2027 Retail Leaders’ Series · The operating model, explained
How customer engagement runs in the AI era.
This report explains CRM 5.0, the operating model that replaces both campaign-led CRM and the AI co-pilot layer bolted on top of it. It shows what changes, where each system in your stack sits, who does what, how decisions are made for each customer and across a retail group, and how the cost of customer engagement changes over time. It starts on your current stack, with no rip and replace.
The shift in four numbers
Why the operating model is changing now
70–80%
of customer messages will be decided by AI for each customer, not sent as campaigns
BCG, 2026
60–90
days and 20+ people to build one campaign today
BCG, 2026
3–5
people can run the same work with AI, in days instead of months
BCG, 2026
20–40%
more engagement and conversion where AI decides, in early results
BCG, 2026
Figures from BCG’s 2026 report “The End of Marketing Campaigns as We Know Them”, which studies how AI is replacing campaign-based marketing. They set the scale of the change this report explains: how much work moves to AI, what it costs today, and what early adopters see.
About this report
The shift
Every martech tool can send. What separates retailers now is the decision behind each action: who, which product, why now, in which tone, toward which commercial outcome.
CRM has moved through four eras, and each solved the problem of the one before. CRM 1.0 recorded who the customers were. CRM 2.0 broadcast the same message to all of them. CRM 3.0 automated campaigns and journeys for segments, built by hand. CRM 4.0, the AI co-pilot era, added AI inside the same tools to tune sends, scores and copy. In every one of them, people still decide what happens to each customer. CRM 5.0 changes that: the team sets objectives and builds skills, codified commercial know-how that an AI CRM Manager applies to every customer, every day. The job is no longer maintaining a stack. It is building the skills that fit the business.
Exhibit 1a
Five eras of CRM: who decides changes only in the last one
Era
What it was
Typical tools
Who decides for each customer
The ceiling
CRM 1.0 · Record
1990s–2000s
A database of customers and contacts
Customer database, contact management
Nobody: the system only stores
Knowing who customers are, not what to do
CRM 2.0 · Broadcast
2000s
The same message to the whole list
Email service providers, newsletters
The marketer, for everyone at once
One message for all; relevance by luck
CRM 3.0 · Automate
2010s–early 2020s
Segments, journeys and triggers built by hand
Marketing automation, CEP, CDP audiences
The CRM team, per segment
Output capped by how many journeys the team can build
CRM 4.0 · Assist
2024 onward
AI co-pilots added to the same tools
Predictive scores, send-time optimisation, AI copy
Still the CRM team; AI tunes what they built
Faster campaigns, same ceiling
CRM 5.0 · Decide
Now
An AI CRM Manager that decides and executes 1:1
An intelligence layer over your existing stack
Maestro, inside goals and guardrails your team sets
Set by your skills and guardrails, not by headcount
Why CRM 4.0 is not the answer
The AI co-pilot looks like progress, and it does make teams faster. But it is added to the same operating model, so it inherits the same bottlenecks. The team still writes the brief, builds the segment, designs the journey and presses send. AI tunes the result after the important choices were made. That is why confidence in martech AI is falling even as adoption grows.
Exhibit 1b
The bottlenecks CRM teams face, and why CRM 4.0 leaves them in place
Bottleneck
What teams face today
What CRM 4.0 adds
Why it does not fix it
Team capacity
Every play is designed, built and checked by people
AI writes copy and variants faster
The team still decides and builds each campaign; output still grows with headcount
Averaging
Customers grouped into segments that get the same message
Predictive scores on each profile
A score ranks likelihood; someone still decides what to do with it
Lost objective
The business goal is narrowed to what the tool can configure
AI optimises what was activated
Tuning starts after the objective was already cut down
Timing
The calendar sets when messages go out
Send-time optimisation
It picks the best hour for a campaign, not whether this is the customer’s moment
Fragmented logic
Rules spread across email, site, app and store tools
Separate AI inside each tool
Each tool’s AI sees only its own channel and data
Cost
Licences that grow with contacts and messages
AI sold as a premium tier
The bill rises; the operating model does not change
Exhibit 1
From campaigns and co-pilots to context-driven decisions
CRM 3.0 · Campaign era · product-centric
Starts from the promo calendar
P&L objective
Offer and campaign per product
Built by hand
Audience and channel
Picked by the CRM team
Send
Output is capped by how many campaigns the team can build
CRM 4.0 · AI co-pilot era · same model, AI on top
Martech AI inside the same tools
P&L objective
Campaign built by the team
Same brief, same canvas
AI tunes the send
Send time, subject line, variant
Send
Faster campaigns, same ceiling; the team still decides and builds
CRM 5.0 · customer-centric
Starts from each customer’s context
Customer, Product and Brand Memory
Maestro applies skills
Micro and macro sub-decisions
Execution-ready output
Product, timing, channel, tone
Your existing channels
↺ Learns from every outcome
Adapted from BCG Exhibit 1 (2026). Brand launches, seasonal moments and transactional messages stay as planned pushes alongside the always-on layer.
