A specialty chemicals manufacturer developing custom adhesive formulations was stuck in a costly cycle. Each new customer application required extensive experimentation to find the optimal blend of 12 chemical components. Their R&D team was running 200 experiments per quarter at a cost of $1,400 per experiment — $280,000 quarterly, $1.12 million annually — just on experimentation costs, not counting the scientist time. Most experiments were guided by educated guesses and one-variable-at-a-time testing, which meant they were systematically missing interactions between components and leaving performance on the table.
We deployed a Bayesian optimization system that modeled the relationship between formulation inputs and performance outputs using a Gaussian process surrogate model. Instead of testing random or sequential combinations, the system intelligently selected the next experiment to maximize information gain. The result: optimal formulations were found in an average of 45 experiments instead of 200 — a 78 percent reduction in experimentation cost. Even better, the optimized formulations performed 12 percent better on key metrics because the system explored the design space more thoroughly than human-guided experimentation ever could.
Bayesian optimization is a powerful but underappreciated technique that AI agencies can deliver to clients facing expensive, time-consuming optimization problems. It applies anywhere the goal is to find the best configuration of parameters when each evaluation is costly. Here is the delivery playbook.
What Bayesian Optimization Is and Why It Matters
Bayesian optimization is a strategy for finding the optimal value of an expensive-to-evaluate objective function with as few evaluations as possible.
The core idea:
- Build a statistical model (surrogate model) of the objective function based on data observed so far
- Use the surrogate model to decide which point to evaluate next — balancing exploration (learning about unexplored regions) and exploitation (sampling near known good solutions)
- Evaluate the objective function at the selected point
- Update the surrogate model with the new observation
- Repeat until the optimization budget is exhausted or a satisfactory solution is found
Why this matters for business:
Many real-world optimization problems involve evaluations that are expensive in time, money, or both. Each evaluation might be:
- A physical experiment (lab test, manufacturing trial, field test)
- A time-consuming simulation (computational fluid dynamics, financial model)
- A business experiment (A/B test, pricing experiment, marketing campaign)
- A resource-intensive process (training a machine learning model, designing a circuit)
In these situations, you cannot afford to evaluate thousands of combinations. Bayesian optimization finds good solutions with 5-10x fewer evaluations than grid search, random search, or manual experimentation.
High-Value Use Cases for Bayesian Optimization
Manufacturing Process Optimization
Optimizing production parameters (temperature, pressure, speed, timing, material ratios) to maximize quality, yield, or efficiency while minimizing cost and waste.
Example: A food manufacturer adjusting 8 baking parameters (temperature, humidity, time, ingredient ratios) to optimize texture, taste, and shelf life. Each baking trial takes 4 hours and costs $800. Bayesian optimization found a superior recipe in 30 trials that manual experimentation had not found in 500 trials over two years.
Product Formulation
Finding optimal ingredient combinations for chemical products, food products, pharmaceuticals, or consumer goods.
Example: A cosmetics company optimizing a moisturizer formulation across 15 ingredients for skin feel, stability, and cost. Traditional design-of-experiments would require 200+ formulations. Bayesian optimization found a Pareto-optimal formulation in 40 trials.
Marketing Campaign Optimization
Optimizing campaign parameters (bid amounts, audience targeting, creative elements, timing, budget allocation) where each configuration takes days or weeks to evaluate.
Example: An e-commerce company optimizing 6 parameters of their Google Ads campaigns. Each configuration needed 2 weeks of data to evaluate reliably. Bayesian optimization found the optimal configuration in 12 iterations (24 weeks) versus the 40+ iterations (80+ weeks) that manual optimization would have required.
Simulation Optimization
Optimizing parameters of expensive computer simulations (engineering design, supply chain configuration, financial models).
Example: An automotive supplier optimizing the geometry of a heat exchanger using computational fluid dynamics simulations. Each simulation takes 8 hours. Bayesian optimization found a design that improved heat transfer by 18 percent in 50 simulations versus 300+ with traditional methods.
