Agentic customer experience: one decision layer for the whole customer
Most customer experience automation automates sending, not deciding. The agentic version puts one decision layer across lifecycle, loyalty, retail and service, and runs it through the tools a DTC brand already owns.
Most of what gets sold as customer experience automation is a campaign calendar with conditions attached. A flow fires when someone abandons a cart. A tier email goes out when someone crosses a points threshold. A survey lands four days after delivery. Each piece works. Together they behave like four teams sending to the same person without ever speaking to each other, because that is exactly what they are.
The agentic version changes the unit of work. Instead of a campaign per segment, the system produces a decision per customer: what is the single best thing to do for this person right now, across every surface where the brand can act, given what we know and what the margin allows. Sometimes that decision is a message. Often it is a reward, a prompt to a shop assistant, or nothing at all.
One boundary before anything else: we design the decision layer. Execution keeps running through the ESP, CDP, loyalty platform and POS the brand already pays for. There is no migration in this article, and there should not be one in the project either.
The seven layers, applied to experience
We described these layers in the agentic CRM piece, where the actions were messages. The layers do not change when the scope widens. The action space does.
| Layer | Question it answers | Widened for experience |
|---|---|---|
| Signals | What is happening? | Adds POS visits, loyalty balances, ticket state, in-store scans |
| Diagnosis | What state is this person in? | One state per human, not one per channel |
| Decision | What is the best next action? | Chooses between message, reward, staff prompt or silence |
| Execution | How does it happen? | Thin adapters into ESP, SMS, loyalty, POS and helpdesk |
| Experimentation | Is it better than nothing? | Holdouts per surface, not only per campaign |
| Learning | What becomes permanent? | Winning decisions promoted into the policy file |
| Human layer | What stays ours? | Margin floors, brand voice, escalation, what never runs |
This is the whole idea of agentic marketing in one line: keep the brain in one place and let every surface borrow it. The moment each tool holds its own logic, the brand starts contradicting itself in public.
Where the decisions land
Lifecycle and CRM
The core domain and the one with the most existing plumbing. The agent decides channel, timing, offer depth, cross-channel frequency and suppression. Next best action marketing lives here in its most familiar form: one person, one queue of candidate actions, one winner scored against margin rather than open rate. The deeper mechanics of this domain are covered in the agentic CRM article.
Loyalty and membership
Loyalty is the richest decision surface most brands never treat as one. A program generates continuous state: balance, tier distance, burn behaviour, reward preference, expiry pressure. An agent uses that state to choose the next reward, the tier challenge worth offering, whether a points expiry prompt would help or annoy, which mission fits the person's actual buying rhythm, and when a referral ask is likely to land. Static rules cannot do this, because the right nudge for a member two visits from a tier is not the right nudge for someone who has not burned a point in nine months.
Retail and in-store
Where a brand sells offline, the decision layer has one more executor: the person at the till. The agent can push a short prompt to POS when a known member checks in, recognise a tier without the customer asking, follow up on a card-linked visit, or bridge an online browse into a store recommendation. This is also the identity work: without a scan, a QR check-in or a phone lookup, the offline half of the relationship stays invisible and every decision is made on half the evidence.
Service
Support is in scope, but as an input first. Ticket state is one of the highest-value signals a decision layer can read, and it is almost never wired into commercial sending. The three decisions worth taking here are narrow and unglamorous: hold commercial pressure while an issue is open, choose a recovery action after a resolved incident, and route a high-value relationship to a human rather than a queue. That is enough. Automating the support function itself is a different project with a different owner.
The decision matrix: agent, rule or human
Not every task deserves an agent. The useful question is not what could be automated, it is what should decide each task. Three modes cover almost everything. Agentify when the decision is repetitive, reversible, high volume and scored against a number the business trusts. Keep a rule when the outcome is legal, financial or brand critical and the right answer never changes. Delegate to a human when the value of the moment is higher than the cost of a person, or when judgement about a relationship is the actual product.
