Table of Contents
To calculate the return on a virtual lab, measure how many experiments and calendar days your team needs today to bring a formulation to specification, then price one experiment and one day of delay. Compare the savings with the full three-year cost, and express the result as ROI, NPV, payback and the break-even reduction a pilot must beat.
Key takeaways
- A virtual lab's return comes from fewer physical experiments, faster time to specification and reuse of past work. Only the first two are easy to price.
- Only part of an avoided experiment is cash. Materials and external test fees are saved; technician and equipment time is freed capacity, worth money only if it is reused.
- Published gains vary widely. One pilot-scale study found no reduction in experiments (Rummukainen et al., 2024); another estimated 3 to 30 times fewer (Narayanan et al., 2025). Use your own pilot figures.
- Calculate the break-even reduction in experiments first. It tells you what the pilot has to prove.
- Count all costs over three years, including data preparation and your own staff's time.
Where does a virtual lab's return come from?
A virtual lab predicts test results from a recipe and proposes the next experiments worth running. Its value comes through several channels, each calculated differently.
| Benefit | How to calculate it | Is it cash? |
|---|---|---|
| Fewer experiments | Avoided experiments × cash cost per experiment (raw materials, consumables, external test fees) | Yes |
| Freed lab capacity | Avoided experiments × technician and equipment time per experiment | Only if the time is reused: more projects, less overtime, fewer outsourced tests |
| Faster time to specification | Days saved × value of one day for that project | Only where a launch, tender or customer deadline depends on it |
| Fewer late failures | Avoided scale-up or field failures × their cost | Rarely provable in the first years |
| Knowledge kept and reused | Projects that start from stored recipes and models | Usually left unpriced |
Keep the last two rows out of the main calculation unless you have history to support them, and list them as upside.
How do you set the baseline?
Everything rests on the baseline: how your lab performs today without a virtual lab. Fix it before the project starts, from comparable past projects.
- Experiments per project: all physical tests until a recipe met the full specification, including failed and repeated tests.
- Calendar days per project: from receiving the specification to a validated recipe.
- Projects per year in the product family you plan to cover.
- Cost per experiment, split into cash and capacity, as described below.
Compare projects of the same family and similar difficulty. Adapting a drilling fluid to a new supplier is not comparable to developing a new mud system, and raw material changes alone can multiply trials, as our guide to raw material variability explains.
How do you price one experiment?
Add up the fully loaded cost of testing one candidate recipe:
- raw materials and reagents, including what is lost during preparation;
- consumables and external test fees;
- technician and engineer hours at a loaded hourly rate;
- equipment time: depreciation, calibration and maintenance spread over the tests run.
Then split the total. Materials, consumables and external fees disappear when a test is not run. Staff and equipment costs do not, unless the freed hours go to other work. Present the two parts separately.
What costs belong in the calculation?
Our guide to industrial AI project cost covers the cost lines in detail. For a virtual lab, check that the calculation includes:
- The software licence or subscription, and deployment on your infrastructure.
- Data preparation: collecting, cleaning and structuring past recipes and results. Data scientists in Anaconda's 2020 survey reported spending about 45% of their time loading and cleaning data (InfoWorld). The FAIR principles (findable, accessible, interoperable, reusable) are a useful target.
- Configuration of targets, constraints and test protocols for each product family.
- Training and adoption for formulators and technicians.
- Confirmation experiments. The lab still tests every recommended recipe, and those tests count against the savings.
- Running costs every year: support, model updates and a named owner on your side. Google researchers found it "common to incur massive ongoing maintenance costs in real-world ML systems" (Sculley et al., 2015).
Which formulas should you use?
Use three measures over the same three-year period:
- ROI = (total benefits − total costs) ÷ total costs.
- Payback period: the time until cumulative net benefits recover the initial investment. It is easy to explain, but it ignores the time value of money and the cash flows after payback (OpenStax, Principles of Finance).
- Net present value (NPV): the present value of the benefits minus the present value of the costs, at your company's discount rate. Projects with a positive NPV should be accepted (OpenStax, Principles of Finance).
Add a fourth measure specific to a virtual lab: the break-even reduction, the percentage of experiments the virtual lab must remove for the result to reach zero. It turns the business case into a claim a pilot can test.
