1. Executive Summary

Every automation platform that uses artificial intelligence carries an operating cost that scales with usage. How that cost behaves as a franchisee grows from five restaurants to fifty is not a detail of implementation — it is the single factor that determines whether the price quoted at signing is the price still being paid two years later.

This paper explains the architectural choice behind Edgebrook’s Automation Framework and why that choice produces a stable, predictable monthly cost. The Framework executes its workflows deterministically: the navigation, extraction, and import steps run as defined, repeatable code paths against Sculpture Hospitality and the connected point-of-sale system. Artificial intelligence is invoked narrowly and only where genuine interpretation is required, such as reading a scanned distributor invoice.

The prevailing alternative in the current market is the agentic pattern, in which an AI model observes each screen and decides what to click. That approach is flexible, and for some problems it is the right tool. For the recurring, high-volume, structurally stable workflows that define Sculpture operations, even a well-optimized agentic implementation consumes more than twenty-five times the execution cost of the deterministic approach for identical output.

The core finding. Industry-wide, the per-token price of AI has fallen roughly 88 percent since 2023. Over the same period, enterprise AI bills have risen sharply. Falling prices are being absorbed, and then exceeded, by rising consumption. Architecture, not vendor pricing, is what determines which side of that trend a platform lands on.

Edgebrook’s tier pricing is set as a flat monthly figure with a per-restaurant rate and a hard ceiling at the highest tier. That structure is only sustainable because the underlying execution cost per restaurant is small, stable, and does not grow with the length or complexity of a workflow run. This paper sets out the evidence for that claim and the methodology behind it.

2. The Cost Creep Problem

A pattern has emerged across enterprise deployments of AI-driven automation over the past two years, and it is counterintuitive enough to be worth stating plainly.

2.1 Prices are falling

The cost of a single unit of AI processing, measured per million tokens, has dropped further and faster than almost any comparable technology input. A tracked index of frontier model pricing sits approximately 88 percent below its early-2023 baseline, with roughly 80 percent of that decline occurring in the twelve months between early 2025 and early 2026. Competition among providers is intense and inference efficiency continues to improve. There is no present indication that this trend is reversing.

2.2 Bills are rising anyway

Despite that decline, organizations deploying AI automation report their spending increasing year over year. Published survey data places the average enterprise monthly AI expenditure at approximately eighty-five thousand dollars, up more than a third from the prior year, with a majority of enterprises reporting margin erosion tied directly to AI workload costs.

The explanation is consumption. Agentic workflows have been measured at consuming between five and thirty times the processing volume of a conventional AI interaction for the same completed task. An agent that reasons iteratively, observes results, and self-corrects may issue dozens or hundreds of model calls to accomplish what a defined script accomplishes in one pass. The unit that appears on the invoice is the task, not the token, and a cheap token multiplied by an unbounded number of calls is not a cheap task.

Why this catches organizations by surprise The multiplier is invisible during a pilot. A proof of concept running two restaurants for three weeks produces a bill small enough that nobody examines its structure. The cost profile only becomes visible at production scale — by which point the architecture is committed and the contract is signed.

2.3 What this means for a franchisee

A franchisee evaluating an automation platform is, whether or not the proposal says so, making a bet on that platform’s cost trajectory. If the vendor’s execution cost per restaurant rises with usage, one of three things follows: the vendor raises prices, the vendor imposes usage caps that were not discussed at signing, or the vendor absorbs the loss until the arrangement becomes untenable. None of those outcomes is good for the operator who built a staffing plan around the original quote.

3. Two Architectures

The workflows at issue are consistent across the Sculpture franchise network: Product Mix import from the point-of-sale system, distributor invoice ingestion, variance reporting, and period close. These can be automated two fundamentally different ways.

3.1 The deterministic approach

Navigation is defined in code. The Framework knows that the Product Mix report lives at a specific location, that the export control carries a specific identifier, and that the resulting file requires a specific set of column transformations before Sculpture’s importer will accept it. Each step executes the same way every time. An AI model is called only where the task requires genuine interpretation rather than navigation, and each such call is a single, bounded request.