Exhibit 2
Three layers: your team directs, Maestro decides, your stack executes
Objectives & guardrails
Your team
Commercial goals, brand rules, contact limits, offer budgets, and the skill library you own.
Intelligence layer
Maestro · AI CRM Manager
Reasons, remembers, decides and executes. Reads Customer, Product and Brand Memory, applies skills and makes the sub-decisions behind every action.
Execution layer
Your existing stack
ESP, MAP, CEP, app, site, CRM, store. Kept as is to start. Becomes a reliable delivery endpoint instead of the place where logic is built, and over time a leaner, API-driven one.
Skills become the retailer’s intellectual property. Channel tools become delivery.
Takeaway
- The value in customer engagement now sits in the decision, not in the send.
- CRM 3.0 and CRM 4.0 share one ceiling: people decide and build everything. AI co-pilots make that faster; they do not change who decides.
- Your stack stays. Planned pushes such as launches and seasonal moments stay too. What changes is the work your team does.
The organising principle
Every action answers three questions at once: who is this customer, what do they need now, and what must the brand achieve. Legacy tools start from the third and push it onto audiences. Maestro holds all three and reconciles them per customer.
The model is organised around customers, not the calendar. Commercial goals do not disappear. They become weights and limits on each decision, instead of the starting point that decides who gets which message. Maestro infers much of the customer picture from what each customer buys, not only from what they declared.
Exhibit 3
Three inputs, reconciled per customer
Who they are
The customer
- Household members each purchase is for
- Life stage, skin or hair type, routine
- Price sensitivity and loyalty
- Where and when they respond
What they need now
The need
- A product about to run out
- A routine with a missing step
- A first buyer with no second purchase
- A recurring product quietly dropped
- A season or life event starting
What the business must achieve
The brand’s goal
- Retention and repeat rate
- Basket size and category growth
- Margin floor and incentive budget
- Stock, launch and seasonal priorities
Maestro reconciles
Sub-decisions per customer
- Act or wait
- Which product
- When
- Which channel
- Which tone
- Incentive or none
When the customer’s need and the brand’s goal pull apart, guardrails decide. Example: no discount for a customer who buys at full price anyway, even when the goal is volume.
Exhibit 4
How each input shapes each sub-decision
Sub-decision
From the customer
From the need
From the brand’s goal
Act or wait
Recent contacts, channel fatigue
Is the need live today
Priority of this objective this period
Which product
Inferred traits, which household member
The item due, missing or dropped
Margin, stock, category to grow
When
Personal cycle, active hours
Time left before run-out or season
Launch dates, promotion windows
Which channel
Where this customer responds
Urgency of the need
Channel cost and reach
Which tone
Life stage, relationship with the brand
Reminder, advice or discovery
Brand voice rules
Incentive
Price sensitivity
Risk of losing the purchase
Incentive budget, margin floor
Takeaway
- The model starts from each customer, yet your commercial goals stay in control as weights and limits on every decision.
- Customer understanding comes from what people buy, not only from what they declared.
How the model runs
BCG describes the destination. This is how it runs in a retail deployment, and how work splits between your team, Maestro and your stack.
What Maestro is
Maestro is an AI CRM Manager: an AI workforce role built for retail. A retailer hires it, onboards it and holds it accountable, the way it would a senior CRM manager. It reasons about every customer, remembers what it learns, decides what should happen next and executes through the channels already in place. Your team sets the objectives, the guardrails and the skills; Maestro does the daily 1:1 work.
