What We Built
V1 was the simplest possible pricing problem: eight products, eight price parameters, demand modelled as a linear function of price with fixed elasticity coefficients. Each unit of demand generates revenue at the set price minus a fixed cost. The agent's job was to find the combination of eight prices that maximised total portfolio profit. No cross-product effects, no competitor dynamics, no customer segments. Just clean elasticity optimisation across a static demand curve.
The agent ran 4,691 experiments using Bayesian optimisation with a Gaussian Process surrogate. Baseline profit (midpoint prices for all products) was £428,000. Final profit was £577,000 — a 35.3% improvement. The surface was smooth and the optima clear. This version confirmed that autonomous pricing optimisation works under controlled conditions before we introduced the complexity that makes it interesting.
V2 introduced the single most important complication in real pricing: cross-price elasticity. The price you set for one product changes the demand for every other product in the portfolio. A premium widget at £90 suppresses demand for the standard widget at £35 if customers perceive them as substitutes. A budget widget at £6 cannibalises the standard widget's volume. The demand matrix became an 8×8 cross-elasticity tensor. Each price decision now has portfolio-wide consequences.
V2 ran 7,203 experiments. The additional complexity of cross-price effects slightly reduced the percentage improvement to 33.7%, but the absolute profit improvement was comparable — and the structural findings were considerably richer. The agent began discovering non-obvious pricing configurations where deliberately suppressing margin on one product improved total portfolio profit through volume spillover effects.
V3 extended V2 with three additional layers: customer segmentation (price-sensitive, quality-focused, value-seeking), seasonal demand variation (four seasons with different elasticity profiles per segment), and competitor pricing reactions. The parameter count jumped from 8 to 32 — four price variants per product to handle seasonal/segment combinations. The search space exploded. The agent had to simultaneously optimise prices that work across segments and seasons while anticipating competitor responses.
V3 ran 23,522 experiments — the largest single version — because the parameter space required it. The complexity penalty was real: profit improvement dropped to 29.2%. But the findings from V3 were qualitatively different. The agent discovered that segment-aware pricing (charging quality-focused customers more while maintaining volume with price-sensitive segments) consistently outperformed a single-price-fits-all approach, even after accounting for the increased optimisation difficulty.
V4 was the structural leap. Two additions changed everything. First, bundles: three pre-defined product combinations available at a discount, each with its own price parameter. Second, structural floor detection: the simulator explicitly modelled cliff thresholds — price points below which demand collapses discontinuously. The agent now had 99 parameters across 8 individual products and 3 bundles, and had to navigate discontinuous demand surfaces where smooth Bayesian optimisation breaks down near the cliffs.
V4 ran only 118 experiments — orders of magnitude fewer than V3 — yet achieved 77.5% profit improvement. The reason is structural: once the agent discovered the bundle economics and the loss-leader dynamic, a small number of high-quality experiments was sufficient to confirm and exploit the finding. Fewer experiments, larger improvement. The agent had found a fundamentally different part of the solution space.
The Numbers
| Version | Products | Parameters | Key Addition | Experiments | Baseline Profit | Final Profit | Improvement |
|---|---|---|---|---|---|---|---|
| V1 | 8 | 8 | Basic elasticity + cliffs | 4,691 | £428,000 | £577,000 | 35.3% |
| V2 | 8 | 8 | Cross-price elasticity | 7,203 | £428,000 | £572,000 | 33.7% |
| V3 | 8 | 32 | Segments + seasons + competitors | 23,522 | £428,000 | £553,000 | 29.2% |
| V4 | 8 + 3 bundles | 99 | Monthly + bundles + structural floors | 118 | £578,320 | £1,026,781 | 77.5% |
The apparent paradox in the table — V4 achieves the highest improvement with the fewest experiments — resolves once you understand what the agent discovered. V1 through V3 optimise prices within a fixed revenue architecture. V4 discovered a different revenue architecture entirely: one where bundles generate more than half of total profit and a loss-leader product at cost floor drives the bundle demand that makes everything else possible.
