Contact

Explore Our Portal

CASE STUDY · AI-READY LABS

Systems for Enabling AI Scientists at the Bench

How SaphoBio wired three HPLCs into one system Claude can run overnight, and got about 14 times more method development per week.

Oct 27, 2026

Talk to a Scientist

View Other Case Studies

Abstract

SaphoBio integrates the instruments a lab already owns into one system an AI scientist can operate, with every run recorded and auditable. In our own lab, Claude now develops HPLC methods overnight. The goal is an AI scientist that proposes improvements and proves them before they reach production.

Cameron Bravo, PhD

SaphoBio Lab Services

LIVE · HPLC METHOD DEVELOPMENT · OVERNIGHT

UV · mAU vs min

Illustrative chromatogram · UV absorbance vs. retention time

~14×

weekly HPLC productivity

unattended overnight runs

HPLCs: one Waters, two Agilent

Level 3

closed-loop lab autonomy

The handoff

~2×

instrument hours on method development

Our scientists prepare solvents, change columns and check results by day. Overnight, Claude runs method development on three HPLCs and starts each run as soon as it has reviewed the last.

DAY · scientists prepare, run and verify

NIGHT · Claude runs from the handoff brief

handoff →

← morning review

Problem

AI can plan experiments, but people still run them

Most AI scientists stop at the plan. A person sets up each run and records the result, so the loop closes at human speed. Every instrument has its own software and file formats, and results land in spreadsheets without the context needed to interpret them.

FIG. 01

The same four steps, open and closed. The difference is who carries each result back to the next plan.

Toolkit

How the toolkit works

Our hardware and software connect the devices a lab already owns into one system, with no new equipment or vendor APIs.

A lab can adopt these layers one at a time, and permissions widen only as the AI scientist proves it can run a class of work reliably.

Instrument control

The AI scientist works each instrument through its own software, on the same screens a scientist uses.

Data and audit

Every run, parameter change and decision is recorded with its context, down to the solvent lot and the installed column. Failed runs are kept too.

Planning

The AI scientist plans each experiment against an objective your scientists set and decides what to run next.

Oversight

Scientists set the strategy, keep instruments stocked and serviced, brief the AI scientist at each handoff, and can stop it at any time.

FIG. 02

Reference architecture. Oversight controls attach at the layer they govern.

Maturity

Five levels of lab autonomy

Most labs sit at level 0 or 1. At level 3, the AI scientist chooses each experiment within a strategy people set. At level 4, it proposes its own improvements and designs the experiments to test them, and scientists decide what reaches production. Level 4 is the toolkit’s goal.

FIG. 03

Maturity levels. What the AI scientist does and what people do at each level.

Proof

HPLC method development in our lab

Claude develops methods for new APIs on our three HPLCs, one Waters and two Agilent, working the vendor software through RustDesk, a remote desktop tool. One of our HPLC experts agrees a strategy with Claude and prepares the solvents and columns it calls for. Within that strategy, Claude chooses the gradient, column, temperature, flow rate and pH for each run until the method passes system suitability. That is level 3.

Claude roughly doubles each instrument’s hours on method development and never waits for a person between runs. Together, that gives about 14 times as much progress per week.

FIG. 04

The method-development loop runs unattended inside an expert-set strategy.

NIGHT 01 · 25 RUNS

BPC-157

0.18 → 0.04

shoulder height, as a fraction of the apex

The peptide tailed on our production methods, with a shoulder holding about 16% of its area. Claude tested three explanations. Neither the gradient nor the ion-pair reagent mattered, but raising the column temperature from 35 to 65 °C shrank the shoulder. That points to a slowly interconverting form of the peptide rather than an impurity, so the shoulder belongs in the potency result. Claude proposed a 55–60 °C method.

NIGHT 02 · 28 RUNS

Atropine sulfate

8 → 6 min

method runtime, with R² 0.99999

Claude shortened a new method from 8 to 6 minutes and cut the injection from 7 to 2 µL, which fixed the detector saturation that had bent its calibration curve. The method is linear from 0.0625 to 1 mg/mL (R² 0.99999), with 0.08% RSD across six replicate injections.

On both nights a scientist approved the plan at handoff, and Claude shut the instrument down before morning.

Acceptance criteria

A method passes when replicate injections of a standard meet the system suitability tests in USP–NF ⟨621⟩ Chromatography (official 1 June 2026).

Test	Shows	⟨621⟩ default
Resolution	API peak separated from its nearest neighbor	Set by monograph or method
Symmetry factor	Peaks symmetric enough to integrate reliably	0.8–1.8 for the standard peak
Plate number	Column separating efficiently	Set by monograph or method
Repeatability	%RSD of peak response across injections	5 injections if limit ≤ 2.0%, else 6
Sensitivity	Impurities can be quantified	S/N ≥ 10 at reporting threshold

Passing methods then go to formal validation under ICH Q2(R2).

Roadmap

Next steps toward level 4

Next, Claude will run serial dilutions on an Opentrons liquid handler and help model melting temperatures of custom oligonucleotides.

After that comes our daily lot release testing for 503A pharmacies: sterility, endotoxin, HPLC potency, pH and antimicrobial effectiveness. Each test we move onto the system adds to the record that more autonomy rests on. At level 4, we will ask Claude where our testing could improve and let it run the experiments to make its case. A change reaches production only if the evidence convinces our scientists.

FIG. 05

Each instrument added widens what the AI scientist can do.

06 · Your lab

Working with SaphoBio

Better models will arrive regardless. What separates labs is whether their instruments, data and controls are ready for an AI scientist, and that is an integration problem we solve. We recommend starting with one well-defined, high-volume assay, as we did with HPLC, and widening from there.

Talk to a Scientist

Visit saphobio.com →

References

King, R. D. et al. The Automation of Science. Science 324:85–89 (2009). doi:10.1126/science.1165620
Williams, K. et al. Cheaper faster drug development validated by the repositioning of drugs against neglected tropical diseases. J. R. Soc. Interface 12:20141289 (2015). PMC4345494
Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624:570–578 (2023). doi:10.1038/s41586-023-06792-0
Lu, C. et al. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292 (2024).
Laurent, J. M. et al. LAB-Bench: Measuring Capabilities of Language Models for Biology Research. arXiv:2407.10362 (2024).
Gottweis, J. et al. Accelerating scientific discovery with Co-Scientist. Nature 655:487–496 (2026). doi:10.1038/s41586-026-10644-y. Preprint first posted as “Towards an AI co-scientist,” arXiv:2502.18864 (2025).
Bartley, B. et al. Building an Open Representation for Biological Protocols. ACM J. Emerg. Technol. Comput. Syst. 19(3) (2023). doi:10.1145/3604568
Autoprotocol specification, autoprotocol.org. SiLA 2 standard, sila-standard.com.
USP–NF ⟨621⟩ Chromatography, official 1 June 2026. doi:10.31003/USPNF_M99380_10_01
ICH Q2(R2) Validation of Analytical Procedures, effective 14 June 2024. EMA