Agentic CRM: turning your existing stack into a decision engine
Agentic CRM is not another tool. It is seven layers (signals, diagnosis, decision, execution, experimentation, learning and human judgment) that turn a DTC brand's existing stack into a system that decides the next best action per customer.
Most DTC brands do not have a tooling problem. They have a deciding problem. The stack already captures more than the team can read: orders, sessions, subscription events, loyalty balances, support tickets, reviews, and a full history of every message the customer has ever been sent. What is missing is the layer that turns all of that into a single answer to a single question, per person, every day: what is the right thing to do for this customer right now?
That is what agentic CRM means in practice. Not a chatbot bolted onto the ESP. Not a new platform to migrate to. A set of agents that read the customer completely, infer state, decide the next best action, execute it across the channels you already run, test continuously, and feed what they learn back into the system. We build this inside the stack a brand already owns, because the differentiator was never the tool.
Layer 1: Signals: what is actually happening
Before deciding anything, the system has to read the whole customer, not the slice that happens to sit inside the ESP. Every one of these sources already exists in a mid-size DTC stack. Most of them are never joined.
| Source | What it carries | What it tells the system |
|---|---|---|
| Purchases | Frequency, ticket, recency, margin | Economic value and buying rhythm |
| Browse | Abandonment, category affinity, intent signals | Direction of interest before it becomes an order |
| Subscription | Skips, pauses, cancellations, cadence changes | Friction with the model, not always with the product |
| Loyalty | Points, tier, rewards redeemed, referrals | Engagement that has not yet converted into a reason to return |
| Support | Tickets, friction reasons, sentiment, incidents | Relationship risk that outranks any campaign |
| Reviews & UGC | Satisfaction, objections, customer language | The words to use and the objections to answer |
| Prior exposure | Emails, SMS, push, offers, campaigns already received | Commercial pressure already spent on this person |
The last row is the one brands skip, and it is the one that breaks everything downstream. An agent that cannot see what a customer already received will re-send, over-discount, and burn the relationship while reporting a healthy attributed-revenue number.
Layer 2: Diagnosis: what the signal means
A CRM groups customers. An agent should infer state. The difference is that a segment describes a shared property; a state describes an individual situation with an implied consequence. 'High-value, lapsed 60 days' is a segment. 'High LTV potential, close to a second purchase, has not received the right stimulus, and should not be discounted' is a state.
The inferences that matter most in a DTC context are consistent across brands:
- Close to a second purchase, but has not yet received the right stimulus. Timing and product fit, not price.
- Losing intent, but not necessarily on price. Discounting here destroys margin and solves nothing.
- High LTV potential, so do not burn the relationship with a discount that resets the reference price.
- An open support ticket changes the priority of the next communication. Sometimes the next best action is a resolution, not a campaign.
- A loyalty member accumulating points without ever finding a reason to come back. The mechanic works; the trigger is missing.
This is where the CRM stops grouping customers and starts understanding individual states. It is also where most 'AI CRM' pitches quietly stop, because inference without a decision layer is just a nicer dashboard.
Layer 3: Decision: what to do now
This is the core of the system, and the part almost nobody builds. The output of this layer is not a segment. It is a decision, per customer, with a defensible reason attached.
- What is the next best action for this person?
- Which channel has priority: email, SMS, push, loyalty, support or none?
- Which incentive makes sense, and which one destroys margin for no reason?
- When is the right moment?
- Who should be excluded from this campaign?
- When is the best action to send nothing at all?
That last question is the one that separates an agentic system from a busier one. Suppression is a decision with a measurable value, and a system that can never choose silence will always over-send, because sending is the only move it knows.
The practical implication: the decision layer needs explicit business constraints, margin floors, contact frequency caps, discount eligibility, brand-voice limits, encoded as inputs, not as afterthoughts. An agent optimising engagement without a margin constraint will find the discount every single time.
Layer 4: Execution: how a decision becomes an action
Campaigns do appear here, as a consequence of a decision, never as the starting point. Once the decision exists, execution is largely mechanical and is where agents deliver the most obvious leverage in hours saved.
- Briefs and copy variants, generated against the customer's own language from reviews and support.
- Product, offer or incentive selection consistent with the margin decision made upstream.
- Flow construction inside the existing ESP or CDP.
- Email, SMS and push orchestration as one sequence rather than three roadmaps.
- Loyalty, referral and post-purchase actions triggered by state, not by calendar.
- Support automations that recover a relationship before it turns into churn.
Note what is not in this list: a monthly campaign calendar. If the calendar still drives execution, the layers above it are decoration.
