Autonomy framework · Reliance Retail × Fynd

01 · The premise

Retail is an inventory risk business. 85-97% of every rupee a retailer earns goes toward acquiring, storing, moving, funding, and selling inventory. The remaining 3-15% is profit — the reward for getting the bet right. This is not a metaphor. It is the arithmetic of every annual report from Inditex to DMart to Reliance Retail. Revenue is selling inventory. COGS is buying inventory. Rent is storing inventory. Employees manage inventory. Everything else is a derivative.

Today, humans place this bet. They forecast demand, select suppliers, set prices, plan promotions, allocate stock, and decide markdowns. AI has entered the conversation — but in most organisations, it is still a surface-level addition. The spreadsheet is better-formatted. The PowerPoint is faster. The forecast has one more input. The bet itself has not changed.

Autonomous Commerce changes this. Not incrementally. Structurally. It is a framework for deciding, per workflow, the right level of autonomy — what should be fully autonomous, what should be AI-assisted, and what should remain human. The question is not “how do we use AI in retail?” The question is “at what level of autonomy should each workflow operate, and what must be true to move it to the next level?”

The approach covers both technology and organisation. On the technology side, the platform connects AI to every workflow and data stream — demand signals, inventory positions, supplier networks, pricing engines, and fulfilment nodes become a single addressable surface. The data catalog (Paper 2) documents the 67 activities, 120 SOPs, and 35 data sources that power these workflows. On the organisation side, each function must make a deliberate choice about its target autonomy level — and the people freed by that transition are redirected into AI-native roles that did not exist before.

AI has entered the conversation — but in most organisations it is still a surface-level addition. The bet itself has not changed.

02 · The risk gates · guardrails

The transition between autonomy levels is not a technology upgrade. It is a risk reduction proof.

Every retail workflow sits at some level of autonomy today. Moving it to the next level requires answering one question: does this move demonstrably reduce risk? Not “can AI do this?” — that is a technology question. Not “will this save headcount?” — that is an efficiency question. The question is whether giving the machine more authority over this workflow produces measurably better outcomes with measurably lower risk. If you cannot prove it, you are not ready. If you can prove it but the organisation will not act on the proof, the problem is not technology. It is governance.

Risk manifests differently across the value chain. Four domains, four categories of risk, one standard: here is the data, here is the A/B test, here is the measurable outcome in rupees. This rule governs everything that follows.

Plan

Inventory holding risk

The risk of building or holding stock the customer will not buy at full price.

Buy

Procurement risk

The risk of paying the wrong price, picking the wrong supplier, or committing capital too early.

Move

Supply chain risk

The risk of stock reaching the wrong store at the wrong time, or not at all.

Sell

Revenue risk

The risk of leaving margin on the table — under-pricing, over-discounting, missing the moment.

03 · The autonomy ladder · six levels

Six levels of autonomy, adapted from the Autonomous Networks taxonomy standardised by TM Forum for telecommunications. The domains change — from network fault management to demand forecasting, from radio access to supply chain — but the architecture of autonomy is universal. Each level defines who decides, who executes, and who bears risk.

L0Manual

Manual Commerce

Humans decide · humans execute · humans bear risk

Humans execute all tasks. Systems are record-keeping tools. Monitoring and data capture only. All risk borne by human judgement and experience.

L1Assisted

Assisted Commerce

Humans decide · AI accelerates · risk unchanged

Humans make all decisions. AI performs specific, repetitive subtasks on pre-configured rules — report generation, data formatting, alert triggering. Risk unchanged. Execution speed improved.

L2Augmented

Augmented Commerce

Humans decide within AI recommendations · risk reduced in narrow domains

Humans own decisions but operate within AI-generated recommendations. Closed-loop execution for narrow, well-defined units. Risk reduced in narrow domains with proven rule sets.

L3Intelligent

Intelligent Commerce

Humans set policy · system self-optimises within guardrails

Humans set policies and handle exceptions. The system senses real-time changes and self-optimises within defined boundaries — adjusting to demand shifts, competitive moves, and supply disruptions. Risk actively managed by the system within guardrails.

L4Agentic

Agentic Commerce

Humans set intent · system coordinates across Plan / Buy / Move / Sell · risk pre-empted

Humans set strategic intent and review outcomes. The system makes cross-domain decisions via predictive analysis and continuous learning. Coordinates across Plan, Buy, Move, and Sell simultaneously. Risk predicted and pre-empted.

