Two kinds of aim
A sniper computes; a fire controller converges. Two modes of decision-making, learned coordinating indirect fire in the Swedish archipelago and used since in credit committees and boardrooms — and an explanation for why the people best trained for precision are often the slowest to move when the problem is not a precision problem.
A sniper's shot is a computed thing. The target is known, the world is briefly stable, and the inputs — range, wind, drop — feed a firing solution that must be exactly right before the trigger moves. An error is a failure.
Indirect fire is the opposite problem. The artillery battery cannot see the target; the space is large, the picture moves, and the perfect solution is not available at any price. So the answer is not a solution — it is a procedure. You fire a ranging round you know might miss, observe the splash, and bracket: over, short, halve the difference, fire for effect. Directionally right and fast iteration beats exactly right and late, because exactly right and late does not exist. The discipline isn't precision. It's correction.
I sat on the artillery side of this — battalion-level fire coordination across amphibious, army, naval and air assets. The lasting lesson wasn't about weapons; it was that neither mode is superior. Competence is knowing which problem you are in. Most institutional failure I have seen since — in credit committees, in boardrooms, in AI programs — is related to mode confusion.
The paralysis of the precisely trained
The question that gives it away is "where do we start?" — and in my experience it is almost always asked by an organisation's best people. Credit officers, engineers, CEOs, controllers: people whose careers were built on being right, in professions where a single wrong number can erase a hundred good ones. Their trained instinct is the sniper's: compute the firing solution first. The strategy document, the vendor matrix, the governance framework, the complete picture — then act.
It is the correct instinct, applied to the wrong class of problem. Entering an unfamiliar market, restructuring a function, adopting a technology whose limits nobody has established yet — none of these hold still long enough to be computed. AI adoption is only the current and most visible example, which is why it is the one people ask about.
The artillery answer is to take a ranging round: one target that actually matters, a bounded budget, a sprint or two. The point of the first round is not to hit. It is to see the splash — what the data actually supports, where the process is weaker than the org chart admits, what people reach for once they have the thing in hand, which parts of the problem turn out to need exactness after all. Very little of that is knowable from the planning room. Most of it is visible from the first correction.
The reversible sprint
The real job in decisions of this kind — I have written this elsewhere and will keep writing it — is distinguishing reversible from irreversible. A "wasted" sprint is a reversible decision (a "two-way-door" decision, to quote Jeff Bezos): the cost is bounded, known in advance, and small — at least not large enough to break the bank. Two weeks and a limited budget, fully written off, is tuition. What is not reversible is delay. Every day you don't start is a day your data does not compound — the corpus you didn't capture, the operating method you didn't begin writing, the corrections you didn't make. The compounding asset was never the tool; it is the accumulated corrections and intelligence on data you DID gather and structure. And you cannot bracket without a first splash.
The banker's aside: price the two honestly. The expected loss of a ranging round is capped at the sprint's budget. The expected loss of a year spent asking "where do we start" is unbounded, and it never appears on any P&L — which is precisely why precision-trained organisations keep paying it.
Where the sniper still rules
None of this abolishes the sniper. Some decisions are genuinely sniper decisions — a covenant test, a mass balance, a tax computation, a dosing threshold. Exactly right or wrong, no bracketing. But notice what those have in common: they are computed, deterministically, from stated inputs. And there the machine's role inverts. AI helps you build the instrument that computes the answer — it should never be the instrument. A language model guessing at a sniper problem is the worst of both modes: artillery dispersion wearing sniper clothes. If the decision needs to be exact, demand a deterministic tool with provenance on every number, and let the AI write — and research — the tool.
So the question to ask of any number that crosses your desk: is this a sniper's number or a fire controller's number? If it's a sniper's, insist on the computation and its sources. If it's a fire controller's, insist the bracket is stated honestly — and correct fast. The failure mode is the silent mix: the five-year projection quoted to two decimals, or the hand-waving where exact arithmetic was available all along.
"Where do we start?" is itself a question asked in the wrong mode. You do not compute your way to a first shot. Pick a target that matters, a budget you would happily "waste", a fortnight, and fire. Read the splash. Correct. The organisations getting this right are not better marksmen — they are better at corrections. And they start to understand the tools now at their disposal.
The harder version of the same discipline is a system that deliberately carries both modes at once — financial statements that articulate to the cent, sitting beside scenario brackets that state their uncertainty and refuse to pretend otherwise. Most consequential work looks like that, and most of the arguments about it are really arguments about which mode a particular number belongs in.
— Oscar · Stockholm, July 2026