an AI architecture perspective, what changes when the model must actually choose the next action rather than simply produce a probability or recommendation?

At Replenit, there are four important parts to that architecture. First is Golden Memory: living records of the customer, product and brand that give the system context beyond an isolated transaction. Where information is incomplete, Replenit can also enrich the available data and generate synthetic context to help close those gaps.

Second are more than 100 industry skills. These are specialized reasoning capabilities for particular categories and situations—a cosmetics routine builder for beauty retailers or an interior architect for furniture and DIY companies, for instance. Alongside those skills, 15 supportive checks can run in parallel around a decision, including substitution, compatibility, duplicate and suppression checks.

Third is workflow ownership. Rather than producing a recommendation and stopping there, Maestro can reason across lifecycle workflows such as replenishment, cross-sell, engagement, win-back, churn, promotion and substitution. Those workflows can operate continuously and in parallel for an individual customer, with Maestro determining which action is appropriate at that particular moment.

Finally, we package that into what we call a Golden Decision Event: a logged, execution-ready decision containing the action and its reasoning, which can be passed directly into the retailer’s CRM, marketing automation, ecommerce, app or data systems across more than 120 platforms.

So the difference is that we are moving from a model that predicts what is likely to happen to a system that can remember the context, bring in the right expertise, evaluate the available actions, commit to one decision and make it executable.

Maestro can autonomously decide whether an individual customer should receive a replenishment reminder, cross-sell, promotion, engagement message, win-back attempt, or potentially no intervention at all. How does the system weigh these competing actions and determine what is genuinely the best decision for that customer at that moment?

The starting point is the customer and the commercial objective. The retailer sets the strategy, priorities and guardrails. Maestro then reasons across the available customer, product and workflow context to determine what should happen next.
That can include deciding which commercial objective should be prioritized, which workflow is relevant, which skill or combination of skills should be applied, which products make sense, and when the action should happen. Retailers can also define priorities around metrics such as revenue, purchase frequency, average order value, margin protection or churn prevention.

Our Theory-of-Mind reasoning adds the individual context: what the behaviour appears to mean, what the customer may actually need and why this may or may not be the right moment. Our specialized skills and the supportive checks can run alongside that reasoning. When the best decision is not to intervene, Maestro doesn’t. Once Maestro reaches a conclusion, it commits the decision and its reasoning as an execution-ready event rather than simply returning a menu of recommendations to a marketer.

Replenit uses a swarm architecture in which multiple specialized AI agents can operate in parallel. What happens when several agents are reasoning about the same customer at the same time, and how do you prevent conflicting recommendations or actions from being executed?

Replenit’s swarm engine doesn’t create agent conflicts because it isn’t about parallelizing customer decisions; it is about running the components of a single decision in parallel. When Maestro reasons about a customer, it activates multiple specialized agents simultaneously: a gating network routes the decision to the right models, a reasoning LLM evaluates customer context and behaviour, enrichment agents synthesize missing customer or product data in real time, and specialized models evaluate product-customer relationships and fit.

Instead of running these sequentially, which would be slower and more expensive, they operate in parallel on the same customer state and then converge on one unified decision about what should happen next. There is no conflict because they are not competing to select different actions; each contributes a different dimension to the same outcome.

For example, enrichment might surface missing customer or product signals, the reasoning layer interprets purchase history and the commercial objective, product-relationship models identify the anchor product and viable substitutes, and scoring weighs the potential business impact. All of that happens in parallel, but produces one result. This level of parallelization is an important architectural advantage. It is difficult to retrofit into engines built as single-tool wrappers, whereas Replenit engineered it into the core from the beginning.

Persistent memory is another important part of the platform, allowing decisions to build on what has happened previously with each customer. How important is long-term memory to genuinely autonomous AI agents, and what safeguards are needed when that memory involves potentially sensitive customer behavior?

Memory is foundational to autonomous decision-making because context is dynamic. When Maestro sees a first baby product purchase, it remembers that the household context has shifted, and future decisions should reflect the parent lifecycle. When it sees initial skincare purchases, it remembers that this customer is building a routine. When purchase patterns emerge around fitness or wellness, it captures the habit being formed.