L5Autonomous

Autonomous Commerce

Humans define purpose · system runs the lifecycle · risk structurally minimised

Humans define purpose, values, and constraints. The system operates the full lifecycle via cognitive self-adaptation — self-healing, self-optimising, continuously learning. Risk is structurally minimised.

The transitions · what each jump actually is

L1 → L2

Tool upgrade. The same people with better instruments.

L2 → L3

Skill upgrade. The same people doing the same jobs faster and with higher leverage.

L3 → L4

Identity shift. The roles themselves dissolve, and the organisation becomes a fundamentally different kind of entity.

04 · Not every workflow should reach L5

This is a design choice, not a failure of ambition. Workflows that are high-frequency, data-rich, and reversible — demand forecasting, pricing, replenishment, allocation — move toward L4-L5. Workflows that are low-frequency, judgement-intensive, and irreversible stay deliberately at lower levels. The right answer is not a uniform target. It is a portfolio of autonomy levels — chosen deliberately, per workflow, governed by provable risk reduction.

L1 – L2

Strategy & vision

What categories to enter, what the brand stands for. AI-informed, human-decided.

Brand & creative direction

AI-generated, human-curated.

L2 – L3

Compliance & ethics

Compliance and ethical judgements stay AI-monitored, human-accountable.

05 · The 90-day target matrix · across the F&L value chain

The matrix maps retail workflows against the six autonomy levels. Read it left to right to see how a single workflow evolves from fully manual to fully autonomous. Read it top to bottom within a value-chain stage to see the breadth of autonomy decisions an organisation must make simultaneously. The shaded cells show the target autonomy state in the next 90 days.

Autonomy framework matrix · Plan and Buy

[1/2] · Plan (Demand Forecasting · OTB Budgeting · Range Architecture · Space Planning · Assortment Planning) · Buy (Sourcing · Costing · Product Development · Order Placement · Vendor Management)

Autonomy framework matrix · Move and Sell

[2/2] · Move (Production Tracking · Quality Assurance · Allocation · Replenishment · Logistics) · Sell (Store Operations · Visual Merchandising · Pricing & Markdowns · Promotions · Omnichannel · Marketing)

Across 21 core retail workflows, L2 (Augmented) is the baseline — every workflow reaches at least this level. Of these, 43% move to L3 (Intelligent) where the system optimises within guardrails and humans handle exceptions. 29% reach L4 (Agentic) where the system makes cross-domain decisions and humans set intent. The remaining 24% operate at L2 where AI recommends and humans decide.

06 · The full workflow table · per-cell capability descriptions

Every cell describes what that workflow looks like at that level of autonomy. Verbatim from the paper. The shaded cells are L0–L3, the workflow's current state for most stages today; clear cells are L4–L5, the targets.

