Reports on Demand: AI as the Implicit Intranet
Point AI at your own code, docs and email instead of building dashboards and wikis
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 68%
Most internal infrastructure is a routing layer between a question and a corpus that already contains the answer. The mechanism here removes the layer. Instead of maintaining an intranet, dashboards and business-integration systems, you unleash AI directly on the raw material a company already has: the code base, hardware designs, the supplier and vendor document store, even company email. Then you ask the question you actually wanted answered. How far are we from shipping? Who is the expert on this subsystem? Which division is short on resources? The AI does the digging and produces the artifact, including a chart, on the spot. Because generation is cheap, the artifact does not need to persist; you rebuild and refresh it on demand. A second-order effect is that hardware, software and AI specialists can each reach 20-30% into the others' work, so the glue between disciplines gets thinner and fewer handoffs are needed.
Origin
Naval gives this as the concrete answer to how AI actually helps at Impossible, a company that runs with no Slack and no project-management software. AI is not deployed as an explicit communication tool; it is used implicitly, as a reader of the company's own raw corpora.
Core principles
- 01Documentation exists to answer questions; AI can answer them from the raw corpus instead.
- 02A dashboard is a frozen query. Regenerate it on demand rather than maintaining it.
- 03The corpus you already have (code, designs, vendor files, email) is the knowledge base.
- 04AI turns specialists into partial generalists, which widens the interfaces between them.
- 05On-demand generation trades latency for maintenance, which is usually a good trade.
How to run it
- 1
Inventory the corpora you already own
List the places where your company's real state lives: the code base, hardware and product designs, the supplier and vendor document folders, and email. These are the inputs, not your wiki.
Pro tip If a fact is only in someone's head, it is not in a corpus. Note those gaps separately.
- 2
Use AI to summarize rather than to document
When you hit complicated code someone else wrote, or a paper you need the gist of, have the AI read it and summarize instead of asking the author or writing a doc.
Pro tip Summarize before you ask a human. It turns a meeting into a two-minute confirmation.
- 3
Ask AI who the expert is
Have it walk the code base and tell you who in the organization is likely to be an expert on a given topic, then route yourself to that person. This replaces the explicit internal directory.
Pro tip This pairs directly with a flat org: the AI does the routing that a hierarchy would otherwise do.
- 4
Point it at the operational corpora, not just code
Unleash it on designs, on the supplier and vendor database or file folder, and if you are willing, on company email. Broader access is what makes cross-cutting questions answerable.
Watch out Company email and supplier data carry real confidentiality and access-control exposure. Decide the boundary deliberately before opening it up.
- 5
Ask for the status artifact, not the status meeting
Ask questions like where are we, how far are we from shipping, who is behind and who is ahead, which division is short on resources, and have it draw the Gantt chart from what it found.
Pro tip Ask for its estimate of reality, not a restatement of the plan. The gap between the two is the report.
Watch out The output is an inference from the corpus, not ground truth. Sanity-check anything you are about to act on.
- 6
Regenerate instead of maintaining
Build dashboards on demand and update them on demand rather than running standing charts and business-integration systems. Accept some latency in exchange for zero maintenance.
Pro tip You may not want to regenerate every single time; cache the ones you look at daily and rebuild the rest fresh.
- 7
Let specialists reach into adjacent disciplines
Have AI people build their own throwaway software harnesses to test something, and hardware people write enough software to bring up a new device, rather than queuing on a specialist.
Pro tip Judge this code by whether it unblocks the next step, not by whether it is production-grade.
Watch out Cross-discipline code is fine for bring-up and testing and usually not fine for production deployment. Keep that line explicit.
In the wild
Rather than collecting status updates, you point AI at the code base, the designs, the vendor documents and company email and ask where the company actually is: how far from shipping, based on the real estimates and timelines, who is behind, who is ahead, and which division is short on resources. It produces the Gantt chart on the spot from evidence in the corpus rather than from what each team reported. The chart is disposable; when you want the current picture, you regenerate it.
→ Leadership gets a resourcing and schedule picture without a reporting layer or a status meeting cycle.
An AI engineer needs a harness to test a model change but has no software engineer free. Instead of queuing, they use AI to write the harness themselves. It is not production-deployment quality, but it is enough to run the test today rather than next week. The same move works in reverse: a hardware engineer writes just enough software to bring up a new device without waiting for the software team.
→ Cross-discipline handoffs shrink and each specialist becomes 20-30% capable in the neighbouring discipline.
Common mistakes
Treating AI as a new communication tool
The value here is implicit, not explicit. Bolting AI onto chat and project-management software adds a layer instead of removing the layers that AI made unnecessary.
Persisting the generated dashboard
Once a generated report is saved and maintained, it becomes another stale artifact to keep in sync. The point is that regeneration is cheaper than maintenance.
Shipping cross-discipline AI code to production
A hardware engineer's software or an AI engineer's harness is good enough to unblock work and usually not good enough to deploy. Blurring that line converts a speedup into a reliability problem.
Is it for you?
Best for
Small technical teams that already hold their real state in code, design files and email rather than in reporting systems.
Not ideal for
Situations needing audited, low-latency, always-consistent reporting, or teams that cannot give AI access to the underlying corpora.
From the transcript
“So, AI can do a lot of that digging for you. You don't need the explicit intranet as much anymore.”
“AI can constantly be doing this data analysis and digging and reporting for you. Reports on demand.”
“So, having AI just lets everybody do a little bit of everything. It makes them more generalist”
From the episode
'Nothing Ever Happens' Is Over