The 77.5% improvement is not the result of 118 experiments finding marginally better prices. It is the result of 118 experiments confirming a structural insight that required bundle mechanics and cliff-threshold dynamics to become visible. The insight couldn't exist in V1–V3 because those simulations didn't include bundles.
The Loss-Leader Discovery
The most counter-intuitive finding across all 35,534 experiments was the budget_widget loss-leader dynamic. In V4, the agent drove the budget_widget price progressively lower — below cost — and total portfolio profit increased at each step. This is not a modelling error. It is a structural property of markets with bundles and cross-price elasticity: a product priced at or below cost can generate positive portfolio profit by driving demand for high-margin companion products and bundles.
The mechanism works as follows. The budget_widget has a very high price-sensitivity coefficient. Small price reductions generate large volume increases. A significant fraction of budget_widget customers also purchase bundles that include high-margin products. The bundle discount makes the bundle attractive to customers who would not purchase the premium or standard widget individually. By driving budget_widget volume to its maximum (through cost-floor pricing), the agent maximised the pool of customers entering the bundle funnel, generating bundle revenue that more than offset the per-unit loss on the budget_widget itself.
The loss-leader effect was only discoverable because V4 included bundles. In V1–V3, pricing the budget_widget below cost would simply reduce portfolio profit proportionally. The bundle mechanism is what transforms a per-unit loss into a portfolio gain. This is why the structural findings from V4 cannot be approximated by running more experiments in V1–V3 — they require a different model architecture.
The quantitative result: driving the budget_widget from £8 (baseline midpoint) down to its cost floor at £3 increased total portfolio profit from £716,117 to £983,099 — a 37.3% gain from moving a single product's price. No other single-product price change in the entire V4 experiment set produced a comparable portfolio effect.
| Experiment | budget_widget Price | Total Portfolio Profit | Change vs Baseline |
|---|---|---|---|
| Baseline | £8.00 | £716,117 | — |
| Sweep 1 | £7.00 | £750,487 | +4.8% |
| Sweep 2 | £6.00 | £773,055 | +7.9% |
| Sweep 3 | £5.00 | £811,700 | +13.4% |
| Sweep 4 | £4.00 | £863,427 | +20.6% |
| Sweep 5 | £3.50 | £913,923 | +27.6% |
| Sweep 6 | £3.00 (cost floor) | £983,099 | +37.3% |
Cliff-Threshold Pricing: The Dominant Strategy
Six of the eight products in V4 ended up priced within 4% of their cliff threshold — the price point below which demand collapses discontinuously. This was not a coincidence or an artefact of the search algorithm. It reflects a structural property of markets with psychological price anchors: the maximum profit for a product with a demand cliff is almost always achieved by pricing as close to that cliff as safely possible.
The intuition: cliff-threshold markets have two demand regimes separated by a discontinuous boundary. Above the cliff, demand is reasonably elastic — price increases reduce volume, but gradually. Below the cliff, demand collapses — customers mentally re-categorise the product (from "reasonable" to "expensive", or from "premium" to "overpriced"). The profit-maximising price is therefore the highest price still in the high-demand regime, which is just at or slightly below the cliff.
The agent discovered this architecture independently, without being told about the cliff structure. It inferred the location of each cliff from the profit landscape and converged on cliff-threshold pricing for six products in the same run. The budget_widget was the exception — priced at cost floor as part of the loss-leader strategy. The premium_widget and standard_widget were the other exceptions — their optimised prices landed significantly above their cliffs, which the analysis identified as grid artefacts where the search had not fully explored that region of the parameter space.
