Audit-ready copilot that turns Claude model sessions into reproducible experiment packets for pharma discovery teams.

1. Claude is now a live interface for quantitative scientific models rather than just a text assistant, which expands the set of scientists who can invoke them.
2. Independent coverage says the interface is the bottleneck for adoption, so workflow control around that interface becomes the natural next layer to buy.
3. Natural-language access removes the coding barrier, which raises the urgency for reproducibility and review before model outputs influence assay decisions.
4. SandboxAQ is already extending the same interface to AQPotency and AQCell, so the problem will compound as more model-backed decisions enter one discovery workflow.

Catalyst. SandboxAQ's Claude launch shows that quantitative models are moving from specialist tooling into a mainstream LLM interface, creating immediate demand for controls that make those outputs usable in real R&D decisions.

The product connects Claude to approved internal and external quantitative models through an MCP control plane built for scientific workflows. Scientists can ask natural-language questions, but each workflow is constrained by assay-specific templates, parameter bounds, required controls, and citation of the underlying evidence. The software records prompt history, datasets, model versions, and uncertainty notes, then packages the result into an experiment brief a bench team or CRO can actually execute. It also compares outputs across model versions and flags when a recommendation relies on out-of-domain assumptions or missing validation. The initial product focuses on potency ranking and next-experiment planning for hit-to-lead programs where each bad handoff costs weeks of assay time.

What's different. SandboxAQ and similar vendors sell model access, while incumbent lab software stores downstream records; this company owns the missing layer in between: governed conversion of natural-language model sessions into executable experiment packets. Because it is model-neutral and workflow-specific, it can sit across internal models, vendor models, and future MCP endpoints instead of betting on one scientific model stack. Over time, its proprietary asset becomes the library of approved assay templates, review policies, and model-to-outcome traces that make scientific AI usable at enterprise scale.

Jobs to be done

flowchart LR
  Buyer[Discovery team] --> Pain[Model outputs are hard to trust and operationalize]
  Pain --> Product[Governed Claude to experiment packet layer]
  Product --> Outcome[Faster and reproducible assay decisions]

Executive takeaways

• Large-pharma scientists are already using copilots as a default interface, but adoption falls in the most regulated and data-messy workflows where trust breaks [5][24].
• The why-now is concrete: Claude-accessible scientific models and MCP widen model access faster than informatics teams can manually govern outputs [1][2][3][4].
• Incumbents are moving fast, but they mostly win when work already lives inside their own stack; the white space is model-neutral capture of external Claude sessions and conversion into assay-ready packets [6][7][13][14][15].
• The product should be sold as governed workflow infrastructure, not as a general chatbot, because the pain is manual handoff, reproducibility loss, and slow CRO or wet-lab packet preparation [8][10][11].

Market definition

The relevant category is governed scientific-workflow software that captures LLM- and model-mediated discovery reasoning, links it to source data and approvals, and exports execution-ready experiment packets rather than loose chat transcripts [3][6][8][17].

Customer and buyer

Primary users are translational pharmacology and medicinal chemistry scientists who can now access advanced models through natural-language interfaces but still need discovery informatics leaders to make outputs reviewable, shareable, and executable. The economic buyer is the Head of Discovery Informatics or VP Computational Chemistry because the purchase touches validation, security, data structure, and cross-team workflow design [5][9][11].

Buying triggers

• A team enables Claude- or MCP-accessible scientific models and suddenly has more people producing model-backed recommendations than it has specialists available to document and review. [1][2][3][4]
• Program leaders realize weekly review decisions still require manual PowerPoint, notebook, and CRO-PDF synthesis before anyone can approve an assay request. [8][10][11]
• An AI pilot reaches security, IP, or compliance review and stalls because provenance, validation, and auditability are not yet reliable enough for scaled rollout. [5][6][9][21][22]

Willingness to pay

Willingness to pay should be high when framed as program infrastructure: incumbent platforms already sell customized enterprise bundles with validation support, while users still spend hours to weeks on report prep, CRO-data wrangling, and institutional-memory recovery. [9][10][11]

Category dynamics

Growth signal 9.9% CAGR

Validation signals

• 89% of surveyed biotech scientists already use copilots or reasoning tools as a first stop for interrogating data.
• Benchling reports that scientific AI report prep can drop sharply when the workflow is structured and source-linked.
• SandboxAQ explicitly says advanced quantitative models can now be accessed through plain-English prompts inside Claude.
• Benchling’s Anthropic integration already markets one-click traceability and automatic audit-log carryover, proving buyers care about this layer.