Machine Learning Hyperparameter Tuning
Optimizing neural network architecture and training parameters where each training run takes hours or days.
Example: A client needed to fine-tune a large language model for domain-specific tasks. Each training run cost $2,000 in compute. Bayesian optimization found optimal hyperparameters in 15 runs versus 50+ with random search.
Technical Architecture
Surrogate Model
The surrogate model is the statistical model that approximates the objective function.
Gaussian Process (GP) regression is the standard choice for Bayesian optimization:
Strengths:
- Provides both a prediction and an uncertainty estimate at every point
- Uncertainty quantification is essential for the exploration-exploitation tradeoff
- Works well with small datasets (which is the whole point)
- Non-parametric — adapts to the data without assuming a specific functional form
Limitations:
- Scales poorly beyond ~1,000 observations and ~20 dimensions
- Requires choosing a kernel function (though Matern 5/2 is a good default)
- Assumes smoothness that may not hold for all objective functions
For higher dimensions or larger datasets:
- Random forests with uncertainty estimation
- Bayesian neural networks
- Deep kernel learning (combining neural networks with Gaussian processes)
- Ensemble methods with uncertainty quantification
Acquisition Function
The acquisition function determines which point to evaluate next. It balances exploration (sampling where uncertainty is high) and exploitation (sampling where the predicted value is good).
Common acquisition functions:
- Expected Improvement (EI): The expected amount by which the next evaluation will improve over the current best. The most popular choice and a good default.
- Upper Confidence Bound (UCB): The predicted value plus a multiple of the uncertainty. Simple and effective.
- Knowledge Gradient: The expected improvement in the overall decision quality from one more evaluation. More computationally expensive but theoretically sound.
- Thompson Sampling: Sample from the posterior and optimize the sample. Simple to implement and works well in practice.
For multi-objective optimization:
- Expected Hypervolume Improvement (EHVI)
- ParEGO (scalarizes multiple objectives with random weights)
- Multi-objective Thompson Sampling
Constraint Handling
Real-world optimization almost always has constraints:
- Material properties must meet minimum standards
- Production parameters must stay within equipment limits
- Budget constraints on total experimentation cost
- Safety constraints that cannot be violated
Approaches:
- Constrained acquisition functions: Multiply the acquisition function by the probability of constraint satisfaction
- Penalty methods: Add a penalty to the objective for constraint violations
- Feasibility-aware surrogate models: Model constraints as separate GPs and only suggest feasible points
Batch Optimization
In many practical settings, you can run multiple experiments in parallel (e.g., multiple lab stations, multiple simulation servers).
Batch Bayesian optimization selects multiple points to evaluate simultaneously:
- Kriging believer: Select one point, update the model with the predicted value, select the next point, repeat
- Local penalization: Penalize the acquisition function near already-selected points
- Batch Thompson Sampling: Draw multiple posterior samples and optimize each one
Delivery Framework
Phase 1: Problem Definition and Design (Weeks 1-3)
Activities:
- Understand the optimization problem: What are the inputs (design variables)? What are the outputs (objectives)? What are the constraints?
- Characterize the evaluation cost: How long and how expensive is each evaluation?
- Map the current optimization approach: How does the client currently explore the design space?
- Define the search space: What are the ranges and types (continuous, discrete, categorical) of each input variable?
- Assess existing data: Does the client have historical data from previous experiments?
- Design the optimization strategy: Surrogate model, acquisition function, batch size, total budget
Deliverable: Optimization specification document with problem formulation, strategy, and success metrics.
