The AI stories this month are not about chatbots. They are about plumbing. Two of them come from Five9, which sells contact center software. The third is Transformer Lab, a company that wants AI to run entire research pipelines. All three share a common shift: AI is moving out of the demo and into the operations layer, where the hard problems are identity, routing, and audit trails rather than prompt quality.
Take Transformer Lab's Primus. The company opened access to the tool today. Primus runs dozens or hundreds of agents, each using a different model sized for its task. Some agents are tiny. Some are frontier models from big-name providers. Transformer Lab claims a single prompt can take a research question from hypothesis to a completed paper in hours or days, without additional human direction.
The claims deserve scrutiny. Transformer Lab says it produced and published 30 "Masters to PhD-level" research papers in 30 consecutive days. The domains span materials science, physics, seismology, protein design, and 3D vision. One paper has already received a citation in research published by the Google DeepMind team. That is the concrete evidence, and it is thinner than it looks. A single citation shows the paper was read and referenced. It does not show the research passed peer review. The company does not explain how it assessed the papers as "Masters to PhD-level." Its terms forbid users from submitting the produced papers to journals, and it asks human volunteers to review them instead.
Academics are skeptical. Alessandra Buccella wrote for The Conversation that "the very existence of science as a source of authoritative knowledge about the natural world fundamentally depends on human life." I understand the discomfort, but I read the DeepMind citation as a real signal. The pace of output is a genuine change in what a small team can attempt. The honest framing: Primus is not replacing scientists. It is replacing the labor of reading, summarizing, and drafting, which is a large slice of what graduate students actually do. Whether that helps science depends on whether publishing practices change to match.
On the contact center side, Five9 rolled out FIS 2.0, its identity service, plus a controlled availability feature called Automated User Management. The identity work matters because contact centers run on permissions. Agents need access to the dialer, the CRM, the knowledge base, and supervisor dashboards. When an agent leaves, that access has to disappear. FIS 2.0 adds strong password policies, multi-factor authentication, and single sign-on. AUM contributes SCIM, the industry-standard API for provisioning users, and a feature called Circles, which automates the assignment of skills and applications.
The value is not exotic. It is that onboarding and offboarding stop being manual tickets. SCIM handles the identity provisioning. Circles handles the profile data. Together they turn a week of administrative work into a configuration. For a large contact center with seasonal hiring, that is real money, and it also makes compliance audits less painful because you can point at an automated deprovisioning log instead of a spreadsheet.
Five9 also announced Fusion for ServiceNow. The integration pipes real-time transcription from Five9's TranscriptStream into the ServiceNow workspace. ServiceNow's Now Assist converts those transcripts into case summaries, resolution notes, and recommended actions. AI Search surfaces knowledge articles at the moment an agent needs them. Five9's unified routing engine uses customer and case data from ServiceNow to send each interaction to the best available agent the first time.
The routing claim is the part I find most interesting. Five9 says it routes based on skills, history, and real-time workload pulled directly from ServiceNow case data. That is the difference between a queue and an actual routing decision. The transcript summarization is useful, but it is becoming table stakes in 2026. The routing integration is the piece that actually moves first-call resolution rates.
The pattern is clear from last week's Five9 Fusion for Salesforce announcement too. Five9 is building the contact center as a data layer and letting the CRM or ITSM platform own the AI surface. That is a deliberate architectural choice. The customer can activate Einstein for Service or Now Assist without ripping out the contact center. Choose your AI, not your telephony.
My practical takeaways:
Treat AI procurement as infrastructure procurement. Ask about integration standards, SCIM support, and routing APIs before you ask about model quality. The model is a switchable component. The wiring is not.
For research automation tools, verify the output pipeline rather than the headline. A citation from a DeepMind team is a data point, not a certification. Check whether the papers can enter the peer review system and what review path exists.
For identity, check whether automation covers deprovisioning as well as provisioning. Most tools nail the onboarding flow and quietly drop the offboarding one. SCIM does both, and Circles-style profile automation is what makes it stick.
The throughline across all three announcements is that AI vendors are now competing on integration quality, not on benchmark scores. Five9 wins by talking cleanly to ServiceNow and Salesforce. Transformer Lab stakes its claim on the breadth of agents it can orchestrate. The model is the commodity. The wiring is the product.
None of this answers whether AI should write research papers. That question has a political answer, not a technical one. But the technical trajectory is legible: AI is being measured by how well it plugs into your identity system, your case data, and your routing logic. If a vendor cannot do that, the underlying model does not matter.