Exhibit 4b
What Maestro does that other approaches do not
Distinction
Most martech and AI add-ons
Maestro
Segment vs individual
Better segments and micro-segments
Reasons per customer, with no segment in the middle
Tune the send vs decide the move
Optimise the campaign you already chose: audience, send time, subject, variant
Decides what should happen next: whether to act, which product, why now, which tone
Co-pilot vs worker
Help a marketer do tasks faster; the human stays at the centre
Owns the work and the outcome, inside your guardrails
Inside one tool vs above the stack
AI that lives inside one vendor’s channel and data
One decision layer above every channel you run
Generic vs retail
Built for every industry, from banks to travel
Retail skills: replenishment, routines, fit, households, markdowns
Feature vs memory
Starts again on each campaign or session
Customer, Product and Brand Memory that learns from every outcome
Exhibit 5
CRM 3.0 vs CRM 4.0 vs agent-native vs CRM 5.0 with Maestro
Dimension
CRM 3.0 · Campaign era
CRM 4.0 · AI co-pilot era
Agent-native (BCG)
CRM 5.0 with Maestro
Starting point
P&L target and promo calendar
The same campaign, already built
Customer DNA, shelf of offers and content
Customer, Product and Brand Memory, read per customer daily
Unit of work
Campaign or journey per audience
The same campaign, tuned
1:1 next-best action
A skill applied to every eligible customer
Who decides
CRM manager picks audience, product, date
Still the CRM manager; AI tunes send time, subject, variant
AI models compile and optimise
Maestro: product, timing, channel, tone, incentive per skill
Customer understanding
Declared attributes, RFM
Predictive scores on the same profile
Rich customer context
Traits inferred from purchases: household, routine, skin or hair type, life stage, cycle
Content
One creative, few variants
AI-written copy variants
GenAI-curated content and incentives
Payloads assembled from your approved assets and offers
Build cycle
60–90 days, 20+ people
Same build; faster copy
Days, 3–5 person pods
New skill live in days; no journey rebuild
Channel tools
Where the logic lives
Logic still in the tool; AI sold as a premium add-on
Delivery endpoints
Delivery endpoints; logic moves into Maestro
Measurement
Opens, clicks, campaign revenue
Lift on opens and clicks inside the tool
Engagement, conversion, CLV
Incremental revenue and retention per skill vs holdout
Team role
Build and operate campaigns
Still builds campaigns; AI assists
Strategists, curators, governors
Skill owners: objectives, guardrails, shelf, results
Exhibit 6
Build and delivery: who does what
Activity
Your team
Maestro
Existing stack
Commercial objectives
Sets and weights them; reviews quarterly
Optimises each skill against them
—
Research and insight
Defines the questions; validates findings
Infers customer traits and patterns from purchases
Supplies order and behaviour data
Skill design
Owns the business problem; approves the skill
Skill designed and fine-tuned with your team
—
Content and offers
Approves tone, assets, offer budgets
Selects and assembles from the shelf per customer
Content library, PIM, promo engine
Compliance and consent
Sets policy: brand, legal, contact limits
Enforces as hard limits on every decision
Stores consent; template checks
Experiment design
Sets the learning agenda
Runs holdouts per skill; reports incrementality
BI reads the results
1:1 decision
By exception only
Product, time, channel, tone, incentive for every eligible customer, daily
—
Collisions across skills
Sets priorities
Resolves automatically
—
Rendering and send
—
Sends execution-ready payload
ESP / MAP, app, site deliver
Escalation
Reviews high-value cases and shelf gaps
Flags gaps and edge cases
—
Per BCG, strategic objectives, budget allocation, shelf strategy and changes to agent autonomy stay with people and are never delegated.
Takeaway
- Maestro is an AI CRM Manager you hire: it reasons, remembers, decides and executes for every customer.
- Your team stops building and checking journeys, and starts owning objectives, skills, the shelf and results.
- Strategy, budget and how much autonomy Maestro has stay with your people.