Phase 2: System Development (Weeks 4-6)
Activities:
- Implement the Bayesian optimization framework
- Build the surrogate model on any available historical data
- Implement the acquisition function with constraint handling
- Build the user interface for suggesting experiments and recording results
- Implement visualization of the optimization progress (convergence plots, surrogate model visualization, design space exploration)
- Test the system with synthetic objective functions that mimic the real problem
Phase 3: Active Optimization (Weeks 7-12+)
Activities:
- System suggests the next experiment(s) to run
- Client runs the physical experiments or simulations
- Results are recorded in the system
- Surrogate model is updated
- System suggests the next experiment(s)
- Monitor convergence and optimization progress
- Adjust strategy if the optimization is not progressing as expected
This phase duration depends on the evaluation time. If each experiment takes 1 day, you can run 30 iterations in 6 weeks. If each experiment takes 2 weeks, the same 30 iterations take over a year. Plan accordingly.
Phase 4: Analysis and Handoff (Weeks 13-14, or whenever optimization concludes)
Activities:
- Analyze the optimization results: What is the best solution found? How does it compare to the previous best?
- Characterize the objective function: What inputs matter most? What are the interaction effects?
- Deliver the surrogate model as a tool for ongoing design space exploration
- Document methodology, results, and recommendations
- Train the client team to use the system for future optimization projects
Common Delivery Challenges
Client Experiment Execution
Your optimization system can only suggest experiments — the client has to execute them. Delays in experiment execution slow the entire optimization.
Manage this:
- Establish a clear cadence for experiment execution (e.g., 3 experiments per week)
- Provide batch suggestions so the client always has queued experiments
- Build a buffer in the timeline for client-side delays
- Make it easy to record results (mobile-friendly interface, minimal data entry)
Noisy Objective Functions
Real-world experiments have measurement noise. Two identical experiments might produce different results. This noise makes optimization harder.
Handle noise by:
- Modeling noise explicitly in the Gaussian process (heteroscedastic noise model)
- Running replicate experiments at the same conditions to estimate noise level
- Using acquisition functions that account for noise (e.g., noisy EI)
- Setting realistic expectations: the optimization will converge more slowly with higher noise
High-Dimensional Problems
Standard Bayesian optimization works well up to about 20 dimensions. Beyond that, the surrogate model struggles to learn the objective function.
Strategies for high dimensions:
- Feature selection: Work with domain experts to identify the most important variables and fix the rest
- Dimensionality reduction: Use techniques like principal component analysis or autoencoders to reduce the effective dimension
- Structured surrogate models: Use additive models or other structured assumptions to handle more dimensions
- Trust region methods: Optimize in local regions rather than the full space
Stakeholder Communication
Bayesian optimization is conceptually unfamiliar to most business stakeholders. They need to understand and trust the process.
Communication strategies:
- Show the exploration-exploitation tradeoff visually: "The system is testing this point because we know very little about this region"
- Track and display the convergence: "After 20 experiments, we have improved from X to Y"
- Compare to what random or manual experimentation would require
- Let domain experts review and veto suggestions that violate domain knowledge the model does not capture
Pricing Bayesian Optimization Projects
Project-based pricing:
- Single-objective optimization system: $50,000-100,000
- Multi-objective optimization with constraints: $80,000-150,000
- Enterprise optimization platform (multiple problems, user management, reporting): $150,000-250,000
Per-optimization-campaign pricing:
For clients who want to use the system for multiple optimization problems:
- Setup and configuration for each new problem: $10,000-25,000
- Ongoing support during active optimization: $3,000-8,000 per month
Value justification: If each experiment costs $1,000 and the current approach requires 200 experiments, that is $200,000 per optimization campaign. Bayesian optimization requiring 50 experiments saves $150,000 per campaign. A $100,000 system that supports multiple campaigns pays for itself on the first use.
Your Next Step
Identify a client who runs expensive experiments — manufacturing trials, lab tests, simulation studies, or long-running A/B tests. Show them the math: how many experiments they currently run, what each one costs, and how Bayesian optimization can find better solutions with a fraction of the experiments. Offer a pilot on one specific optimization problem where you can demonstrate clear time and cost savings. The combination of faster results and lower costs makes this an easy sell for any organization that regularly faces expensive optimization challenges.