Dynamic Pricing in the Age of AI Commerce
PRICING AUTOMATION DECISION
Automate a pricing policy, not an unconstrained target
Dynamic pricing should implement a rule the business can explain: which products may change, which inputs are permitted, how often, within what limits and with whose approval. An optimisation model without those boundaries can create customer, margin, legal and competition risk.
01 – PURPOSE
Choose one objective
Specify whether the rule addresses expiry, capacity, seasonal clearance, supplier cost or another measurable business condition.
02 – CONSTRAIN
Set hard boundaries
Protect floor price, advertised terms, customer classes, update frequency, evidence quality and approval thresholds.
03 – MONITOR
Watch real outcomes
Track margin, conversion, cancellations, complaints, overrides, stock exposure and unequal effects rather than revenue alone.
Pilot one transparent, low-risk rule
Choose a narrow product group, document the policy and legal review, simulate historic outcomes and require approval before any live price changes.
Dynamic pricing changes an offer in response to defined conditions. Those conditions could include time, stock age, demand, capacity, supplier cost or a published customer segment. AI may forecast or recommend a price, but the business remains responsible for what customers see and pay. The safest starting point is a narrow commercial problem with explicit limits, not a general instruction to maximise revenue.
Use the ecommerce integrations service for controlled workflow design, the ecommerce service for catalogue and checkout implementation, and the future-commerce resource hub for related governance. Pricing, consumer-protection and competition obligations require qualified legal review for the actual market and policy.
1. Write the pricing-policy canvas
| Policy element | Decision to record |
|---|---|
| Scope | Products, channels, customers and locations included or excluded |
| Objective | One primary outcome and the customer value rationale |
| Inputs | Approved fields, source, freshness and quality tests |
| Boundaries | Floor, ceiling, step size, frequency and protected periods |
| Authority | Recommendation, approval or automatic execution level |
| Recovery | How to stop, roll back, honour or correct an affected offer |
Avoid combining clearance, margin recovery, competitor response and personalised offers in the first policy. Different objectives need different safeguards and evaluation. State what success would look like and what evidence would cause the rule to be suspended.
2. Establish the lawful and customer-facing basis
South Africa’s Consumer Protection Act is an official starting source for obligations affecting consumer transactions and displayed offers. Competition law may also be relevant when algorithms use competitor data or interact across markets. The Competition Commission has published discussion of digitalisation and algorithmic pricing. Do not convert this guide into a legal conclusion; obtain advice on price display, promotions, discrimination, collusion risks, data use and sector-specific rules.
3. Approve and test every input
- Current landed or replacement cost from the controlled finance record
- Available and ageing stock after reservations and quality holds
- Capacity or booking availability with a reliable timestamp
- Published promotion dates and eligibility rules
- Competitor information collected through an approved, lawful method
- Customer segment only where purpose, transparency and fairness are reviewed
Reject stale, missing or impossible inputs rather than allowing a model to fill the gap. A sudden cost of zero, negative stock or duplicated competitor offer should stop the recommendation. Keep input lineage so an unexpected price can be reconstructed.
4. Set commercial guardrails
| Guardrail | Example control |
|---|---|
| Margin | Never recommend below the approved economic floor without named approval |
| Movement | Limit percentage and number of changes per period |
| Consistency | Revalidate cart and checkout under the published terms |
| Fairness | Prohibit protected or unjustified proxy variables |
| Evidence | Expire a recommendation when its inputs are no longer current |
| Authority | Route high-impact changes to finance or commercial review |
5. Simulate before publishing
Replay the proposed rule against historic records without assuming the historic outcome will repeat. Inspect extreme prices, frequent changes, loss-making orders, stockouts, promotion conflicts and effects across customer groups. Add synthetic tests for missing cost, zero stock, sudden demand, bad competitor data and an integration outage. A model that performs on average can still fail dangerously at the edge.
6. Preserve explanation and override
For each recommendation, store the policy version, inputs, timestamp, proposed price, current price, constraint results, approver and final action. The explanation should name the real driver rather than generating a plausible story after the fact. Give authorised users a reasoned override and analyse override patterns: repeated rejection may show that the rule misunderstands a product or business constraint.
7. Monitor customer and operational outcomes
Measure gross margin, units, conversion, cancellations, returns, complaints, abandoned carts, stock ageing and manual interventions. Compare customers or channels where the policy could create unequal results. Review whether displayed and charged prices remain consistent. Sample individual price journeys as well as aggregate results, because a stable average can hide a short-lived extreme offer or repeated changes affecting one customer group. NIST AI RMF can support governance and risk review, but it is not a substitute for consumer, competition or sector-specific legal advice.
Worked pilot
A retailer has a small set of seasonal products with confirmed end dates and high remaining stock. The policy may recommend one markdown step each week within an approved floor, but cannot use individual customer behaviour or competitor scraping. Finance confirms the cost basis; commercial staff approve every recommendation; checkout validates the current published price; and the business honours completed orders. The team monitors margin, sell-through, complaints and overrides. Only after the policy behaves predictably does it consider a second product group or more frequent automation.
Sources checked
- South African Government: Consumer Protection Act
- Competition Commission South Africa: Digitalisation and algorithmic pricing discussion
- NIST: Artificial Intelligence Risk Management Framework 1.0
Reviewed by
Mitrend Digital editorial team
2026-07-17
Evidence used for this page
Reviewed against South Africa’s Consumer Protection Act source text, Competition Commission material on digitalisation and algorithmic pricing, and NIST AI RMF guidance. Includes an original pricing-policy canvas and controlled pilot.
Turn the guide into a practical next step
This resource provides general implementation guidance. Verify platform settings, tax, legal, payment and operational requirements against the current business context before making a live change.