The architecture
Maestro sits between the data you already have and the channels you already pay for. It replaces no system. It replaces the manual work of deciding.
Most retailers already own the pieces of this model: customer and order data, a product catalogue, brand rules, and channels that send email, push and onsite content. What is missing is the layer that decides what each customer should receive. This section shows where that layer sits, in two views: how data flows through Maestro to your channels (Exhibit 7), and what each system you already own does once Maestro is in place (Exhibit 8).
That is the starting point. Once decisions live in one layer, the channel tools only have to deliver, which changes what they are worth paying for. Section 05 covers that shift.
Exhibit 7
Data in, Maestro decides 1:1, existing tools deliver
1 · Data in
Retailer data
Already in place
2 · Maestro decides 1:1
Maestro
AI CRM Manager behind the stack
3 · Execution-ready output
Existing channel tools
Delivery endpoints
↺ Outcomes and holdout results flow back into every skill
Guardrails set by your team
- Brand
- Frequency
- Eligibility
- Budget
- Autonomy
Exhibit 8
What each system does in the new model
Layer
System
Role before
Role in the new model
Hands over / receives
Data
Warehouse, CDP, order and POS data
Store data; build campaign audiences
System of record; feeds customer, order and behaviour data
Sends orders, events, consents, IDs
Product
PIM, catalogue, stock and price feeds
Product blocks in templates
Product truth: attributes, availability, margin, price
Sends catalogue, stock, price, promo mechanics
Brand
Guidelines, content library, offer rules
Briefs and manual QA
The shelf: approved tone, assets and offers Maestro can use
Sends assets, offer budgets, exclusions
Intelligence
Maestro
Did not exist; logic lived in the MAP and in people’s heads
AI CRM Manager: Memory, skills, 1:1 sub-decisions, orchestration
Receives data; sends execution-ready decisions
Delivery
ESP / MAP / CEP, ecommerce site, app, CRM, clienteling, store
Journey builders, audience tools, premium AI modules, static banners, manual lists
Execution endpoints for the same decision: send it, show it on PDP or in the app, or hand it to staff for high-value cases
Receives who, what, when, channel and payload per customer
Measurement
BI, warehouse
Campaign reports
Skill-level incrementality; learning fed back
Receives decision log and outcomes per skill
Takeaway
- No rip and replace to start. Your MAP or CEP becomes a delivery endpoint, and one decision layer serves every channel.
- Once decisions leave the channel tools, those tools only need to deliver, which changes what they are worth paying for.
Cost and value
Replenit starts on the stack you already run and is measured on results from the first skills. Over time it changes what you need to pay your stack for.
Most engagement platforms price on volume: profiles stored, messages sent, events and data points processed, seats, plus premium AI modules on top. The bill grows with your customer base whether results grow or not. Once decisions are made in one intelligence layer, the channel layer only has to deliver. Delivery is a commodity: bulk email through a cloud API costs around $0.10 to $0.23 per 1,000 messages.
Exhibit 9
You pay for volume today; in the new model you pay for delivery and for outcomes
Cost driver
Typical martech pricing
In the new model
Profiles and contacts
Priced by profile tier, rising with the customer base
Customer Memory lives in the intelligence layer; delivery needs no profile tiers
Messages sent
Priced per message or volume band
Sent through API-driven delivery at cost per message
Events and data hits
Priced per event or data point processed
Processed once in Maestro, not billed again in every tool
Premium AI modules
Extra tiers per tool: predictive, send-time, agents
Decisions made once, for every channel
Operations
Team time to build and maintain journeys
Team time on objectives, skills and results
Decision logic and know-how
Locked in vendor journeys; rebuilt when you switch
Your skill library, portable across delivery vendors
Who carries the risk
You: the licence is paid whatever the result
Replenit: measured per skill against holdout
Delivery cost: AWS SES public pricing, $0.10 to $0.23 per 1,000 outbound emails by plan. Martech pricing models vary by vendor and contract.
