Everyone's raving about AI literature review tools as a magic bullet for grant writing. I decided to put Iris.ai through its paces on a recent proposal for a cloud security project. My goal wasn't just to find papers, but to genuinely identify a defensible research gap—something these tools often gloss over. Here's my actual process, warts and all.
First, the setup. I didn't just dump in a broad topic. I started with three key "seed documents": our own preliminary abstract, a recent seminal paper in the field, and the specific call text from the funding body. I used the "Focus" tool to extract what Iris.ai thought were the core concepts. It was… okay. It heavily favored the terminology from the seminal paper, almost ignoring the nuances in our abstract. I had to manually adjust the weightings.
My workflow then looked like this:
* **Initial broad search:** Used the extracted concepts to run a broad discovery query across their database. The volume was high, but relevance was mixed. The ranking seems to prioritize citation count over conceptual similarity, which is a problem for emerging niches.
* **Filtering is key, and costly:** To get anywhere useful, you immediately need their filters. I iterated by:
* Excluding obvious outlier disciplines (e.g., biomedical).
* Setting publication date to the last 3 years.
* Using the "keyword in title" filter to narrow scope.
* This is where the "efficiency" promise starts to crack. Each filter iteration feels slow, and you burn through your monthly "analysis units" quickly.
* **Synthesizing the gap:** I exported the top 50 results and moved to the "Data extraction" module. I defined a custom set of data points to pull: *Methodology*, *Target Environment* (e.g., multi-cloud, hybrid), *Claimed Novelty*. The extraction was imperfect, requiring manual verification, but it did cluster papers visually. The gap became apparent not from a single missing paper, but from a clear cluster of papers using Method A in Context X, and a total absence of Method A in our target Context Y. That was the usable insight.
The final value was solid, but the path there wasn't as smooth as the marketing suggests. The resource consumption for a single, proper gap analysis was substantial. For a small team or a PI on a tight budget, the cost per actionable insight needs careful calculation. It saved me time versus a manual snowball search, but it introduced its own overhead in learning the tool and managing its quirks.
Question everything
The weighting issue with seed documents is such a real pain point. I've seen similar behavior in sales engagement platforms when they try to auto-score leads - they'll latch onto the loudest signal and ignore the nuance you actually care about.
Your note about filtering being costly rings true, both in time and in platform credits. That "volume over relevance" initial result reminds me of doing a broad LinkedIn Sales Navigator search - you get a huge list, but the real work is building those boolean strings to carve it down. Does Iris.ai let you save those filtered views, or do you have to start from scratch each session?
spreadsheet ninja
You hit the nail on the head with the LinkedIn comparison. That initial data dump is just raw material.
To your question: No, you can't save filtered views in a practical way. You get a "workspace" where your results live, but any major change in your search parameters or a new session often means rebuilding your logic. It's not a persistent, reusable filter set.
The core problem is the same as in sales tech: these platforms prioritize showing you they *can* find things over whether those things are *useful*. You spend your credits, and your time, doing the vendor's job of quality control.
The analogy to sales engagement platforms is apt, but it actually undersells the problem in the research context. In sales, a loud but irrelevant signal is just noise. In literature analysis, it's actively misleading. When the tool latches onto the most frequent terminology from a seminal paper, it isn't just missing nuance, it's reinforcing the existing consensus you're trying to move beyond.
You're correct about the lack of saved filter states. In practice, I've resorted to a manual, external logging system to track the effective boolean-like strings. The cost isn't just in rebuilding the logic, but in the irreproducibility it introduces. If you can't precisely recreate a view, you can't systematically compare how a research gap shifts when you adjust the weighting of, say, "zero trust" versus "microsegmentation." That's a critical failure for a tool marketed on systematic discovery.
The underlying issue is likely a naive TF-IDF approach in their concept extraction, not a semantic one. It's prioritizing term frequency within your seed docs, not conceptual relevance to your unique combination of them.