Regulatory & technical constraints

• Electronic records in regulated environments need validated systems, retrievable records, and secure time-stamped audit trails.
• Approval workflows need signature meaning, timestamp, and non-repudiation features if they are to substitute for manual review records.
• AI governance increasingly requires documented risk management, human oversight, and traceability rather than opaque model outputs.
• Model recommendations are only as reliable as the structured data and contextual metadata feeding them.

Competition is real but fragmented. Benchling and Dotmatics own structured-system-of-record territory, Schrödinger owns DMTA collaboration, BenchSci owns evidence-grounded disease-biology copilots, and SandboxAQ owns the model-access wedge. None of them clearly owns cross-model, external-Claude session capture plus experiment-packet governance as a standalone layer [6][7][12][13][14][15].

Why incumbents do not win by default

• Cloud platforms. Claude and MCP increase access, but the platform itself does not become the scientific system of record or approval workflow by default.
• ELN and informatics suites. Benchling-like platforms win when work already happens inside their data model, but they do not automatically capture external sessions or model decisions that occur outside their boundary.
• Model vendors. SandboxAQ-class vendors differentiate on predictive models and licensing, not on neutral provenance, cross-model comparison, or reviewer sign-off across the broader workflow.
• Computational design suites. Schrödinger-class tools are strong for in-platform DMTA collaboration, but their collaboration model is still coupled to their design environment rather than a portable packet layer across mixed systems.

This company should be built as a governed experiment-packet layer for large-pharma hit-to-lead teams that are starting to use Claude-accessible quantitative models in weekly compound-prioritization work. The first user is a translational pharmacology or medicinal chemistry lead who wants model-backed next-step recommendations without waiting on a computational specialist. The economic buyer is the Head of Discovery Informatics or VP Computational Chemistry because the purchase touches validation, model access policy, and CRO handoff. The MVP should not try to replace Benchling, Dotmatics, LIMS, or the upstream model vendor; it should capture external Claude sessions, enforce approved assay templates, and export reviewer-signed packets that a bench team or CRO can execute. The best wedge is U.S.-led small-molecule potency-ranking and next-experiment planning inside top-30 pharma because the workflow repeats weekly, the handoff pain is visible, and the validation surface is narrower than broader scientific AI governance. Research supports real demand for traceability and workflow control, but it does not yet prove buyers will fund a neutral layer before waiting for Benchling, Dotmatics, or SandboxAQ to extend their products. The company can win if it becomes the fastest path from free-form model query to audit-ready assay packet across mixed model stacks, then compounds into packet templates, review policies, and packet-to-outcome traces. The board-level question is whether early pilots show enough cycle-time reduction and specialist-time savings to create budget pull before incumbents close the gap.

Problem

• As Claude and MCP expose quantitative models to more scientists, weekly hit-to-lead decisions create irreproducible prompts, hidden assumptions, and manual packet prep before anyone can approve an assay request.
• Discovery informatics teams must govern model-backed recommendations across external sessions, multiple model vendors, and CRO handoffs, but ELNs and model vendors mostly capture records inside their own stack after the fact.
• Each bad handoff can cost weeks of assay time and erode trust in scientific AI, so buyers need reviewability before they need broader automation.

Solution

• Capture Claude session context, approved datasets, model version, assumptions, and reviewer sign-off in assay-specific packet templates for potency ranking and next-experiment planning.
• Export an execution-ready experiment brief for internal labs or CRO partners while blocking packet generation when data, parameter bounds, or provenance are incomplete.
• Compare outputs across model versions and surface out-of-domain warnings so informatics leaders can scale model access without surrendering control.

