Research
Three research lines, one question: what happens between AI agents and people.
We take that question where it is hardest to answer and most costly to get wrong. Everything is published with its limits stated: 900 Reports are preregistered, independently verified and published with an error rate, from the next Report on. 900 Notes are fast, and say what they did not do.
Line 1
How agents perform in real work
Our first Report concluded that the binding constraints on enterprise AI are organizational rather than algorithmic. The Agentic Enterprise tests that, with its kill criteria written down first.
- The AI Execution Gap900 Report · published March 2026 · PDF below
- Nearly half of executives pulled back on AI agents. Half of what?900 Note 001 · published 9 August 2026
- The Agentic EnterpriseEvidence review · protocol approved · registration on OSF next · no sources collected yet
Line 2
AI and mental-health safety
A chatbot answering someone in psychosis is one of the hardest tests of how an AI system behaves with a person, and a psychiatrist can judge what a safe answer is.
- How chatbots answer psychosis, in SwedishLed by Dr. Pontus Inerup · a Swedish replication of a published benchmark of how chatbots answer prompts that show psychosis · protocol in preparation
Line 3
AI-assisted evidence synthesis
Whether AI-assisted review of medical evidence holds up when people make every judgment that carries weight.
- Drug repurposing for treatment-resistant depressionDr. Pontus Inerup, principal investigator · grant application in preparation
A subject is listed here only when it has a named owner. Ideas without an owner are not listed, because a list of directions nobody is working on is a prospectus rather than a pipeline.
Two kinds of data, kept apart
Research data
From our studies, including those with patients, under academic and ethical governance. It never feeds our products.
Deployment data
From the agents we run, under each client’s contract. It improves our agents, and it is never sold.
The AI Execution Gap
Why about 95% of enterprise AI fails or stalls, and what the top 5% do differently to reach measurable ROI
The number everyone quotes is not one number. "Failure" means something different in each survey it comes from. One measures organizations with no measurable profit-and-loss return. Another measures proof-of-concepts scrapped before production. Another measures projects abandoned after the pilot. The report reconciles the definitions, then asks why spending accelerates while outcomes stay this uneven. Its answer is that the binding constraints are organizational rather than algorithmic.
What this report is not
It was produced before our current method. It was not preregistered. It was not checked by an independent verifier. It carries no published error rate. Our next Report will carry all three.
900 Notes
What we are thinking about.
A Note is usually a few hours of work. No systematic search, no independent verification, nothing preregistered. Every Note states what it did, what it did not do, and what it could not check, in a box you do not have to go looking for.
A Note may raise a question. It may not settle one. If a Note reaches a conclusion that would change a client's decision, an investment or a public policy position, it stops being a Note and becomes a Report.
Nearly half of executives pulled back on AI agents. Half of what?
Two figures traveling as a single sentence turn out to come from two different surveys, of two different populations, with revenue thresholds an order of magnitude apart. Neither survey defines the thing it is counting.
Read the Note →
More as they land. Monthly is the goal.
Two series, labeled before you read a word.
| 900 Notes | 900 Reports | |
|---|---|---|
| What it is | What we are thinking about, and why. Fast. | What we found, and how confident we are. |
| Cadence | Monthly is the goal | Two or three a year |
| Preregistered | No | Yes, on OSF |
| Independently verified | No | Yes, and the error rate is published |
| Systematic search | No, and it says so | Yes |
| Can it settle a question? | No | Yes |
Corrections
Errors are corrected in the published artifact, in place, dated, with the change visible. An error that changes a headline conclusion retracts the piece publicly, with reasons.
Corrections to date: none. We have published one Report and one Note. When the list fills, it fills here.
Found something wrong? hello@900labs.com
Get them when they publish.
Notes as they land, monthly as the goal. Reports two or three times a year. Nothing else.