Reports are getting cheap. Decisions are not.

This is the first strategy1 Review. Each month it takes one strategic question and works it through in the open: the argument, the strongest case against it, one number traced to its source, and what we measured.

We start with a question about our own trade. AI has made the strategy report easy to produce. If the report is cheap, what is a buyer of strategy actually paying for? Our answer is the lead. Our red team's answer, five ways an AI-made strategy goes wrong, follows it. Then we set out how we will measure the speed we claim, before we publish a single figure.

There are no published cases in this issue. We would rather say so on the first page than fill the space with numbers nobody can check.

For a long time the expensive part of strategy was the document. Research, interviews, modelling and drafting went into a report over several weeks, and the fee followed the hours. The report was the thing you could hold, so it was the thing you paid for.

That part is getting cheap. A language model can read the sources on a market, summarise them, draft options and argue against each of them. We use one in our own work every day: Claude reads, drafts and explains; our own code calculates.1Show source for note 1 Work that took weeks of reading and drafting can now take days. If the report were the product, its price would only fall from here.

But a board was never really buying a report. Three things stay expensive, and AI does not make any of them cheap on its own.

• A model of your own market. A report describes a market on the day it was written; a deck freezes the strategy at the moment it is presented. What a company needs is a model of its own markets, competitors, capital, context and decisions that changes when the evidence does. That is the Strategy Twin. Building it is careful work, and its structure is confirmed by a person before anything is built on it.
• A decision someone is accountable for. A model can draft a recommendation. It cannot answer for it. The decision has to be argued against, chosen and signed by people with the authority to change it. The output is a signed decision memo, not a deck.
• The work of carrying it out. Strategy that sits in a document changes nothing. A decision has to reach the systems where a business runs: CRM opportunities, budgets, briefs and run-sheets. Then the results have to come back, so the next decision starts from what actually happened.

That is why our work runs as a loop with people at the gates. Nothing advances past a gate without a signed approval record.2Show source for note 2 And nothing is written to a client's systems until the named owner approves a dry-run of exactly what will change.3Show source for note 3

What does this mean for how you buy strategy? If the report is cheap, judge the work on what is not: the model, the decision and the execution. Ask who signs. Ask where each number came from and how sure the team is of it. Ask how the plan fails. Ask what happens in the second month, when a competitor moves or a KPI drifts.

It is also why we do not sell hours. Our engagements have a fixed scope, never hours.4Show source for note 4 When the report costs less to make, a fee that follows the hours measures the wrong thing.

The lead argues that what stays valuable is the model, the decision and the execution. Our red team's job is to argue the other side. Its strongest case is this: an AI-made strategy can look finished and be wrong in ways a reader cannot see. The writing is fluent, the numbers are precise and the options are tidy. None of that is evidence.

These are the five failure modes we design against, and the control we use for each.

1. 01

   1. Confident numbers with no source

   What goes wrong

   A fluent paragraph carries a figure nobody can trace back to where it came from.

   The control

   The evidence ledger. Every number and claim links to a source, a date and the agent that produced it, and the narrative cannot add a fact that is not in the ledger.

2. 02

   2. Estimates that read like facts

   What goes wrong

   A model's guess sits beside an audited figure, in the same type, with the same weight. Or a language model does the arithmetic.

   The control

   Confidence grades and code. Every key figure carries a grade, and a C-grade figure is labelled as an estimate wherever it appears. Calculations run in our own deterministic code; the LLM explains, it does not calculate.

3. 03

   3. Options nobody argued against

   What goes wrong

   The first plausible option becomes the plan because nothing pushed back on it.

   The control

   The red team. Four critics, competitor, regulator, investor and customer, argue against every option before it reaches the Council, and two judges check the result. “How this fails” is a required section.

4. 04

   4. Plans no one signed

   What goes wrong

   A recommendation arrives with no owner, so no one can be asked later why it was chosen.

   The control

   Three human gates. A named person signs at Model (H1), Decide (H2) and Deploy (H3), and Domain Fellows sign the parts that carry professional liability.

5. 05

   5. Changes nobody approved

   What goes wrong

   An AI system writes a plan straight into the systems a business runs on.

   The control

   Governed writeback. Before any write, the change passes a policy check, the named owner approves a dry-run of exactly what will change, and the approval is recorded in a hash-linked audit trail.

Where is the lead wrong? Controls reduce these failures; they do not remove them. The lead is wrong if the controls become ritual: a gate signed without reading, a grade attached without checking, a red team that only argues against the options nobody wanted. Every strategy we ship carries its strongest counter-arguments and the conditions under which it fails5Show source for note 5, and that rule is only as good as the attention behind it.

It is also wrong for a buyer who needs a report and nothing more. If no one has to model, sign or run anything, the cheap report may be exactly the right purchase.

One figure this month, and it is our own. When a specialist gives a baseline in calendar days, we convert it to working hours at 8 hours a working day, and we show the conversion.6Show source for note 6 Any baseline written in days passes through this number before it becomes an index, so it is worth showing where it comes from.

The figure

8 working hours to 1 working day.

Where it came from

Our published method for the Compression Index. It is a convention we set, not a measurement, so the source is our own written rule.

How old it is

Published October 2026, the month of this issue.

Its grade

A, verified primary source. The grade says we have stated our own rule correctly. It does not say every specialist works an 8-hour day.

What would change it

A revision of the published method. A baseline written in working hours needs no conversion, so the figure does not touch it.

No new published cases this month.

Speed is the easiest thing for an AI strategy firm to claim and the hardest for a buyer to check. We say strategy1 turns weeks of strategy work into days. Before we print a single number that supports it, here is how the number will be made.

The measure is the Compression Index: a specialist's baseline time divided by strategy1's time for the same deliverable.7Show source for note 7 An index of 5 would mean we took a fifth of the time the specialist estimated. That is the arithmetic, not a result.

• Written first. An independent specialist estimates the same scope in writing before we begin, and the estimate is stored with the case evidence. No baseline, no index.
• Like for like. Same scope, same acceptance criteria. If the specialist's estimate covered more than we delivered, we cut the baseline to match, or we do not publish an index.
• All of our time. Agent runs, strategist work, human review and Domain Fellow sign-off all count, taken from our time log and the timestamps on signed approvals.
• Waiting treated the same. Time spent waiting for the client's data or approvals is excluded on both sides, or on neither.
• Medians, with outliers shown. Each case is reported on its own first. Then we report the median per engine and per quarter, with the number of cases beside every median, and every case stays in, slow ones included.

A case is counted only when four things exist8Show source for note 8: a written baseline dated before work started, a deliverable the client or the specialist accepted, our time record for the same scope, and the client's written permission to publish, with an anonymisation pass reviewed by a Strategy Partner. Internal demos and estimates we wrote ourselves do not count.

An index measures time, not quality. Every published case will also show who validated the output, the red-team findings and what the work could not do. No case is published yet, and Report #1 of the India Strategy Compression Report is being prepared. When a case clears every step, Compression watch will report it, linked to its baseline document.