Illustrative worked example
Every figure in this example is invented to show the method. None comes from a real lab, a KABAS client or a KABAS price list. Replace each one with your own data.
| Assumption | Illustrative value |
|---|---|
| Formulation projects per year in the product family | 20 |
| Experiments per project to reach specification (baseline) | 40 |
| Experiments per year | 800 |
| Cost per experiment, cash part (materials, consumables, external fees) | 20,000 DZD |
| Cost per experiment, capacity part (staff and equipment time) | 40,000 DZD |
| Initial cost (licence, deployment, data preparation, training, internal time) | 12 million DZD |
| Running cost per year (support, updates, owner's time) | 4 million DZD |
| Benefit in year 1 (ramp-up) | 50% of full benefit |
| Discount rate | 10% |
Base case: 25% fewer experiments, freed time reused. The lab avoids 200 experiments a year, worth 200 × 60,000 = 12 million DZD at full rate.
- Year 0: −12.0 million DZD
- Year 1: 6.0 − 4.0 = +2.0 million DZD
- Year 2: 12.0 − 4.0 = +8.0 million DZD
- Year 3: +8.0 million DZD
Total benefits are 30 million DZD against 24 million DZD of costs. That gives a three-year ROI of 25%, an NPV of about +2.4 million DZD at 10%, and payback about 2 years and 3 months after go-live.
Same case, cash only. If the freed technician and equipment time is not reused, the cash saving is 200 × 20,000 = 4 million DZD a year. That only covers the running cost, so the initial 12 million DZD is never recovered and the three-year result is −14 million DZD.
Same case, with the value of time. Suppose 4 of the 20 projects each year are tender responses or launches for which finance values each day gained at 50,000 DZD, and the virtual lab saves 15 days on each. That adds 3 million DZD a year at full rate. The three-year ROI rises to about 56%, the NPV to about +8.5 million DZD, and payback falls to under 2 years.
Sensitivity to the reduction in experiments, with freed time reused and no value of time:
| Fewer experiments | Full annual saving | 3-year net result | 3-year ROI | NPV at 10% | Payback after go-live |
|---|---|---|---|---|---|
| 10% | 4.8 M DZD | −12.0 M DZD | −50% | −12.2 M DZD | Not within 3 years |
| 20% | 9.6 M DZD | 0 | 0% | −2.4 M DZD | 3 years |
| 25% | 12.0 M DZD | +6.0 M DZD | 25% | +2.4 M DZD | About 2.3 years |
| 40% | 19.2 M DZD | +24.0 M DZD | 100% | +17.1 M DZD | About 1.4 years |
In this example, break-even is 20% fewer experiments without discounting and about 22.5% at a 10% discount rate. That is the figure the pilot has to beat, and it moves with every assumption above.
What does published evidence say about fewer experiments?
Published results vary widely, and few match routine industrial formulation work:
- Bayesian optimisation, a common method behind a virtual lab, outperformed expert chemists and engineers in both average number of experiments and consistency on a reaction optimisation task (Shields et al., 2021).
- In a pilot-scale wood delignification study, Bayesian optimisation reached conditions comparable to a 15-run Box-Behnken design but "did not enable a decrease in the number of experiments" (Rummukainen et al., 2024).
- In cell culture media, the authors report 3 to 30 times fewer experiments than they estimated a standard design of experiments would need. No DoE campaign was run in parallel (Narayanan et al., 2025).
- A benchmark across five materials data sets measured acceleration against a random sampling baseline (Liang et al., 2021). Experienced formulators do better than random, so such factors overstate the gain against your own team.
- Robotic studies found nine suitable recipes for a formulated liquid product within 15 working days (Cao et al.) and ran 688 experiments in eight days (Burger et al., 2020). Most industrial labs have no robots, so their saving comes from fewer tests, not faster ones.
The method also suits a particular profile: tests that take minutes or hours to evaluate and fewer than about 20 continuous variables (Frazier, 2018). If your tests are cheap and fast, the return will be small whatever the model. Our comparison of design of experiments and Bayesian optimisation explains when each one fits.
Which mistakes inflate the result?