| Task | Mode | Guardrail |
|---|---|---|
| Send or suppress a lifecycle message | Agent | Frequency cap, blackout windows, global holdout |
| Choose channel and timing | Agent | Per channel caps and consent state |
| Offer depth and reward selection | Agent | Hard margin floor set in the policy file |
| Next best action ranking | Agent | Scored on margin, every decision logged with a reason |
| Points expiry and tier progress nudges | Agent | Loyalty state must be fresh, never over a stale balance |
| Store staff prompt on member check-in | Agent proposes, human executes | Staff can dismiss, dismissals feed back as signal |
| Hold commercial pressure while a ticket is open | Rule | Non negotiable, no agent override |
| Consent, opt out and data retention | Rule | Legal, versioned, audited |
| Discount ceilings and price integrity | Rule | Set by finance, agent proposes only inside the band |
| Tier qualification and reward eligibility | Rule | Program terms are a promise, not a variable |
| Brand voice, creative and campaign concepts | Human | Agent selects among approved assets, never writes the promise |
| Recovery for a high value complaint | Human | Agent flags and prepares context, a person decides |
| Churn save on a top decile relationship | Human | Owned account, agent surfaces the trigger |
| Support resolution itself | Out of scope | Ticket state is an input, not a domain we automate |
The last row is deliberate. Support belongs in the matrix as a signal and as an escalation path, not as the core of the system. Brands that start by automating deflection end up with a cheaper helpdesk and the same commercial engine they had before. The compounding value sits in lifecycle, loyalty and retail, where every decision changes the state that feeds the next one.
Two rules keep the matrix honest over time. First, tasks move only in one direction and only with evidence: a human decision becomes a rule once the pattern is stable, a rule becomes an agent decision once there is a holdout showing the agent beats it. Second, nothing moves without a reversal path. If the agent cannot be switched back to the rule in one deploy, it is not ready to own the task.
Concrete examples: email, SMS and loyalty
The matrix is easier to use when each channel has its own examples. Below are the typical tasks we see in DTC brands, split by what the agent decides, what the policy file locks as a rule, and what a human still owns.
| Email task | Mode | Why it lands there |
|---|---|---|
| Send or suppress a lifecycle message | Agent | Scored against frequency cap, ticket state, recency and margin |
| Choose subject line from approved pool | Agent | Selects among variants the brand has already written and cleared |
| Personalise send time per recipient | Agent | Predicts open-to-conversion window from behaviour |
| Audience segmentation for a campaign | Agent proposes, human approves | First run needs brand sign-off; agent can own once holdout validates |
| Consent and unsubscribe handling | Rule | Legal requirement; no override, no optimisation target |
| Discount ceiling per segment | Rule | Finance sets the band; agent chooses inside it |
| Brand voice and campaign concept | Human | Agent never writes the promise; it selects from approved assets |
| Escalation after a deliverability incident | Human | Operator calls the ISP or ESP; agent flags the drop |
| SMS task | Mode | Why it lands there |
|---|---|---|
| Send or suppress a text | Agent | Higher interruption cost means suppression is even more valuable |
| Choose the SMS window per recipient | Agent | Local time, recent opens, and purchase intent predict the best slot |
| Send a restock or low-stock alert | Agent | Triggered by inventory and preference signals |
| Send a high-priority fraud or security alert | Rule | Compliance and trust; fired by the event, not scored |
| Opt-in and consent refresh | Rule | Regulatory; time-based and channel-specific |
| Write a flash-sale message from scratch | Human | Creative concept; agent can only assemble from approved copy |
| Handle an angry reply to an SMS | Human | Relationship risk; agent routes to the right operator with context |
| Loyalty task | Mode | Why it lands there |
|---|---|---|
| Choose the next reward or mission | Agent | Predicts preference from burn history, margin and tier distance |
| Prompt a referral ask | Agent | Timed to positive moments; scored by relationship quality |
| Offer a tier challenge | Agent | Dynamic goal based on predicted spend and cost of the tier |
| Points expiry warning | Agent proposes, rule validates | Policy decides whether a warning is sent; agent picks the timing and reward |
| Tier qualification and reward eligibility | Rule | Program terms are a promise; the agent does not change them |
| Tier reset and annual requalification | Rule | Program governance; fixed dates and thresholds |
| Design a new tier or reward archetype | Human | Strategic product and brand decision |
| VIP recovery after a bad experience | Human | Agent surfaces the trigger and context; a person decides the make-good |
Worked example: one customer, one day, four decisions
A member, eleven orders in two years, 340 points, one delivery complaint opened yesterday, browsed the new range this morning, walks into the flagship store at 18:40. Four systems have an opinion. Only one of them should win at any moment.
| Moment | Typical stack | Decision layer | Why |
|---|---|---|---|
| 09:00, browse | Browse abandonment email | No send | An open complaint outranks a commercial nudge |
| 12:00, ticket resolved | Nothing, or a CSAT survey | Recovery credit, no discount | Repairs the relationship without training a discount habit |
| 16:00, points expiry job | Bulk expiry warning | Hold until the store visit | The reward is worth more attached to a real occasion |
| 18:40, store check-in | Nothing, staff sees a stranger | Staff prompt: VIP, complaint resolved, 340 points to burn | The one moment where a human can outperform any message |
| 21:00, post visit | Generic thank you flow | Points confirmation, next tier in one visit | Closes the loop with the state the customer just changed |
None of the right-hand column needs a model that does not exist. It needs the ticket, the points balance, the browse event, the POS scan and the send log to sit in one place, and something with permission to say no to a scheduled send.