Exhibit 10
No rip and replace now; a leaner, API-driven downstream over time
Now
Maestro on top of your stack
- Skills run through your current MAP, CEP, site and app
- Nothing replaced, nothing migrated
You pay for: current stack + outcomes
Next renewals
Logic leaves the channel tools
- Journeys, audience tools and premium AI tiers become optional
- Downgrade to core delivery tiers on your timing
You pay for: core delivery + outcomes
Long run
Centralised API-driven delivery
- Email, SMS and push through messaging platforms built for API sending (CPaaS, e.g. Vercom)
- Site, app and store read the same decisions by API
You pay for: delivery at cost + outcomes
Each step is optional and timed by the retailer, usually at contract renewal. Skills and Memory stay the same through every step.
Exhibit 11
Results in the first quarter; a stronger model every quarter after
Horizon
Outcome you see
What happens to the stack
What the team does
First quarter
2 to 3 skills live; incremental revenue measured against holdout
Unchanged; Maestro sends through current tools
Sets objectives, guardrails and shelf
First year
Skill library across the lifecycle, channels and brands
Journey and AI tiers reviewed at renewal
Moves from production to skill ownership
Long run
Skills compound as your IP; cost per outcome falls
Downstream centralised on API-driven delivery
Directs the model; grows the skill library
Exhibit 12
Replenit owns the outcome, not just the software
A tool you buy
A manager you hire
You configure it
You onboard it
You run campaigns with it
It owns the work
Judged on features
Judged on performance and P&L
Paid whatever the result
Accountable for the result
Adds to the team’s workload
Removes work from the team
What ownership means in practice
- Measured per skill. Every skill runs against a holdout, so the incremental result is visible, not attributed.
- Traceable. Every decision is logged with its reason and can be audited.
- Accountable for the number. Replenit is judged on the commercial outcome, not on usage or seats.
Takeaway
- No rip and replace today, and no reason to keep paying premium, volume-based prices for tools that only deliver. You set the timing, usually at renewal.
- Results are measured from the first skills, and Replenit is accountable for the outcome, not for usage.
Group-level decisioning
In a retail group, the same customer shops several brands or banners. When each brand decides alone, the shared customer is over-contacted and brands compete for the same wallet.
BCG flags the same failure: engines tuned to one product line overpromote it, bombard shared customers and trade lifetime value for short-term conversion. Maestro adds a group layer above the brands. Each brand keeps its own goals, shelf and tone. The group layer decides which brand acts next for each shared customer.
Exhibit 13
Three decision levels, one customer
Group
Group CRM / CDO
Sets
Group objectives, shared customer view, group-wide contact caps, priority between brands, group incentive budget
Maestro decides
Which brand speaks next, total contact pressure, cross-brand routing of a need
Brand
Brand CRM team
Sets
Brand goals, brand shelf and tone, brand skills, brand consent
- Fashion
- Beauty
- Pharmacy
- Home
Maestro decides
Which brand skill applies; product and offer from that brand’s shelf
Customer
Maestro
Reads
One Customer Memory across brands: what they buy where, for whom, and how often
Maestro decides
Act or wait, product, timing, channel, tone, incentive
Group rules sit above brand rules. A brand skill runs only inside the room the group layer leaves it.
Exhibit 14
Decisions only a group layer can make
Group decision
What Maestro weighs
Example
Which brand speaks next
Each brand’s live need for this customer and its value to the group
A due skincare refill outranks a fashion newness message this week
Total contact pressure
Contacts from every brand in the same window
One group-wide weekly cap instead of one cap per brand
Cross-brand need
A purchase at one brand that signals a need at another
Baby care at the pharmacy banner makes nursery items relevant at the home banner
Cannibalisation
Where the group earns more on the same need
Route a repeat purchase to the higher-margin banner, within brand rules
Shared incentive budget
Group budget vs brand budgets
Spend an incentive only where it changes behaviour group-wide
Consent and data use
Per-brand consent and group data agreements
Only brands with consent make contact; others inform the decision where the law allows
Exhibit 15
One shared customer, one week (illustrative)
Signals by brand
- Pharmacy: baby care purchases started
- Beauty: skincare refill due
- Home: browsed nursery furniture
Group sub-decisions
- Beauty speaks first: the need is due and certain
- Home waits: group cap reached this week
- Pharmacy: no promotion; buys at full price
Output per brand
- Beauty: refill reminder in the app
- Home: nursery range by email next week
- Pharmacy: no contact this week
Takeaway
- Brand-by-brand tools cannot see a customer shared across your brands.