#	Workflow	L0Manual	L1Assisted	L2Augmented	L3Intelligent	L4Agentic	L5Autonomous
Plan
1.1	Demand Forecasting	Spreadsheets and instinct	AI summarises trends	ML forecasts, human reviews	Continuous signal engine	Forecast triggers actions	Demand is sensed, not forecast
1.2	OTB Budgeting	Excel, quarterly cycle	AI formats the deck	AI proposes OTB based on sell-through, planner reviews	Auto-reallocating constraint	Cross-format OTB, human sets intent	Real-time financial envelope
1.3	Range Architecture	Hand-built option plans	AI generates trend reports	AI-generated options, buyer chooses	System generates full range plans	Human scans and asks “what if?”	Signal to order in days
1.4	Space Planning	Fixed planograms, regional clusters	AI suggests layouts	Weekly GMROF optimisation	Continuously optimised per store	Daily VM briefs per store	Every store is a market of one
1.5	Assortment Planning	Periodic, cluster-level	AI surfaces bestsellers	Store-cluster, AI recommends	Store-level, demand-driven	Cross-format cannibalisation detection	Unified demand sensing
Buy
2.1	Sourcing	Personal network, trust-based	AI benchmarks prices	System scores and recommends	Real-time supplier graph	System selects, negotiates, orders	Autonomous within guardrails
2.2	Costing	Instinct-based negotiation	System tracks cost indices	Target cost per SKU pre-negotiation	Predicted landed cost pre-spec	Predictive, real-time margin	System-to-system pricing
2.3	Product Development	Multi-month physical cycle	AI generates concepts	Digital prototyping, partial	Majority digital approval	Spec-to-order in days	Generative design, direct-to-production
2.4	Order Placement	Bulk order, over-ordering	Forecasts inform quantities	Partial order, reserve capacity	Rolling orders, weekly adjustment	Trend-to-order, human reviews	Continuous, no order window
2.5	Vendor Management	Quarterly reviews, spreadsheets	Auto-generated scorecards	System flags at-risk vendors	Predictive risk identification	System manages operationally	Self-governing ecosystem
Move
3.1	Production Tracking	Email and phone calls	AI monitors milestones	Digitised milestone tracking	Predictive completion modelling	Delay predicted, mitigation auto-identified	Orchestrated, not tracked
3.2	Quality Assurance	Warehouse inspection, high defects	Risk-scored inspection triggers	Selective QC replaces blanket	Pre-production confidence scoring	Quality built in, not inspected	Zero-defect, system-to-system
3.3	Allocation	Once per cycle, A/B/C grading	AI considers store-level demand	SKU-store level, one-time	Continuous, demand-signal driven	Chase and replenish are one action	Inventory is a network
3.4	Replenishment	Manual, weekly, store calls	Auto-trigger, static thresholds	Smart slow-seller / fast-seller	Dynamic, real-time thresholds	Self-healing supply chain	Nervous system, autonomous
3.5	Logistics	Carrier negotiation, 3PL routing	Optimised routing, TAT tracked	Ship-from-store, intelligent routing	Cross-network path optimisation	Auto-rerouting on disruption	Invisible, self-optimising
Sell
4.1	Store Operations	Weekly visits, monthly reviews	Real-time dashboards	Same-day KPI deviations	Every store a market of one	Daily briefs, guardrails not instructions	Demand-sensing node
4.2	Visual Merchandising	Brand guidelines, periodic	AI recommends placement	Data-driven display rules	Daily recommendation per store	Autonomous daily planograms	Continuous, every fixture optimised
4.3	Pricing & Markdowns	Calendar-based clearance	SKU-level markdown recommendations	Dynamic, product-driven timing	Continuous, store-specific	Every SKU, every store, every day	Price is a real-time function
4.4	Promotions	Central, uniform across stores	AI recommends where to promote	Cluster-level promotion variation	Opportunity-detected, instant approval	Continuous micro-campaigns	No campaigns, individual conversations
4.5	Omnichannel	Online and offline separate	Unified inventory	Intelligent fulfilment routing	Single pool, channel-agnostic	Cross-format lifecycle management	Shopping disappears as effort
4.6	Marketing	Calendar-based campaigns, broadcast	AI writes copy, resizes images	Data-driven segmentation, automated bidding	Micro-campaigns triggered by real-time signals	Autonomous audience, creative, deploy, measure	Continuous individual conversations

Verbatim from Operationalizing Autonomous Commerce for Retail · §1.4 · pp. 3-5. The paper extends the table to Asset (5.1-5.4), Finance (6.1-6.3) and HR (7.1-7.3) — not shown here as they sit outside the F&L Plan/Buy/Move/Sell value chain Impetus targets.

07 · How we measure progress

Three measurements decide whether a workflow is ready to move to its next level. Each one carries a rupee number. None of them is a feature count.

Measurement 01

Risk reduction

Per the four risk gates — how much inventory, procurement, supply chain, or revenue risk does the workflow now actively avoid? Quoted in rupees of capital at risk, not in tickets closed.

Measurement 02

Throughput uplift

Decisions per hour at the new level vs the old. The L2 → L3 jump is a skill upgrade — same people, faster cycle, higher leverage. We measure that leverage directly.

Measurement 03

Defect / anomaly rate

What share of the system's decisions get reversed, escalated, or written off? L3 holds at low single digits within guardrails. L4 must beat that rate before the org accepts the identity shift.

The matrix is a workflow tool. Its deepest consequence is organisational. Once you set the target level per workflow, the org chart follows — because every role is defined by the workflows it touches. The move from L2 to L3 does not change the org chart. The planner is still a planner, the buyer is still a buyer — faster, better-informed, operating at higher leverage. The move from L3 to L4 changes what the organisation is. Planning teams compress from dozens to a handful of strategic minds who set intent and handle exceptions. Buying teams become acquisition strategists who manage supplier relationships and review AI-generated decisions. Supply chain teams become strategic operators who handle the disruptions that require human judgement.

Not everyone will step up. The honest reality is that L4 demands a capacity and ambition that not every person on every team possesses. The commitment is that the organisation will redirect, not displace — into work that did not exist before. The condition is equally clear: its people must become AI-native.

From human-drivento autonomous.