3.2 The agentic approach

The model is given a goal and a browser. At each step it receives a screenshot of the current screen, reasons about what it sees, and decides what to click. It then receives the resulting screen and repeats. Because the model carries no memory between calls, the entire accumulated history of the run is resubmitted at every step. A forty-step workflow therefore does not submit forty screenshots — it submits the first screenshot forty times, the second thirty-nine times, and so on.

That compounding is the central mechanic. Processing volume grows with the square of the step count, not in proportion to it. A workflow that doubles in length quadruples in cost.

3.3 Structural comparison

Property	Deterministic Framework	Agentic Loop
Cost per run	Fixed and known in advance	Varies with step count and screen complexity
Cost growth	Linear with restaurant count	Quadratic with workflow length
Reproducibility	Identical execution every run	May take a different path on each run
Failure mode	Deterministic and diagnosable	Non-deterministic, harder to reproduce
Retry cost	Only the failed step	Entire accumulated run to that point
Deferred execution	Eligible for batch processing	Not eligible; requires live session state
Audit trail	Every action logged as code	Reasoning may vary between runs
Best suited to	Stable, repeated, high-volume workflows	Novel or frequently changing interfaces

The agentic approach is not inferior in general. Where a target interface changes frequently and selectors break constantly, the flexibility justifies the cost. Sculpture Hospitality, Toast, and Restaurant365 are mature enterprise applications with stable interfaces and defined release cycles — precisely the case in which determinism wins on cost, speed, and reliability simultaneously.

4. What the Architecture Costs the Operator

Edgebrook’s pricing is fixed and published: $350 per month at Tier 1, $750 at Tier 2, and $12 per restaurant at Tier 3, capped for now at $2,400 for a franchise of 200 restaurants. This section explains why those figures are possible, and what the same service costs when it is delivered on the prevailing alternative.

4.1 The efficiency difference

Both architectures perform the same four workflows, at the same weekly cadence, against the same systems, and produce the same result. The difference is how much processing each one consumes to get there.

Twenty-six times less processing for the same result The Edgebrook Framework completes the four core workflows using approximately one twenty-sixth of the processing volume required by an agentic implementation. The ratio holds at every tier, from twenty restaurants to two hundred.

The mechanic behind that figure is straightforward. An agentic system carries no memory between steps, so the entire accumulated history of a workflow run is resubmitted every time it takes an action. A forty-step run does not process forty screens; it processes the first screen forty times, the second thirty-nine times, and so on. Deterministic execution performs each step once, because the sequence is already known.

That comparison assumes a competently built agentic system with caching enabled and optimization applied. Unoptimized implementations consume several times more. The comparison is deliberately made against the strongest version of the alternative rather than the weakest.

4.2 What that means for what you pay

Processing is one input among several. Deployment, integration with Sculpture and the point-of-sale system, ongoing support, and hosting are broadly comparable whichever architecture a vendor chooses. Holding those constant and varying only the cost of execution gives the price a comparable vendor would need to charge to deliver the same service. The figure below places the two side by side on the same scale, tier by tier.

Figure 1. Monthly price to the operator for the same service under each architecture — same deployment, integration and support in both cases; the difference is the cost of executing the workflows. See Appendix A.

Tier	Restaurants	Edgebrook price	Agentic vendor	Difference
Tier 1	20	$350	$570	1.6×
Tier 2	50	$750	$1,300	1.7×
Tier 3	200	$2,400	$4,620	1.9×

The disadvantage grows with the size of the operation. At twenty restaurants the difference is roughly sixty percent; at two hundred it approaches double. This is the reverse of the volume economics an operator expects, and it follows directly from the architecture. Edgebrook’s cost of serving each additional restaurant is negligible, so the per-restaurant rate can fall as the estate grows. On an agentic system that cost is substantial and does not fall, so the rate cannot.

4.3 Why the gap widens over time

Three mechanics push agentic processing volume upward over the life of a deployment, and none of them applies to deterministic execution.