| Product | Cost | Reference Price | Cliff Threshold | Optimised Price | Margin | Annual Demand |
|---|---|---|---|---|---|---|
| premium_widget | £25 | £80 | £80 | £70 | 64.3% | 8,760 |
| standard_widget | £12 | £40 | £40 | £35 | 65.7% | 10,512 |
| budget_widget | £3 | £8 | £3 (cost floor) | £3 | 0% | 52,560 |
| pro_gadget | £30 | £90 | £80.56 | £80 | 62.5% | 4,380 |
| basic_gadget | £10 | £32 | £28.17 | £28 | 64.3% | 8,760 |
| organic_essential | £8 | £25 | £20.10 | £20 | 60.0% | 13,140 |
| standard_essential | £5 | £18 | £14.54 | £14 | 64.3% | 17,520 |
| luxury_item | £50 | £200 | £202.63 | £200 | 75.0% | 2,190 |
Bundle Economics
The most significant structural change between the baseline and the optimised solution was not individual product pricing — it was the transformation of the profit architecture. In the baseline configuration, bundles accounted for 18% of total profit. In the optimised configuration, bundles accounted for 56% of total profit. The total bundle profit grew from £103,933 to £577,122 — a 455% increase in absolute terms.
This rebalancing happened because the loss-leader strategy on budget_widget dramatically increased the volume of customers entering the bundle funnel. More customers purchasing budget_widget meant more customers eligible for bundles that included budget_widget as a component. Bundle discounts made those combinations attractive relative to individual purchases. The agent discovered that sacrificing individual product margin to drive bundle volume is the dominant strategy when bundles carry higher per-customer revenue than individual purchases.
| Revenue Stream | Baseline | Optimised | Change |
|---|---|---|---|
| Individual product profit | £474,387 | £449,659 | −5.2% |
| Bundle profit | £103,933 | £577,122 | +455% |
| Total portfolio profit | £578,320 | £1,026,781 | +77.5% |
The individual product profit actually declined slightly in the optimised solution — from £474,387 to £449,659. This is a direct consequence of the budget_widget loss-leader strategy: zero margin on a high-volume product reduces individual profit. The gain comes entirely from the 455% increase in bundle profit. A business evaluating this optimisation by looking only at individual product margins would incorrectly conclude that performance had declined.
| Bundle | Products Included | Bundle Discount | Total Bundle Profit | Net Contribution |
|---|---|---|---|---|
| starter_bundle | budget_widget + basic_gadget + standard_essential | 10% | £189,450 | +£152,380 |
| value_bundle | standard_widget + organic_essential + basic_gadget | 12% | £221,890 | +£178,920 |
| premium_bundle | premium_widget + pro_gadget + luxury_item | 8% | £165,782 | +£145,822 |
Five Structural Findings
Across all four versions and 35,534 experiments, five structural patterns repeated with enough consistency to warrant treatment as generalisable findings rather than simulation-specific observations.
Portfolio Pricing Is Not the Sum of Individual Optimal Prices
The price that maximises profit for a single product in isolation is almost never the price that maximises portfolio profit when cross-price elasticity is present. V2 demonstrated this clearly: the individually optimal price for several products produced lower portfolio profit than a configuration where some products were deliberately under-priced to drive volume for higher-margin companions. Treating each product as an independent pricing problem is structurally incorrect in any catalogue with meaningful cross-price effects.
Bundle Mechanics Create a Qualitatively Different Optimum
The addition of bundles in V4 did not simply add a few more parameters to the existing optimisation problem. It created an entirely different solution landscape with a qualitatively different optimal strategy. The best solution in a no-bundle world is a locally reasonable set of individual prices. The best solution in a bundle-enabled world is a counter-intuitive configuration — loss-leader pricing on one product — that would look like a mistake in a no-bundle framework. Businesses evaluating pricing optimisation without modelling their bundle economics will find local optima that are globally sub-optimal.
Cliff-Threshold Markets Have a Dominant Pricing Posture
In markets where demand has discontinuous cliff thresholds (psychological price anchors, tier boundaries, reference price effects), the profit-maximising price is consistently at or just below the cliff, not at the unconstrained elasticity optimum. This means that for six of eight products in the V4 simulation, the correct strategy was not to maximise the price-volume tradeoff along a smooth curve, but to identify the cliff and price just below it. The cliff is the dominant structural feature of the demand surface — finding it matters more than optimising the curve above it.