What actually gets built
The components are few and boring. The order matters more than the sophistication.
- A joined customer record: one identity per human across orders, sessions, loyalty, POS and tickets. Not the email address as a join key.
- An exposure log: every message, reward and discount already given, with timestamp, channel and value. The cheapest table to build and the most commonly missing.
- A state layer: predicted next order, margin contribution, discount dependency, tier distance, relationship risk, all recomputed on a schedule and all explainable.
- A decision service with a policy file: the agent proposes, the policy disposes. Margin floors, frequency caps, blackout windows and brand rules live in version control.
- Thin executors: adapters into the ESP, SMS, loyalty platform, POS and helpdesk. Deliberately dumb, so a vendor swap never touches the brain.
- A holdout from day one: 5 to 10 percent global, plus per-initiative holdouts, agreed before the first decision ships.
Ninety days, in order
| 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 live | That the diagnosis matches what operators already suspect |
| First domain | 6-9 | Suppression, frequency caps, next best action on lifecycle | Incremental lift with less volume, not more |
| Second domain | 10-12 | Loyalty or retail decisions on the same core | That the decision layer generalises beyond messaging |
Suppression before personalisation, always. In most brands the fastest measurable win is not a smarter message, it is fewer wrong ones, and it usually pays for the build before the interesting layers go live.
Five ways this fails
- Treating it as a marketing project. If loyalty, retail and support are not in the room, the agent decides on a partial customer and is confidently wrong at scale.
- Ignoring offline. Where stores exist and identity is not captured at the till, half the relationship is missing from every decision.
- Buying the agent before joining the data. The layer that pays is layer one, and it is the one everyone wants to skip.
- No exposure log. Each surface optimises in isolation and collectively over-contacts every good customer the brand has.
- No holdout and no margin input. Engagement-optimised systems converge on discounting, then report attributed revenue that nobody can defend.
What to measure
| Measure this | Not this | Why |
|---|---|---|
| Incremental revenue vs holdout | Attributed revenue | Attribution credits actions 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 |
| Identified visits in store | Footfall | Offline decisions are impossible without identity |
| 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 it
We start with the signal surface: what the brand already captures, what it silently loses, and what has to exist before anything is allowed to decide. Then we build the state and decision layers against real margin constraints, and connect them to the activation stack already in place, including the loyalty platform and the till.
No new platform to migrate to, no retainer producing six campaigns a month. One layer that reads the customer completely and acts on what it sees, with your team holding the constraints.
Frequently asked questions
- What is customer experience automation?
- Customer experience automation is the practice of running customer touchpoints, messaging, loyalty, service and in-store follow-up, without a human triggering each one. Most implementations automate execution against fixed rules. The agentic version automates the decision itself: what to do for a given person right now, across every surface, within business constraints.
- How is agentic customer experience different from marketing automation?
- Marketing automation executes rules a human wrote in advance, one flow per channel. An agentic layer decides which action should apply, on which surface, at what moment, and whether doing nothing is the better answer. It also arbitrates between channels, so loyalty, lifecycle and retail stop competing for the same customer.
- Is this the same as agentic CRM?
- Agentic CRM is the lifecycle and messaging part of it. Agentic customer experience uses the same seven-layer model but widens the action space to loyalty rewards, in-store prompts and service-aware suppression. Same brain, more surfaces.
- Do we need to replace our ESP, CDP or loyalty tool?
- Almost never. The decision layer sits above those tools and uses them as executors. A typical mid-size DTC stack of Shopify, a warehouse, an ESP, a loyalty platform, a POS and a helpdesk is enough. Migrations delay the only work that changes results.
- How does it handle in-store touchpoints?
- Through identity capture at the till, a scan, a QR check-in or a phone lookup, and a thin adapter that can surface a short prompt to staff. Once an offline visit is attached to the customer record, in-store becomes another surface the decision layer can act on and measure.
- How do you prove it works?
- A global holdout of 5 to 10 percent set up before the first decision ships, plus per-initiative holdouts on each surface. Results are reported as incremental revenue against that holdout, never as platform-attributed revenue.