Our AI Approach
More AI doesn't make a better product. The right AI, in the right place, does.
AI models will change every six months. Your data foundation shouldn't have to.
Why the data matters most
Same question. Two very different answers.
Ask an AI model one simple question about your study history. What comes back depends on the data underneath it.
The question
“Show every prior study with a loss of red cell mass.”
Against free-text history
- Misses studies that say “anemia” instead
- Invents a study that doesn’t exist
- Cites nothing you can check
Against controlled, traceable data
- Reads “anemia”, “RBC decreased” and “erythrocytes, decreased” as one term
- Returns every matching study
- Links each result to its source record
Same model. The difference was the data.
Levels of AI use
Ask instead: how much does the AI decide?
Labels like machine learning, generative AI and agentic AI describe how something is built, not who makes the decision. We describe every AI feature by how much it decides.
- L0
None
- AI
- Nothing
- Human
- Does it all
- L1Most value
Assist
- AI
- Finds, searches and summarizes on request
- Human
- Does the work
- L2Most value
Draft
- AI
- Writes a first draft
- Human
- Reviews every item
- L3Most value
Recommend
- AI
- Analyzes and flags, with cited evidence
- Human
- Decides
- L4
Delegated
- AI
- Runs a defined task end to end
- Human
- Approves at checkpoints
- L5
Autonomous
- AI
- Acts with no human in the loop
- Human
- Not for GLP decisions
Most value today sits at L1–L3.
Finding, drafting and flagging with cited evidence save real time, and the scientist still decides.
Choose the level per task, not per organization.
A report draft and a raw study measurement carry different risks. Each task gets its own level.
L0 is a choice, not a gap.
Where the answer must be exact, no AI is the right answer. And L5 is not for GLP decisions.
Promise 1 of 3
Grounded
The AI only uses what it is allowed to know: approved facts, fixed settings and reviewed prompts.
Settings
Fixed per task, recorded with every output
- Which model and version
- Temperature
- Which tools it may use
Prompts
Treated like SOPs
- Written by experts
- Reviewed and versioned
- Under change control
Data
The facts the AI may use
- One controlled term per finding
- Full context: species, dose, day, instrument
- Provenance for every answer
Reference sources
Ranked, when the AI uses them
- GLP studies and FDA reviews first
- Then peer-reviewed literature, rated for reliability (Klimisch)
- Each source’s tier shown on every output
Promise 2 of 3
Scientists stay in control
AI earns trust when scientists can see it and change it.
“AI should amplify the study director’s judgment, not replace it. No black boxes.”
Visibility
For any AI output, see the data it used, the prompt and settings, and citations back to source records.
Changeability
Adjust prompts and parameters within guardrails. Every change lands in the audit trail.
Curated skills
Packaged AI workflows for common tasks. Experts write them once; each is versioned and validated like any method.
Promise 3 of 3
Risks handled up front
Every AI step is something to validate, monitor and explain to QA and regulators. So we name the risks and design for them.
- Hallucination
- Grounding plus mandatory citations
- Over-reliance
- A level chosen per task; humans decide
- Unvalidated change
- Settings and prompts under change control
Where each level belongs
- L0where the answer must be exact
- L2–L3where a draft or a flag saves time
- L4only where the data can carry it
Our products
Where our products sit today
The level is set per task, not per product. L0 is a choice, not a gap.
Not used today.
- TurboLIMSProtocol report text
- TurboToxicologyReport
- TurboPKTKReport
- TurboBioAReport
- TurboToxicologyPriority scores
Not used today.
Not used. Not for GLP decisions.
L0 tasks such as study data and PK analysis use no AI by design: the answer must be exact. No product uses L5.
What's next · Vision, not product
Where this can go
These are ideas we are working toward with the industry. They are not features you can buy today.
Virtual control groups
- Control matching: pick historical controls that match on strain, sex, age, vendor, diet, facility and instrument, and show why each qualified
- Drift monitoring: flag shifts from an analyzer change, a new colony or a season before the data is used
- Fewer concurrent controls: an auditable, statistically justified case per study that serves the 3Rs
New approach methodologies
- Connect NAMs to animal history: line up in vitro and organ-chip readouts against structured in vivo results
- Weight of evidence: gather NAM, in silico and limited in vivo evidence into one traceable package for the study director
Open study-data APIs
“A FHIR for nonclinical data”
Today SEND is a file, produced at the end of the study. Standard APIs would keep the data live while the study runs, in CDISC controlled terminology.
- Study
- Subject
- Dose
- Observation
- Specimen
- Finding
- Live visibility for sponsors
- Any AI can connect
- Easy sharing between sponsors and CROs
- No lock-in
An invitation to the industry, not a product.
Every one depends on the foundation: instrument-level traceability and controlled terminology. That is what TurboPlatform builds, from protocol to report.
One question to ask
“Ask any vendor, us included: why is AI in this step, at this level, and who decides?”
Talk to us about AI in your study workflow
We will show you where AI helps, at what level, and where we chose not to use it.