Length compounds. Because the full history is resubmitted at every step, processing volume grows with the square of the step count. A workflow that doubles in length quadruples in cost, and production workflows accumulate edge-case handling — and therefore steps — over their lifetime.
Retries are expensive. A failure late in a run discards the entire accumulated volume of that run and the work is processed again. Deterministic execution fails at an identifiable step and resumes from it.
Deferred processing is unavailable. Batch execution, which halves processing cost across the industry, requires that work be deferrable. A loop dependent on live browser state cannot be deferred, so that discount is out of reach.

The figures above therefore describe the two architectures on the day they are deployed. Every subsequent month of workflow refinement moves them further apart, and that movement has to be absorbed somewhere: in the vendor’s margin, in a price revision, or in a limit on how often workflows may run.

What this means at the negotiating table A vendor offering a capped monthly price on an agentic architecture is doing one of three things: pricing the cap defensively high to cover the exposure, limiting execution volume in terms that appear after signature, or absorbing a loss that will eventually require correction. Edgebrook’s $2,400 ceiling is none of those — it is a consequence of an execution cost that is small and does not move.

5. What This Means for Pricing

Edgebrook’s tiers price on a graduated (banded) scale, the way marginal tax brackets work. Tier 1 is a flat monthly rate. Above it, each additional restaurant adds a fixed amount — but each tier’s rate applies only to the restaurants that fall inside that tier’s band, never retroactively to the whole estate. Because a lower tier rate is never charged against the restaurants below it, the monthly total only ever rises: crossing into a higher tier can never lower an operator’s bill. Each tier includes the four core workflows deployed and operating.

Band	Restaurants	Rate above previous band	Cap (max) reached at	Blended rate at cap
Tier 1	1–20	Flat $350	$350 · any count 1–20	$17.50
Tier 2	21–50	+$15 per restaurant	$750 at 47 restaurants	$15.00
Tier 3	51–200	+$12 per restaurant	$2,400 at 188 restaurants	$12.00

The band rates are set so the blended (all-in, per-restaurant) price lands on the published figures — $17.50 at 20 restaurants, $15.00 at 50, and $12.00 at 200 — while the total climbs smoothly in between. An operator pays for the size it actually runs rather than a single flat tier price. The $2,400 ceiling is reached at 200 restaurants. The figure below plots the whole scale, with two worked examples directly beneath it.

Figure 2. The graduated scale. Tier 1 is flat at $350 up to 20 restaurants. Tier 2 starts at $350 and adds $15 for each restaurant over 20, reaching the $750 cap at 47 and holding flat through 50. Tier 3 starts at $750 and adds $12 for each restaurant over 50, reaching the $2,400 ceiling at 188 and holding flat through 200 restaurants. The monthly total only ever rises.

Worked examples — how a Tier 2 and a Tier 3 bill are built

Tier 2 — a 35-restaurant operator

Tier 1 base (first 20 restaurants): $350.00
Restaurants 21–35 — 15 × $15.00 per restaurant: $225.00
Monthly total: $575.00 — a blended $16.43 per restaurant. (Still below the $750 cap, which is reached at 47.)

Tier 3 — a 120-restaurant operator

Tier 2 max (first 50 restaurants): $750.00
Restaurants 51–120 — 70 × $12.00 per restaurant: $840.00
Monthly total: $1,590.00 — a blended $13.25 per restaurant. (Still below the $2,400 cap, which is reached at 188.)

Each restaurant is billed at its own band’s rate and the bands add up — so the total always climbs and the blended rate slides down toward $12 as the estate grows, exactly as the graph shows.

Crossing a tier never lowers your bill Because each band’s rate applies only to the restaurants inside it, growth is always monotonic — at every tier boundary the total steps up, never down:

Tier 1 → Tier 2: 20 → $350 ($17.50 each); the 21st restaurant adds $15 → $365 ($17.38 each) — just above 20, never below.
Tier 2 → Tier 3: 50 → $750 ($15.00 each); the 51st restaurant adds $12 → $762 ($14.94 each) — just above 50, never below.
Ceiling: the $2,400 cap is reached at 188 restaurants and holds through 200 restaurants.

