The Hard Truth About AI formulation in R&D

The recent Q&A with Mohamed Badaoui Najjar, PhD by Institute of Food Technologists (IFT) resonated with me. Indeed, it reminded me our journey from last 5 years. Mohammed's point about the 0.011% error threshold in food R&D highlights one of the reason why traditional AI formulation approaches hit fundamental limits.

The Scaling Laws That Don't Scale

In most AI applications, more data = better performance. In R&D formulation:

• 80% of data is incomplete, incorrect, or unusable
• A 30-year-old company might have 5,000 SKUs but only 10 clean data points per formula
• Stability data takes 18 months to generate—by then, the market's moved on
• Digital transformation initiatives take 3+ years to show ROI

Building AI that SCALES for this environment requires fundamentally different thinking. You can't pattern-match your way to innovation when patterns barely exist.

The Stage-Gate Complexity Formulation operates in multi-dimensional decision spaces:

• Gate 1 decisions need broad exploration (RAG works)
• Gate 2 requires precise constraint validation (fine-tuned models)
• Gate 3 demands SME judgment (no AI suffices)

Beyond the LLM Limitations

Yes, context degradation is real—Stanford shows accuracy dropping below 40% at 10,000 tokens. But that's just one piece. The deeper challenges:

1. Multi-Constraint Orchestration Every formulation decision touches regulatory, cost, processing, sensory, and stability constraints simultaneously. No monolithic AI handles this complexity reliably.

2. Confidence Without Classification A suggestion could be based on thousands of documents or pure hallucination—formulators can't tell the difference. Indeed, even LLMs expert struggle on this.

3. Expertise Integration How do you merge 30 years of formulator intuition with AI recommendations? Most systems override human judgment when they should augment it.

The breakthrough comes from orchestrating these approaches—knowing when to trust data, when to defer to SMEs, and when to acknowledge unknowns.

Our Journey and What's Still Ahead

After 4 years and multiple patents, we've solved pieces of this puzzle:

• Dynamic orchestration between data-driven and expertise-driven decisions
• Context-aware routing that matches problem types to solution approaches
• Confidence calibration that preserves formulator trust

But we're honest: this is just the beginning. The unsolved challenges dwarf what we've built:

• Real-time learning from production failures
• Cross-category knowledge transfer
• Handling novel ingredients with zero historical data

The Path Forward

The industry needs to embrace new technologies while understanding its limitation. We're not building AI to replace R&D expertise. We're building systems that understand when to use which tool, when to ask for help, and when to get out of the way.

The companies that succeed won't be those waiting for perfect AI. They'll be those building imperfect systems that improve through use.

Turing Labs - world's first AI platform purpose built for R&D to formulate winning formulations.

Turing Labs - www.turingsaas.com