Segment Complexity Reduces Optimisation Efficiency but Not Output Quality
V3's introduction of customer segments, seasonal variation, and competitor reactions increased experiment count by 3.3× compared to V2 while reducing percentage improvement from 33.7% to 29.2%. However, the absolute profit levels and the structural insights about cross-price effects remained valid. The reduction in percentage improvement reflects not a worse agent, but a harder problem: a 32-parameter space with non-stationary demand is genuinely more difficult to optimise than an 8-parameter stationary space. The implication for deployment: more complex pricing environments require more experiments, but the quality of the structural findings does not degrade proportionally.
The Value Is in the Map, Not the Prices
The specific prices recommended by the V4 agent — £3 for budget_widget, £80 for pro_gadget, £200 for luxury_item — are valid for the synthetic demand parameters used in the simulation. Real markets have different elasticity coefficients, different cliff locations, different cross-price effects. What transfers across real markets is the structural map: which products have cliff thresholds and where they are, which products are candidates for loss-leader treatment, how bundle economics change the optimal individual pricing strategy, and what the shape of the profit surface looks like near the optimum. This map is the durable output of the optimisation process. The specific prices are transient.
Structural Floor Analysis
The structural floor analysis classifies each product by its pricing position relative to its cliff threshold. Red classification indicates the product is either at its cost floor (loss-leader by design) or priced within 4% of its cliff threshold. Yellow classification indicates the product's optimised price is more than 4% above its cliff — either because the agent found a higher-profit configuration or because the grid search did not fully explore the cliff region.
The dominance of Red classifications is not a cause for concern — it is the expected output of a correctly-functioning optimiser in a cliff-threshold market. Six products at cliff, one at cost floor, and two Yellow artefacts is the fingerprint of an agent that has correctly identified cliff-threshold pricing as the dominant strategy and exploited it across the portfolio.
Industry Applications
The structural findings from this simulation are not specific to the synthetic product catalogue used. The mechanisms — cliff-threshold demand, cross-price elasticity, bundle economics, loss-leader dynamics — are present in most multi-product markets. The following industry contexts have directly analogous structures.
E-Commerce
Multi-SKU online retailers face exactly the portfolio pricing problem modelled in V2–V4. Cross-price effects between product variants (size, colour, tier) are real and measurable from purchase data. Bundle mechanics are directly implementable as multi-item promotions or "frequently bought together" pricing. Cliff thresholds correspond to psychological price points (£9.99 vs £10.00, £49 vs £50) that are well-documented in retail pricing research. The loss-leader dynamic is already used implicitly by sophisticated retailers — the optimisation adds structural rigour to what is currently done by intuition.
SaaS
SaaS pricing tiers are a direct analogue of the cliff-threshold structure. The difference between a Free, Starter, and Professional tier is not a smooth demand curve — it is a stepped structure with discontinuous adoption boundaries at each tier transition. Bundle mechanics correspond to add-on modules or seat packages. The loss-leader equivalent is the free tier, which drives bundle revenue through upsell and expansion. The V4 finding that optimal portfolio profit requires deliberately pricing one product at zero margin has a direct SaaS parallel in freemium architecture.
Grocery and FMCG
Grocery pricing is one of the most extensively studied cliff-threshold markets. Price points for staple goods (bread, milk, eggs) are well-known to have discontinuous demand cliffs at round-number thresholds. Cross-price elasticity between private-label and branded alternatives is measurable and significant. Bundle mechanics correspond to multi-buy promotions and meal-deal configurations. The loss-leader finding — that pricing a high-volume, high-elasticity product at cost floor to drive associated product sales — is standard practice in category management, but typically applied by intuition rather than through systematic portfolio optimisation.
Luxury and Premium
Luxury markets invert the standard elasticity assumption: demand for some products increases with price (Veblen goods) up to a threshold, then collapses if the price signals insufficient exclusivity. This is a different cliff structure — an upper cliff rather than a lower cliff — but it is modelled by the same cliff-threshold architecture. The luxury_item in V4 (£200 optimised price, cliff at £202.63) demonstrates this: the correct strategy is to price just below the cliff where exclusivity signals are intact, not to maximise margin by pushing above it.