Layer 5: Experimentation: which hypothesis deserves a test
This is the most Loiale layer of the seven. Not a campaign calendar. A hypothesis engine that runs continuously over flows, offers, timing and segments.
| Question | What the agent evaluates |
|---|---|
| What to change | Timing, offer, audience, channel, frequency or mechanic |
| Which test has the most upside | Expected effect size against reach and margin exposure |
| How to allocate traffic | Holdout size that produces a readable result without damaging revenue |
| When a result is solid enough | Confidence and duration thresholds set before the test runs |
| What becomes permanent | Which learning is promoted into a standing rule of the system |
"Six campaigns a month is a content plan. Continuous testing over flows, offers, timing and segments is a growth engine. Only one of them compounds."– Loiale
The compounding effect comes from the last row. A test that produces a result and no rule is a report. A test that produces a rule the decision layer applies from then on is an upgrade to the system itself.
Layer 6: Learning and operations: how the system compounds
After every action, the loop has to close. This layer is the difference between a system that improves and a system that merely runs.
- Measure incremental impact against a holdout, not surface-level attribution.
- Detect anomalies: a drop in repeat rate, excessive commercial pressure, a flow that quietly stopped working.
- Convert results into new rules or new hypotheses.
- Prioritise the weekly backlog by expected value, not by whoever asked loudest.
- Produce reporting that explains what changed, why, and what happens next.
Anomaly detection is the underrated one. Most retention programmes do not fail dramatically; they decay. A flow degrades over eight weeks and nobody notices because the aggregate number is carried by something else. An agent watching flow-level incrementality catches that in days.
Layer 7: The human layer: what must never be delegated blindly
Agentic does not mean removing the operator. It means moving the operator up a level: from executing decisions to defining the conditions under which decisions are made. The team still owns:
- Business objectives and margin limits.
- Brand voice and promotional sensitivity.
- Reputational risk.
- Which experiments should never be run at all.
- When a decision needs human judgment rather than local optimisation.
That last point is the honest one. An agent optimises what it can measure, inside the window it can see. It will happily trade a long-term relationship for a short-term conversion if nothing stops it. The human layer is the constraint set, and it is not optional infrastructure; it is the reason the system stays aligned with the business rather than with the metric.
The maturity ladder
| Stage | Unit of work | Who decides | Typical ceiling |
|---|---|---|---|
| Rules | Trigger and send | Whoever built the flow, once | Breaks as soon as the catalogue or cohort changes |
| Segments | Audience and campaign | CRM manager, weekly | Capped by how many segments a human can maintain |
| Decisions | Next best action per customer | Agents inside human constraints | Capped by data quality and margin discipline |
Almost every brand we audit sits between rules and segments, with a tool stack that could support decisions today. The blocker is rarely the model. It is that signals are not joined, prior exposure is not tracked, margin is not available per action, and no holdout exists to prove any of it worked.
A worked example: one customer, one day
Abstractions are easy to agree with and hard to build. Here is the same day for one customer, run through a rules stack, a segment stack and a decision stack. The customer is real in shape if not in name: second-time buyer, 47 days since last order, above-median basket, one open support ticket about a delayed delivery, 320 loyalty points unredeemed, opened the last three emails and clicked none.
| Stage | What the system does | What it costs |
|---|---|---|
| Rules | Day-45 winback fires with a 15% code, because that is what the flow says | Discount handed to a customer who was never leaving, on the day she is angry about a delayed parcel |
| Segments | She sits in 'lapsed 30-60d, high value' and receives the weekly campaign for that audience | Right audience, wrong moment; the support ticket is invisible to the ESP |
| Decisions | Suppress all commercial sends; escalate the ticket; once resolved, send a points-redemption reminder tied to the category she browsed twice this week | Zero margin spent, relationship recovered, next purchase triggered by intent rather than by a calendar |
Nothing in the third row requires a model nobody has. It requires that the support ticket, the loyalty balance, the browse event and the send log land in the same place, and that something is allowed to say no to a scheduled send.
Reference architecture: what actually gets built
The seven layers map onto a small number of concrete components. None of them is exotic. What makes the difference is that they are wired in this order, rather than bolted on wherever there was budget.
1. A joined customer record
One identity per human, resolved across order, session, subscription, loyalty and ticket data. Usually the warehouse or the CDP, rarely the ESP. If email address is still the join key and guest checkout exists, half the signal surface is already lost.
2. An exposure log
Every message, offer and discount ever sent to that person, with timestamp, channel and incentive value. This is the single most commonly missing table and the cheapest one to build. Without it, contact frequency caps and discount eligibility cannot be enforced, only hoped for.
3. A state layer
Derived attributes recomputed on a schedule: predicted next order date, margin contribution, discount dependency, channel responsiveness, relationship risk. These are the inputs the decision layer reasons over. Keep them auditable; a state you cannot explain is a state you cannot defend when it produces a bad send.