- A group decision layer protects that customer from over-contact and stops your brands competing for the same wallet.
Skills and guardrails
A campaign is used once and archived. A skill is a reusable unit of commercial judgement that runs every day, learns from outcomes and stays with the retailer.
Exhibit 16
How a skill reaches a decision
Objective
Capture the repeat purchase before a competitor does
Macro sub-decisions
- Is this customer eligible today?
- Which skill has priority?
- Which commercial goal applies?
Micro sub-decisions
- Which product and pack size
- Which moment and channel
- Which tone; incentive or none
- Logged with its reason
Exhibit 17
Example skill library
Skill
Business objective
What Maestro learns and uses
Sub-decisions per customer
Household understanding
Grow share of the household basket
Which household member each purchase is for
Whose need comes first; product and framing for that member
Replenishment timing
Win the repeat before a competitor
Each customer’s own consumption cycle per product
When to remind; pack size; incentive or not
First to second purchase
Turn new buyers into repeat buyers
What the first basket says about need and price sensitivity
Which second product; timing; offer or no offer
Product-level churn prevention
Stop silent drop-off on one product
An active customer quietly stopping one recurring item
Churn or switch; win-back product and message
Newness and seasonality curator
Sell new and seasonal ranges
Fit between the customer’s profile and new or seasonal items
Who sees which new item, and when
Ready to subscribe
Move habitual buyers to autoship
Regularity and stability of repeat purchases
Who to invite; which product; which incentive
Skills belong to the retailer. Their data, guardrails and logic are never used to train models for, or serve, anyone else.
Exhibit 18
Guardrails your team sets once
Guardrail
Enforced as
Brand
Approved tone, assets and claims only
Frequency
Contact caps per customer, channel and week
Group
Group-wide contact cap and priority between brands
Eligibility
Consent, exclusions, category restrictions
Budget
Incentive budget per skill; no discount where not needed
Autonomy
What runs alone; what escalates
Exhibit 19
The team’s new rhythm
Cadence
What the team does
Weekly
Read skill results vs holdout; approve shelf assets; review escalations
Monthly
Add, retire or re-tune skills; adjust guardrails
Quarterly
Re-weight commercial objectives; plan where to grow the skill library
Takeaway
- Skills are your intellectual property, and they get better with every outcome.
- Guardrails mean autopilot is never unsupervised: control stays with your team.
Getting started
Adoption does not need a re-platform. It starts with one objective, a few skills and the channels already in place.
Pick the commercial objective
One measurable goal: repeat rate, second purchase, category growth.
Retailer
Connect data
Orders, customers, catalogue and consents from warehouse or ecommerce.
Retailer IT + Replenit
Select two or three skills
Where the retailer loses most value today.
Retailer + Replenit
Set guardrails and the shelf
Brand tone, contact caps, offer budget, approved assets.
Retailer
Connect the delivery endpoint
Maestro output into your ESP, MAP or CEP, app or site.
Retailer IT + Replenit
Run with holdout
Skills live; incrementality measured against a control group.
Replenit
Review and expand
Read results; add skills, channels and brands.
Retailer + Replenit
Before you decide
Five questions to ask your team
How many of our customer actions start from the calendar, and how many from the customer’s own moment?
BCG expects 70 to 80% of customer messages to be decided by AI for each customer, not sent as campaigns.
Where does our decision logic live today, and who maintains it?
If it sits in several tools and a few people’s heads, it does not scale and it is not owned.
What do we pay per profile, per message and per AI module, and what do those tools actually decide?
If they mostly deliver, the premium is paying for something else.
If a customer shops several of our brands, who decides which brand speaks next?
Without a group layer, the answer is usually nobody.
Which commercial know-how would we lose if a key CRM person left tomorrow?
Skills keep that know-how inside the business.
Takeaway
- Start with one objective, two or three skills and the channels you already have.
- Use the five questions to see where your team stands today.
The job is no longer maintaining a stack. It is building the skills that run it.
Keep your stack as the execution layer. Put the decisions in the hands of an AI CRM Manager that works inside your goals and guardrails.