B2B and Enterprise
B2B pricing has structural cliffs at procurement approval thresholds. Purchases under £10,000 may be approved at a manager level; purchases over £10,000 require director sign-off; purchases over £50,000 require board approval. These approval thresholds create demand cliffs as predictable as the psychological price points in consumer markets. Bundle mechanics correspond to contract structures and volume discounts. The cross-price effects between products in an enterprise software or services portfolio are real and frequently ignored in standard B2B pricing practice.
Capital Allocation Implications
The V4 findings have direct implications for how businesses should think about the relationship between pricing, product investment, and revenue architecture.
Bundle infrastructure is a high-return investment. The 455% increase in bundle revenue in V4 was made possible by having functional bundle mechanics in the simulator. In a real business, this corresponds to the systems, inventory, and fulfilment capability to offer multi-product bundles reliably. The V4 results suggest that the marginal return on bundle infrastructure investment — measured as the incremental profit available from bundle-enabled pricing optimisation — is likely to exceed the return on standard product margin improvement work. A business without bundle capability cannot access the optimum that V4 found.
Loss-leader products require deliberate capital coverage. The budget_widget at cost floor generates zero margin per unit. At 52,560 units of annual demand, that is a significant volume of product moving at no individual contribution. The capital required to fund that inventory, absorb the per-unit loss, and carry the working capital load must be explicitly budgeted. Businesses that optimise pricing without modelling the working capital implications of high-volume zero-margin products will discover the capital constraint before they discover the portfolio benefit.
The cliff map is a strategic asset. Knowing where demand cliffs are located across your product portfolio is more valuable than knowing the optimal price at any given moment. Prices need to be re-optimised as costs, competitor prices, and customer preferences change. The cliff locations — which correspond to stable psychological anchors, procurement thresholds, or category perception boundaries — change much more slowly. A business that has mapped its cliff structure has a durable strategic asset that remains valid across multiple pricing cycles.
Segment-aware pricing generates structural margin uplift. V3's finding that segment-aware pricing (different price/value propositions for price-sensitive vs quality-focused customers) outperforms single-price strategies has direct capital allocation implications. Building the capability to deliver segment-differentiated pricing — whether through tiered products, channel-specific pricing, or customer cohort identification — generates structural margin uplift that compounds across the customer base. The optimisation complexity is higher, but so is the return.
Summary
Across 35,534 experiments and four simulation versions of increasing complexity, an autonomous pricing agent achieved profit improvements ranging from 29.2% to 77.5% above baseline. The headline result — 77.5% improvement in V4 — was not the product of marginal price optimisation. It was the product of discovering a different revenue architecture: one where bundle revenue exceeds individual product revenue, and where a loss-leader product at cost floor is the structural enabler of that bundle revenue.
The five structural findings — portfolio interdependence, bundle architecture discontinuity, cliff-threshold dominance, segment complexity scaling, and structural map durability — are not specific to the synthetic parameters used in this simulation. They reflect underlying market structures that are present in any multi-product business with cross-price effects, bundle capability, and demand cliffs.
The cliff-threshold finding deserves particular emphasis. Six of eight products ended up priced within 4% of their demand cliff. This was not a coincidence. In cliff-threshold markets, the cliff is the dominant feature of the demand surface — finding it and pricing against it is more valuable than optimising any other aspect of the price function. Most pricing models smooth over discontinuities because they are mathematically inconvenient. This is the wrong abstraction. The discontinuities are the information.
The agent's value in this framework is not in producing a price list. It is in producing a structural map: cliff locations, bundle economics, loss-leader candidates, cross-price dependencies. That map is the durable output of the optimisation. The specific prices are its current expression.
The agent's value is in the structural map it produces — cliff boundaries, bundle economics, loss-leader conditions — not in the specific prices it recommends. Interested in applying this to your pricing? Start an optimisation enquiry →