EXPOSED
The modal natural sciences manager splits time between screen work that AI now does well — drafting grant narratives and progress reports, summarizing literature, building budget spreadsheets, reviewing protocols for completeness — and work that resists it: deciding which research lines to fund or kill, hiring and developing scientists, walking the lab floor on safety and compliance, and owning results in front of funders, regulators, and executives. No license is required for the role itself, so the protection is organizational accountability rather than statute. The judgment tier is real and consequential, but the documentation and reporting load that fills much of the week is exactly what large language models compress.
Mixed — a routine tier and a judgment tier. Grant-narrative drafting, quarterly progress reports, literature landscape summaries, and budget rollups are now first-drafted by a model in minutes, but portfolio triage — killing a project because the assay reproducibility is drifting and the PI is defending it too hard — plus headcount decisions and IACUC/IRB negotiation stay manual, which lands this at mixed rather than resistant.
Some physical or field component. A working R&D or lab manager still walks the floor for chemical hygiene plan compliance, signs off on hood certifications and waste manifests, tours field sites and pilot plants, and shows funders the bench — but the majority of the week is Teams, Excel, and eLN review, so the physical component is recurring rather than defining.
Certification preferred, not legally required. The management title requires no license; what protection exists comes from being the named PI, the responsible official on an EPA or select-agent registration, or the signatory on an institutional assurance — role-attached duty rather than a credential a statute reserves for you, and many holders carry none of it.
Some relationship component. Program officers at NIH/NSF/DOE and internal executives fund the person as much as the proposal, and senior scientists stay because of who they report to — but the deliverable is data, milestones, and publications that survive on their own merit, so the relationship shortens review cycles without being the product.
Exists to be accountable for ambiguous calls. You decide which of six programs gets the next FTE on incomplete preclinical data, whether a safety incident is reportable, whether to disclose an anomaly to a sponsor before the readout, and whether a scientist's finding is solid enough to bet a milestone payment on — calls with no procedure, real dollars, and your name on the outcome.
Has AI actually changed your work?