Counting experiments your team would never have run, taking a published gain as your own assumption, comparing the pilot with a harder past project, leaving out confirmation tests and failed recipes, counting freed staff time as cash, and assuming full benefit from the first month.
How do you test the assumptions in a pilot?
- Choose one product family and fix the baseline from past projects.
- Price one experiment with your finance team, split into cash and capacity.
- Agree with sales or finance which projects carry a value of time, and how much one day is worth.
- List all costs for three years, including your own staff's time.
- Calculate the break-even reduction and set the pilot's pass mark above it.
- Run the pilot, counting every physical test and logging predicted against measured results, as described in our AI formulation optimisation guide.
- Recalculate ROI, NPV and payback with the measured figures before deciding on a rollout.
To present the result to management, use the structure in how to build an AI business case.
What we have seen at KABAS AI
Our Virtual Lab is deployed in two formulation environments we can describe publicly. Neither case study publishes a financial ROI; both report the operational measures the calculation starts from.
At ENAD Shymeca, the published result for detergent formulation is a 40% reduction in physical testing cycles, with faster time to market and lower material consumption. The implementation went through discovery, configuration, migration of historical formulation data, training and continuing support: each of these is a cost line to budget.
At ADFC (Sonatrach Group), the case study reports a significant reduction in physical testing cycles, faster responses to technical specifications and more competitive proposals per drilling section. In the calculation above, faster responses belong to the value-of-time line.
The measurable outputs are experiments and days. Turning them into money depends on your own costs, so build the calculation with your finance team. For help setting the baseline and the pilot pass mark for your lab, contact us.
Sources
- OpenStax, Principles of Finance, 16.1 Payback Period Method
- OpenStax, Principles of Finance, 16.2 Net Present Value (NPV) Method
- InfoWorld, For data scientists, drudgery is still job one (Anaconda 2020 State of Data Science survey)
- GO FAIR Foundation, FAIR Principles
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015
- Shields et al., Bayesian reaction optimization as a tool for chemical synthesis, Nature 590, 2021
- Rummukainen et al., Traditional or adaptive design of experiments? A pilot-scale comparison on wood delignification, Heliyon 10, 2024
- Narayanan et al., Accelerating cell culture media development using Bayesian optimization-based iterative experimental design, Nature Communications, 2025
- Liang et al., Benchmarking the performance of Bayesian optimization across multiple experimental materials science domains, arXiv 2106.01309, 2021
- Cao et al., Optimization of Formulations Using Robotic Experiments Driven by Machine Learning DoE, Cell Reports Physical Science 2
- Burger et al., A mobile robotic chemist, Nature 583, 2020
- Frazier, A Tutorial on Bayesian Optimization, arXiv 1807.02811, 2018
Frequently asked questions
What is a good ROI for a virtual lab?
There is no universal figure. The return depends on how much one experiment costs you, how many formulation projects you run each year, how much a day of delay is worth, and what the virtual lab costs over three years. Calculate the break-even reduction in experiments for your own lab and check in a pilot whether you beat it.
Does every experiment avoided save money?
Only part of it. Raw materials, consumables and external test fees are saved in cash. Technician and equipment time is freed capacity, which is worth money only if it is reused for more projects, less overtime or fewer outsourced tests. Show the two parts separately in the business case.
Which costs should be included in a virtual lab business case?
Include the licence or subscription, deployment, preparation and migration of past formulation data, configuration of targets and test protocols, training, the confirmation experiments the lab still runs, and yearly running costs such as support, updates and the time of a named owner. Count your own staff's time as well as the vendor's.
How long does a virtual lab take to pay back?
It depends entirely on your figures. In the illustrative example in this article, payback after go-live ranges from about 1.4 years to more than 3 years depending on the reduction in experiments. Expect a ramp-up period in the first year while the models and the team learn.
Can we use published speed-ups in our business case?
Not as your assumption. Published results range from no reduction in experiments to estimates of 3 to 30 times fewer, and many compare against random sampling or an estimated design of experiments rather than an experienced formulation team. Use them to understand the range, then measure your own result in a pilot.
How do we measure the reduction in experiments reliably?
Fix the baseline before the project from comparable past projects in the same product family. During the pilot, count every physical test until a recipe meets the full specification, including confirmation tests, replicates and failures, and log each predicted value next to the measured one.