A naive “lower tier rate × every restaurant” scheme would instead bill a 51-restaurant estate 51 × $12 = $612 — less than the 50-restaurant one. The banded scale removes that inversion by construction.

Three properties of this structure follow directly from the architecture described above.

The rate declines as the operation grows. Because each band’s rate is lower than the last, the blended per-restaurant price falls steadily as the estate grows — $17.50 at Tier 1, $15.00 at 50 restaurants, and $12.00 at 200 — roughly thirty-one percent lower at the top than at the bottom.
The ceiling is real. Tier 3 is capped at $2,400 per month, reached at 200 restaurants. That commitment is only safe to make because the marginal execution cost of each additional restaurant is known and small.
The price is insulated from consumption growth. Because execution cost does not scale with workflow complexity or run length, there is no mechanism by which ordinary growth in usage forces a price revision.

A platform built on the agentic pattern cannot offer a capped tier of this kind without either pricing the cap defensively high or accepting an open-ended liability. Where flat pricing is offered on that architecture, it is typically accompanied by execution quotas, fair-use limits, or overage provisions that surface only after deployment.

5.1 What is included and what is additional

Each tier includes the four core workflows, and the deployment cost covers bringing all four into operation. Workflows beyond those four constitute additional scope and are addressed by formal proposal before any work begins. The economics described in this paper apply equally to additional workflows, which is why they can be quoted at a predictable cost rather than as an open-ended engagement.

6. The Reliability Dividend

Cost is the most measurable advantage of deterministic execution but not necessarily the most important one. Sculpture Hospitality functions as the system of record for beverage inventory, recipe costing, and variance analysis. The output of these workflows feeds ordering decisions and period financials.

Deterministic execution produces the same result from the same input every time. That property makes the automation testable in the conventional sense: a change can be validated against a regression suite before it reaches a production tenant, and a failure can be reproduced on demand rather than inferred from logs. A workflow whose execution path may differ between runs cannot offer the same assurance, and in a system feeding financial close, that distinction is material.

This is a discipline carried over from enterprise quality-engineering practice, where reproducibility is the precondition for every other guarantee a platform makes about its own behavior.

Appendix A: Assumptions and Methodology

The comparative figures in Section 4 derive from the following modeled assumptions. They are stated explicitly so that they can be examined, challenged, and replaced with measured production values.

A.1 Workload

Workflow	Runs / month	Input volume	Output volume
Product Mix import	4	5,000 tokens	1,000 tokens
Invoice ingestion	4	4,000 tokens	1,000 tokens
Variance reporting	4	8,000 tokens	1,500 tokens
Period close	4	10,000 tokens	2,000 tokens

A.2 Agentic comparison basis

Forty interaction steps per workflow run.
Approximately 1,750 tokens per screen capture, consistent with published image tokenization rates.
Full conversation history resubmitted at each step, yielding roughly two million cumulative input tokens per run.
An eighty-five percent reduction on repeated input is assumed throughout — the approximate effect of prefix caching within a single warm run. Implementations without caching cost several times the figures shown; they are excluded so the comparison is made against an optimized alternative.
Batch processing discounts are excluded from the agentic scenario, as live browser state precludes deferred execution.

A.3 Limitations

These are modeled estimates, not measurements. Actual volumes vary with restaurant size, transaction count, and invoice format. Scanned invoice documents consume materially more processing than text-based equivalents. Point-of-sale exports for high-volume locations may exceed the figures above.

The relative conclusions are robust to substantial variation in these inputs, because the difference between the architectures is one of order of magnitude rather than degree. The absolute figures should be replaced with instrumented production values before being used for any purpose beyond comparison.

A.4 Sources

Token pricing and caching behavior: published model-provider pricing documentation, August 2026.
Frontier price-index decline: tracked index of median frontier model pricing, July 2026 snapshot.
Enterprise spend growth and margin-erosion figures: published industry survey data, 2025 to 2026.
Agentic consumption multiplier of five to thirty times: industry analyst research, March 2026.

Edgebrook Software, LLC — Deterministic Automation and the Economics of Predictable Pricing. Modeled comparison for evaluation purposes; absolute figures should be replaced with instrumented production values.