4. A decision service with a policy file
The agent proposes, the policy disposes. Margin floors, frequency caps, discount eligibility, brand-voice rules and blackout windows live in version control, reviewed like code. Every decision writes its reason. If you cannot answer 'why did this person get this message', you do not have a decision layer, you have a black box.
5. Executors per channel
Thin adapters into the ESP, SMS provider, push, loyalty and helpdesk. Deliberately dumb: they take a decision and carry it out. Keeping intelligence out of the channel tools is what lets you swap a vendor without rebuilding the brain.
6. A holdout, from day one
A global holdout of 5 to 10 percent, plus per-initiative holdouts, decided before the first send rather than after the first good week. Without it the system produces attributed revenue instead of incremental revenue, and nobody can tell the difference.
How we phase it in ninety days
| Phase | Weeks | What ships | What it proves |
|---|---|---|---|
| Signal audit | 1-2 | Joined record, exposure log, gap list | Whether the data can support decisions at all |
| State layer | 3-5 | Derived states, margin per action, holdout in place | That the diagnosis matches what operators already suspect |
| First decisions | 6-9 | Suppression, frequency caps, next best action on two journeys | Incremental lift with less volume, not more |
| Execution and tests | 10-12 | Channel adapters, hypothesis backlog, weekly readout | That the loop closes and learning becomes rules |
Notice the order. Suppression before personalisation. The fastest measurable win in most brands is not a smarter message, it is fewer wrong ones. Cutting commercial pressure on the wrong people usually pays for the build before the interesting layers are live.
Five ways this fails
- Buying the agent before joining the data. The model then hallucinates state from a partial view and is confidently wrong at scale.
- No exposure log. The system optimises each message in isolation and collectively over-contacts every good customer you have.
- No margin input. Engagement-optimised agents converge on discounting, because discounting always wins the metric they were given.
- No holdout. Six months later there is a deck full of attributed revenue and no way to answer whether any of it was incremental.
- Learning that never becomes a rule. Tests run, results are read, nothing changes in the decision policy, and the system stays exactly as smart as it was on day one.
What to measure
| Measure this | Not this | Why |
|---|---|---|
| Incremental revenue vs holdout | Attributed revenue | Attribution credits sends to purchases that would have happened anyway |
| Contacts per conversion | Total sends | Efficiency of pressure, not volume of activity |
| Margin per contacted customer | Revenue per campaign | Keeps discounting honest |
| Repeat rate by cohort | Blended repeat rate | Aggregate hides decay in newer cohorts |
| Decisions with a recorded reason | Automations live | Auditability is what makes the system safe to expand |
| Rules promoted from tests | Tests run | Measures compounding, not busyness |
How Loiale builds this
We work inside the stack you already run. First we audit the signal surface and fix what is missing. That is layer one, and skipping it is why most AI retention projects stall. Then we build the diagnosis and decision layers against your real margin constraints, wire execution into the ESP, CDP and loyalty tools you already pay for, and stand up the hypothesis engine that keeps the system improving after we hand it over.
No new platform to migrate to. No agency retainer producing six campaigns a month. A decision engine that reads your customers completely and acts on what it sees, with your team holding the constraints.
Frequently asked questions
- What is an agentic CRM?
- An agentic CRM is a customer system where software agents read every available signal, infer each customer's state, decide the next best action within business constraints, execute it across channels and feed results back as rules. The difference from a traditional CRM is that the unit of work is a decision per customer rather than a campaign per segment.
- Is agentic CRM just marketing automation with AI?
- No. Automation executes rules a human wrote in advance. An agentic system decides which rule should apply, to whom, through which channel, at what moment, and whether the right action is to send nothing at all. Suppression as a first-class decision is the clearest dividing line.
- Do we need to replace our ESP or CDP?
- Almost never. The decision layer sits above the tools and uses them as executors. Klaviyo, Braze, Iterable, Shopify, a warehouse and a helpdesk are enough for most mid-size DTC brands. Migrations delay the only work that actually changes results.
- What data do we need before starting?
- A joined customer record across orders, sessions, subscription, loyalty and support, plus an exposure log of every message and incentive already sent. Margin per product or order is the third input. Brands missing the exposure log should build it first; it is cheap and it unlocks suppression and frequency control immediately.
- How long does it take to see results?
- The first measurable win usually comes from suppression and frequency control in weeks six to nine: less volume, equal or higher incremental revenue. A full loop with continuous experimentation is typically live within ninety days.
- How do you prove it worked?
- A global holdout of 5 to 10 percent set up before the first decision ships, plus per-initiative holdouts. Everything is reported as incremental revenue against that holdout rather than as platform-attributed revenue.
- What should never be delegated to an agent?
- Business objectives and margin limits, brand voice and promotional sensitivity, reputational risk, experiments that should never run, and any decision where long-term relationship value outweighs a measurable short-term gain. Those live in a policy the team owns and reviews.