Why we win

• The wedge is model-neutral governance between Claude sessions and the experiment record, not another model vendor or ELN replacement.
• The buying trigger is specific: once a team enables Claude-accessible scientific models, review meetings generate more model-backed recommendations than specialists can manually document.
• Wet-lab and CRO handoffs create measurable ROI through faster packet preparation, lower specialist mediation, and fewer review loops.
• Reusable packet templates, reviewer policies, and packet-to-outcome traces can compound into switching costs if the startup lands early reference programs.

Milestones

flowchart LR
  Wedge[Hit-to-lead packet wedge] --> MVP[Governed Claude session to packet MVP]
  MVP --> Proof[Faster approvals and cleaner CRO handoffs]
  Proof --> Expansion[Multi-workflow scientific AI control layer]

Founding team

Experiment roadmap

Risk assessment

What must be true

• At least 3 of the first 10 target pharma accounts treat external-Claude packet governance as a budgetable problem this year.
• The first 3 pilots cut packet-preparation time by at least 50% without increasing scientific rework.
• More than 80% of packet fields can be auto-generated from session and model context with less than 15 minutes of reviewer edits.
• At least 2 early pilots convert into annual production contracts of $250k or more within 6 months of pilot completion.
• At least one production customer expands to a second program or adjacent workflow within 12 months.

Open diligence questions

• Will discovery informatics leaders buy a neutral packet layer now or wait for Benchling, Dotmatics, or SandboxAQ to extend their stack?
• Is potency-ranking and next-experiment planning the best first workflow, or does another packet type convert faster with buyers?
• What KPI opens budget first in practice: specialist-time savings, cycle-time reduction, packet completeness, or compliance posture?
• How much manual review must remain for buyers to trust the system in assay-planning decisions?
• Can the company secure enough rights to build packet-to-outcome benchmarks without triggering IP or data-sharing objections?

Model sanity

• Revenue engine. Base-case revenue is driven by turning 2 early pilots into production, adding 3 more programs through Y2, and ending Y3 with 9 live governed programs at roughly $300K ACV.
• Must go right. Pilot-to-production conversion must stay near the 50% business-plan target and at least one account must expand to a second program within 12 months.
• Model breaks if. If buyers wait for incumbents and the cycle drifts toward 9 months, downside cash falls toward roughly $0.1M and the company needs a bridge before Y3 ends.
• Next-round proof. The next financing is supported once the company exits seed with 5 production programs, one repeatable expansion motion, and proof that deployments stay under 10 weeks.

Scenarios

Sensitivity

flowchart LR
  Targets[Target pharma programs] --> Pilots[Paid pilots]
  Pilots --> Production[Production governed programs]
  Production --> Expansion[Second-program expansion]
  Expansion --> Revenue[Subscription and usage revenue]
  Revenue --> GrossProfit[72% gross profit]
  GrossProfit --> Opex[Hiring plus compliance and delivery spend]
  Opex --> Cash[Ending cash]

Flags: The model assumes pharma buyers fund a neutral packet-governance layer now instead of waiting for Benchling, Dotmatics, or SandboxAQ to extend their stack. · Revenue per FTE only clears the low end of software benchmarks by Y3, so any extra services work or early hiring can meaningfully worsen burn efficiency. · The base case ends Y3 with 2 live pilots still converting, so next-round quality depends on maintaining the modeled pilot-to-production cadence. · Gross margin can slip quickly if validation packages, CRO exports, or data-mapping work remain bespoke account by account.

• Platform encroachment. Scientific model vendors or Claude platform providers could add basic provenance and workflow packaging themselves. Mitigation: Start with model-neutral approvals, cross-vendor comparisons, and CRO handoff workflows that single-model vendors do not own.
• Slow validation cycles. Pharma teams may require months of shadow-mode evidence before trusting a new layer in assay planning. Mitigation: Launch on one potency-ranking workflow, benchmark against historical campaigns, and prove cycle-time reduction before broader rollout.
• Scientific liability. A bad recommendation or poorly bounded model output could undermine trust quickly in a high-stakes R&D environment. Mitigation: Keep a human approval gate, expose uncertainty and source provenance, and block packet export when inputs fall outside validated ranges.