A large pharmaceutical company runs a high-throughput screening lab where technicians manually inspect assay plates for failures. An assay is a lab test that measures a biological response; a plate holds dozens or hundreds of wells, each a tiny experiment. Roughly 20 to 30 percent of runs fail (a widely cited industry range, exact rates vary by assay type), and each failure means reagents wasted, instrument time lost, and scientists rerunning work. Someone proposes an AI system to flag failing plates early. Good idea? You cannot answer that with enthusiasm. You answer it with a business case.
This lesson teaches you to build that case: pick a bottleneck, quantify the baseline, estimate the lift, and check whether your data can actually deliver it.
Start with a bottleneck, not a technology
The most common mistake in biotech and medtech AI is starting from "we should use AI" and hunting for a place to apply it. Reverse this. Start from a measurable pain point.
Good bottlenecks share three traits:
They are expensive or slow today. High assay failure rates, slow biomarker image analysis, long document review for regulatory submissions.
They generate data as a byproduct. If the process already produces images, sensor logs, or structured records, you have training material.
The output feeds a decision. AI earns money by changing what someone does next, not by producing a dashboard nobody acts on.
Example bottlenecks worth a business case:
Pathologists spending hours counting cells on digitized tissue slides.
QC (quality control) teams manually reviewing manufacturing batch records.
R&D scientists prioritizing which compounds to test next from millions of candidates.
Quantify the baseline
You cannot claim savings without knowing today's cost. Build the baseline from real operational numbers, not vendor slides.
For the assay failure example, break it down:
Number of plates run per year.
Current failure rate.
Cost per failed plate (reagents + instrument time + labor to rerun).
Worked calculation
Assume a lab runs 50,000 plates per year at a 25 percent failure rate. That is 12,500 failed plates. Suppose each failure costs roughly 200 dollars in reagents, instrument time, and labor (illustrative figure, you must measure your own).
Baseline annual failure cost
= 50,000 plates x 25% failure x $200
= $625,000 per year
That 625,000 dollars is your target pool. AI cannot recover all of it. Some failures are caused by bad samples or reagent lots that no early detection can fix. So next you estimate the lift.
Estimate the lift honestly
The lift is the fraction of the baseline cost the AI actually captures. This is where business cases go wrong: people assume near-perfect performance.
Ask two questions:
1. What share of the problem is addressable by AI? If the AI predicts failure early enough to abort a run and save reagents, but only 60 percent of failures are detectable from early-cycle data, your ceiling is 60 percent, not 100 percent.
2. What accuracy will the model realistically hit? A model that catches 80 percent of the detectable failures, at that 60 percent addressability, captures 48 percent of the pool.
Captured value
= $625,000 x 60% addressable x 80% model recall
= $300,000 per year (gross, before AI costs)
Now subtract the cost of the AI itself: vendor licensing or build cost, integration with lab instruments (via LIMS, a Laboratory Information Management System that tracks samples and results), cloud compute, and ongoing monitoring. If those run 120,000 dollars per year, your net is 180,000 dollars. Still attractive. But notice how far that is from the naive "save 625,000 dollars" pitch.
Always present three scenarios: conservative, expected, optimistic. Decision-makers trust a range far more than a single confident number.
Data readiness: the make-or-break variable
Here is the truth most vendors skip: the lift you estimated is only reachable if your data can support it. A brilliant model on unusable data delivers zero.
Check data readiness across four dimensions:
Volume. Do you have enough labeled examples? For a failure-detection model you need many past runs tagged as pass or fail. A few hundred is usually too few; thousands is a more realistic floor for tabular sensor data, more for images.
Labels. Are outcomes recorded reliably? If technicians did not log *why* a plate failed, you have data but no ground truth.
Consistency. Did the instruments, protocols, and file formats stay stable? A model trained on one plate reader may not transfer to another.
Access. Can the data leave the silo it lives in, given patient privacy and IP (intellectual property) constraints?
For clinical or patient data, access is governed by real regulation. In the US, that is HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation. (the Health Insurance Portability and Accountability Act), which restricts use of protected health information. In Europe, it is the GDPR (General Data Protection Regulation) plus the EU AI Act, which as of 2026 phases in obligations for high-risk AI systems, and many medical AI tools fall into that category. Build compliance time into your case; it is not free.
The FDA maintains a public list of AI-enabled medical devices it has authorized, which is a useful reality check on what has actually cleared regulatory review: FDA AI-enabled medical device list.
Framing the case for decision-makers
Package your analysis so a non-technical executive can act on it. A strong one-page case answers:
What bottleneck? One sentence. "25 percent assay failure rate costs us an estimated 625,000 dollars a year."
What does AI do? "Flags likely failures early so runs can be aborted and reagents saved."
Expected net value? The three-scenario range, net of AI cost.
What is required? Data readiness gaps, integration work, regulatory review, and a timeline.
How will we know it worked? The metric you will track post-deployment (failures caught, reagents saved) and when you will review it.
That last point matters. Too many AI projects launch with no agreed success metric, so nobody can say whether they worked. Define the number before you sign the contract.
Vérification des acquis
1. Why does the lesson insist you should start an AI project from a bottleneck rather than from a decision to 'use AI'?
2. According to the lesson, why does an AI output that produces 'a dashboard nobody acts on' fail to justify a business case?
3. A team wants to build a business case claiming an AI system will save money on assay reruns. What must they establish first, and why?
CHOIX MULTIPLES
4. Select ALL correct answers about the traits that make a bottleneck a good candidate for an AI business case.
Sélectionnez toutes les réponses correctes.
CHOIX MULTIPLES
5. Select ALL correct answers describing sound practices when quantifying the baseline for an AI business case.
Sélectionnez toutes les réponses correctes.
Common ways the case falls apart
Even a well-built case can collapse in practice. Watch for these:
The lift assumes the workflow changes, but nobody changes it. If the AI flags a failing plate but the technician's incentive is to finish runs regardless, the savings never materialize. Adoption is part of the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →, not an afterthought.
Model drift erodes the lift. Reagent lots change, instruments get serviced, protocols evolve. A model trained in 2025 may quietly degrade through 2026. Budget for monitoring and retraining, typically an ongoing cost, not one-time.
The baseline was guessed, not measured. If you pitched 625,000 dollars but the real figure was 300,000 dollars, your net value halves and the project may not clear the bar.
Regulatory scope creep. A tool marketed as "decision supportdecision supportTechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.Voir la définition complète →" can quietly become a diagnostic, which shifts it into a regulated device category and adds months. Clarify the intended use early.
A quick reality check on where AI fits
Not every bottleneck deserves an AI solution. Prefer AI when the task is pattern recognition over large data (images, signals, high-dimensional screens), the volume is too high for humans to keep up, and errors are tolerable or reviewable. Prefer simpler automation or process fixes when the problem is really about missing standard operating procedures, broken handoffs, or a single hard business rule. AI applied to a process problem just adds cost.
Key takeaways
Start from a measured bottleneck, not from the technology. The best AI cases begin with a number someone already worries about.
Separate the pool, the addressable share, and the model performance. Multiply them to get realistic captured value, then subtract AI costs to get net value.
Data readiness is the gating factor. Volume, reliable labels, consistency, and lawful access under HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation., GDPR, and the EU AI Act determine whether any lift is reachable.
Present a scenario range and a success metric agreed before launch, and budget for monitoring, retraining, and adoption, which are ongoing costs.
All figures here are illustrative or cited estimates as of 2026; measure your own operation before committing capital.