In our last piece we argued that AI diligence is only as good as the evidence underneath it. That still holds. But a complete evidence base doesn't write the memo on its own, and we've watched plenty of strong source sets turn into thin analysis.
Hand two associates the same data room and you can get two very different reads. One summarises what the documents say. The other notices that net revenue retention looks healthy only because two large accounts expanded, and asks what happens to the thesis if they don't renew. The evidence was identical. The judgment applied to it wasn't.
The same gap shows up with AI. Frontier models are fluent, fast and widely read, and they will produce a confident answer from whatever you give them. What they don't do by default is work through a problem the way a diligence team does, with the intermediate steps that make the conclusion defensible. At Binocs we treat that reasoning as its own layer, sitting on top of the evidence layer, and we engineer it as deliberately as we engineer search.
Ask a general purpose model to assess a target's competitive position and you will usually get something that reads well. But if you look closer, the steps a consultant would insist on are often missing. Nobody defined the market before sizing it. The peer set was assumed, not built. A growth rate appears without the cohort analysis that would tell you whether it's durable.
We see three recurring problems. The first is skipped work: the model jumps to a conclusion because nothing forces it to do the analysis that should come before. The second is a missing lens. A model can retrieve a churn figure, but it has no built-in sense of what churn looks like for a vertical SaaS business at that scale. The third is drift. Run the same question twice with a slightly different prompt and the structure, emphasis and even the conclusion can move.
None of this is a criticism of the models. They are built to be good at everything, and diligence asks them to be very good at one specific kind of work. Closing that gap takes structure around the model, and that structure is what the analysis layer provides.
Three things the analysis has to deliver
Deeper insights. A deal team doesn't need a summary of the data room. It needs to know which metrics matter for this business, what good looks like against peers, which numbers are red flags, and what all of that means for the investment thesis. Insight is the step from fact to implication.
Output consistency. If two runs on the same evidence produce different conclusions, neither is useful for the investment committee. Consistency here means the same analytical method on every run: the same workstreams, the same frameworks and the same standard of rigor, so differences in the output trace back to differences in the inputs.
Stronger evidence. Every conclusion should sit on visible work. A partner reviewing the memo should be able to see how the analysis got from the source material to the recommendation.
How we build it: four parts of the analysis layer
1. A consulting-grade proof-of-work harness
The harness decides how rigorously each task has to be performed before the system is allowed to conclude anything. Binocs breaks the analysis into the same structured workstreams a consulting team would run, such as market definition, sizing, competitive landscape, customer economics and risks. Within each one, the intermediate work has to be done and shown.
In practice that means the peer set gets built before margins are benchmarked against it, and the market gets defined before anyone quotes a growth rate. If the work underneath a conclusion hasn't been done, the conclusion doesn't get written. This is the part of the layer that matters most for consistency, because it fixes the method, and it's also what makes the reasoning visible to whoever reviews it.
2. A domain intelligence engine
Retrieving the right information is half the job. The other half is interpreting it the way a PE investor or diligence consultant would. Our domain intelligence engine combines a Diligence Knowledge Graph with SME judgment and established industry frameworks and benchmarks.
That lens is what tells the system which metrics matter for a given business model, what looks good for a company of that size and sector, what counts as a red flag, and how a finding bears on the investment thesis. A 15% customer concentration figure means one thing for an enterprise software vendor and something quite different for a contract manufacturer. The engine is how the analysis knows the difference.
3. Dynamic context engineering
Generic mega-prompts tries to tell a model everything at once, and the model ends up weighing a lot of instructions that have nothing to do with the task in front of it. We do the opposite. For each analytical task, Binocs assembles the specific context, frameworks, evidence and instructions it needs, based on the industry, the company and the use case.
Say the task is assessing unit economics for a healthcare services roll-up. That step receives the benchmarks and frameworks for healthcare services, and nothing about software churn or retail margins. With a narrow, clear brief, it works like a specialist analyst rather than a generalist, so the output is sharper and more consistent from one run to the next.
4. A continuous learning and evaluation engine
Frontier models improve horizontally, getting a little better at everything with each release. Binocs improves vertically, on the specific work PE and consulting teams need done.
Every failure or issue we find becomes an eval with its own ground truth. We test our analytical instructions, workflows and agents against that growing set and refine them when they fall short. A mistake caught on one engagement becomes a test the system has to pass on every future one, so improvement is systematic and measurable rather than anecdotal.
Where it changes the decision
Take a B2B software target reporting 115% net revenue retention. A general-purpose pass would likely call that strong, cite a benchmark showing it's above median, and move on.
The analysis layer handles it differently. The harness requires the retention figure to be broken down by cohort and customer size before any conclusion is drawn. The domain engine knows that for a business at this scale, concentration in the expansion base matters as much as the headline number. Context engineering means the workstream is working with the frameworks for SaaS retention specifically, not a generic view of revenue quality.
What comes out is a more useful finding. Retention is above peer median, but most of the expansion comes from a handful of accounts, and the cohorts from the last two years are retaining noticeably worse than older ones. The headline is true. The thesis built on it needs testing:
- Are recent cohorts weaker because of product fit, pricing or a change in the customer mix?
- What does retention look like if the largest expanding accounts flatten out?
- Is the growth plan relying on expansion behaviour the newer customers haven't shown yet?
The call is still the deal team's. They're just making it with the analysis a senior associate would have done if they'd had the time.
What a diligence-grade analysis system should demonstrate
Each part of the layer pulls on the three outcomes differently. The harness does the most for insight depth and consistency, the domain engine for insight, context engineering for sharpness and consistency, and the learning engine for consistency over time.
If you're evaluating any AI system for diligence analysis, Binocs included, these are the questions we'd put to it. Does it show the intermediate work behind each conclusion, or only the conclusion? Does it apply sector-specific benchmarks and red-flag criteria, or general ones? Does it give the same structure and method on repeated runs? And when it gets something wrong, is there a way for that mistake to stop recurring?
What changes for the deal team
The evidence layer gets the right material on the table. The analysis layer makes sure that material is worked through properly: decomposed into the right workstreams, read through a PE and consulting lens, and held to the same standard every time.
For an associate, that means starting from a first draft that already has the cohort cut, the peer benchmarks and the red flags called out, and spending their time on the questions that need a person. For a partner, it means a memo where each conclusion can be traced through the analysis behind it as well as to its sources.
Models will keep getting smarter in general. What a deal team needs is a system that keeps getting better at diligence in particular, and that's the layer we've chosen to build.