Each purchase updates what the system understands about that customer—not just demographics, but what drives them, what behaviors they’re establishing, and what life moments are unfolding—so the next decision isn’t generic, it’s rooted in the customer’s actual context. That compounding memory is what makes decisions smarter.

The safeguard is straightforward: we retain purchase events and the contextual patterns that emerge from them—lifecycle signals, habit formation, purchasing drivers, and category affinity—with strict retention policies and pseudonymization where possible. The data scope remains tight. We track what happened and what context shifted, nothing more. Without these guardrails, persistent memory becomes a liability, not an asset.

Replenit is designed to sit on top of the technology retailers already use rather than requiring them to replace their existing stack. What have you learned from deploying with more than 30 enterprise retailers about the biggest technical or organizational barriers to giving AI meaningful decision-making authority?

The clearest technical problem we have encountered is not a shortage of data but the condition of that data. Enterprise retailers often have a great deal of data that’s fragmented across systems, inconsistent or otherwise incomplete. That is why enrichment became an important part of the underlying architecture. The system has to create enough meaningful context to reason even when the underlying dataset is imperfect.

Organizationally, the harder question is trust. Giving AI meaningful decision authority touches several functions at once. Data and IT teams want explainability, security and low integration burden. Marketing teams need confidence that customer communication will remain on-brand. CRM teams want to retain strategic control rather than feel that an AI system is taking the customer relationship away from them.

Our approach has therefore been to separate strategy from execution. People define the objectives, brand directive and guardrails, while Maestro owns the repeatable individual decisioning and execution inside those boundaries. Every committed decision is designed to be traceable, and the system works on top of the existing stack rather than forcing a platform migration.

As AI becomes responsible for optimizing revenue and retention, how do you prevent an autonomous system from optimizing too aggressively for short-term metrics, such as pushing unnecessary promotions or purchases, at the expense of customer trust and long-term brand value?

Maestro optimizes against the KPIs humans define, not metrics it chooses for itself. If you define the goals as customer lifetime value and retention rate, that is what it optimizes for, rather than transaction velocity or promotion volume. If aggressive short-term tactics hurt retention, they fail against the actual KPI. The system learns that trade-off directly because each outcome feeds back into memory, so strategies that damage long-term retention become demonstrably worse for the business goal that has been set.

The constraint is structural. Maestro works within the guardrails you establish. It can’t optimize around them; it operates inside them. Every decision is traceable, so you can audit why Maestro chose to send or hold back, which KPIs weighted that choice, and what the expected outcome was. Without clear KPI definition, guardrails, and decision transparency, an AI system could push too aggressively. But Replenit’s model is transparent and designed to avoid that. The business defines what matters, Maestro executes toward those goals with full auditability, and every outcome helps determine whether the strategy is actually working.

Foundation models are improving rapidly and increasingly capable agents will become available to almost every retailer. Over the long term, where do you believe the defensible advantage in AI decision-making will come from: proprietary models, customer-level memory, domain-specific reasoning, feedback loops, infrastructure, or something else entirely?

Foundation models will become commoditized, so everyone will have access to increasingly capable reasoning engines. The defensibility isn’t the model itself, it’s what you know about each customer and how fast you compound that knowledge. Maestro’s advantage is customer-level memory that deepens with every purchase and every outcome. After six months of outcomes, the system knows your customer better than a generic model ever will. Not just demographics, but what drives them, when they’re receptive, what triggers churn.

That memory also creates continuity. Switching to another system means losing two years of learned context. Alongside that, Replenit has domain-specific skills encoded for retail, including understanding lifecycle stages, purchase velocity patterns, replenishment triggers, category affinity in beauty or groceries. But skills can be copied. The real advantage is the feedback loop: Maestro makes a decision, captures the outcome, updates memory, and uses that learning to  make the next decision smarter. That loop is continuous and compounding. Over time, a system running on your data becomes progressively better than a fresh system on generic models.

Infrastructure also matters. Swarm concurrency and owned models keep unit economics better, but that’s table stakes, not moat. The defensible advantage is the combination of customer data gravity (switching costs), continuous learning (outcomes feeding back into future decisions), and domain-specific reasoning (retail-specific context), all working together. In isolation, each is copyable. Together, they’re hard to replicate fast enough to catch up.

Thank you for the great interview, readers who wish to learn more should visit Replenit.