Most “competitive intelligence” in a young company is the founder reading at midnight. That is not a layer. That is a habit that dies the week the company gets busy.
What a layer is
A strategic intelligence layer is standing. It has named inputs, a place claims live, a cadence where they meet a decision, and a rule for killing a claim that has expired.
A briefing is a file. A layer is a system you can point at when someone asks “what do we believe about this competitor, and how old is that belief?”
SENSE is the five-move version I publish. Use it when models are already in the loop and you still decide from anecdote.
- Scan. Collection with a thesis. Watch the competitor surfaces, regulatory dockets, and customer signals your strategy actually depends on. Not the whole internet.
- Evaluate. Turn raw signal into a sourced claim with a confidence grade, instead of another summary nobody can act on.
- Navigate. Make the claim change something real: the watchlist tightens, the assumption register updates, the strategy moves.
- Simulate. Rehearse the future before it arrives. Ask what would happen if, and pressure-test the plan against it.
- Execute. Meet a decision or it never happened: a one-page brief in the Monday review, an alert the moment it matters.
Skip Simulate and you execute into a void. Skip Evaluate and you scan yourself into a trance. Skip Navigate and you have a prettier briefing habit.
The unit of work is one claim with a confidence grade, not another unread summary. An AI agent can draft the claim. It cannot own the Monday decision.
What a claim looks like
Write five lines. If you cannot, you do not have a claim. You have a vibe.
- Assertion. “Competitor B will ship usage-based pricing this quarter.”
- Source. The job post dated 3 August, the support-forum thread from last Tuesday, the customer call on Thursday. Name them.
- Strength. What would make you raise or drop the grade. One sentence.
- Grade. A working mark you are willing to say out loud. High / medium / low is enough. This is an authorial mark, not a validated scale.
- Expiry. A date. After that date the claim is dead unless someone re-opens it with new source.
That register is a thin evidence ledger. What was asserted, which source supports it, how strong that source is, what would retire the claim. It is not a bibliography dumped at the end of a memo.
Monday then asks one question: which live claim changes this week’s call? If none of them can, stop collecting.
Ninety days: Diagnose, Deploy, Evaluate
The book calendar is three gates, not four artifacts standing in for the protocol. Artifact 8 on the worksheets is the source. Use those templates. Do not invent a second kit.
SENSE also has no Decide phase. The layer prepares the choice. A person owns it.
Days 1–30 — Diagnose. Build nothing that looks like a product. Name the bottleneck. Write one sentence you can fail: this layer watches X, claims Y, feeds decision Z in meeting W. Inside the month you may draft a ten-signal watchlist and an assumption register, but they are notes, not a live system. Gate: if you cannot say X, Y, Z, and W out loud, you do not deploy.
Days 31–60 — Deploy. One system. Simplest architecture. Wired into a meeting that already exists. Instrumented from day one. A watchlist with owners, a register with grades and expiry, a chair, a 45-minute loop. Gate: the layer changed at least one real decision, or it confirmed no action with an evidence trace. A digest that arrives nowhere fails this gate.
Days 61–90 — Evaluate. Score the system against the kill criterion you wrote in week one. Then decide, in writing: expand, adjust, or kill. A retired claim stays visible for a week so people can object. Then it leaves the register. Silence is how zombie beliefs run a company.
An agentic workflow can fill rows. It cannot invent X, Y, Z, and W. If you cannot name the decision this layer will change, you are not ready for more agents. You are still in Diagnose.
A Monday that uses the layer
Picture a founder who sells to ten customers and is deciding whether to match a rumoured price cut.
Without a layer, Monday is a pile: three Slack screenshots, a half-remembered call, a tab left open since Thursday. The decision happens in the shower. Nobody can reconstruct why.
With a layer, Monday has four objects on the table.
- The watchlist row: “Competitor B pricing page, Friday check, founder-owned.”
- The live claim: “B is testing usage-based pricing on new accounts only.” Grade medium. Expiry 1 September. Sources named.
- The cadence rule: this claim may change the discount exception process, not the public price list.
- The kill rule: if Friday’s check shows the old list price and no new-account toggle, the claim dies in public.
The meeting is twenty minutes. The output is one sentence in the decision log: “No public match. One named exception if a customer brings a written B quote before 1 September.” Then people go back to work.
That is Navigate plus Execute. Scan and Evaluate happened earlier in the week. Simulate happened as a single question: “If B’s test is only on new accounts, what breaks if we cut list price for everyone?” The answer was “margin on the eight customers who were not leaving.” So the layer refused the dramatic move.
When the layer is fake
| Tell | What you actually built | First fix |
|---|---|---|
| A hundred “signals,” no owner | A reading list | Cut to ten. Name an owner or delete the row |
| Beautiful summaries, no grade | A briefing habit | Force assertion, source, grade, expiry |
| Claims that never reach Monday | A research hobby | One loop. If it cannot enter, stop collecting |
| No retired claims in ninety days | A museum of vibes | Kill in public. Date every row |
| Agents writing memos nobody reads | Midnight reading at scale | One Monday question. Everything else is optional |
What this is not
It is not a dashboard wall. A wall of charts with no claim, no owner, and no expiry is still midnight reading, only more expensive.
It is not a promise that agents replace judgment. SENSE exists so a model can produce a sourced claim instead of a vibe.
It is not a CHORUS substitute. A team that cannot hand work across a boundary will not become strategic by adding a watchlist. Install the coordination protocol if the week already got slower when you added agents.
It is not a surveyed success rate. Ninety days is an install calendar from AI Agents for Startup Strategy, not a study of how many founders finish it.
The book introduction teaches SENSE as a system. This page is the calendar. Start with ten signals and one Monday question. Everything else is permission to stay.