Cost Sanity-Checks & Estimates

Start with the hard limits, because they define everything else. Claude cannot price your project. It has no live feed of material or labour rates, no knowledge of your local market, contractor margins or this month's steel and cement prices, and it cannot measure quantities off a drawing - recall from Module 1.4 that it cannot reliably read dimensions off an image at all. Worse, when it lacks a figure it does not say "I don't know"; it generates a plausible one. A rate of "around Rs 2,500 per square foot for a mid-range residence" may be roughly right for one city and badly wrong for another, and Claude presents both with identical confidence. Treat any specific number it volunteers as fiction until you have confirmed it at source.

With that firmly established, two legitimate uses remain, and they are useful. The first is the order-of-magnitude sanity-check: not "what does this cost?" but "is this number the right size?" When a client says the neighbour built an identical house for a figure that sounds impossibly low, or a quote lands that feels off, Claude can reason about whether the number is even plausible for the described scope - catching a missing zero, a confused unit, or a budget that cannot possibly buy the brief. The second is setting up the estimate: Claude builds the skeleton - the line items, the elements, the formula that turns areas and rates into a total - which is real, tedious structuring work, and then you pour in verified rates. In both cases Claude handles the shape of the calculation and you own every actual figure. That boundary is the whole lesson.

It helps to understand why the invented numbers are so convincing. Claude has read enormous quantities of cost data, guidance and discussion, so the rates it produces sit in a genuinely plausible range - they are not random. That is exactly what makes them dangerous: a wildly wrong figure you would catch, but a figure that is plausibly-but-materially wrong slides straight into a document and a client's expectations. The generic average it recalls may be right for a mass-market build in one city and forty per cent out for a bespoke house in another, and nothing in the phrasing tells you which. So the rule is not "distrust Claude when it sounds unsure" - it almost never sounds unsure - but "never let a specific figure it volunteers stand as fact." Treat confidence as decoration, and verification as the work.

The sanity-check is the safest and often the most valuable way to bring Claude near money. The question is never "how much?" but "does this smell right?" - and for that, order-of-magnitude reasoning is exactly what a language model can do usefully, because you are asking about the size of a number, not its precision.

Suppose a developer client insists a 200-square-metre house can be built for a total that works out absurdly low per square metre. You can hand Claude the scope and the figure and ask: "For a house of this description, is this total in a plausible range, generous, or impossible? Reason about the drivers." Claude will walk through what such a figure implies per unit area and flag that it looks well below any normal band for the finish level described - not a quote, but a red flag worth raising before the brief is built on a fantasy. It is superb at catching the gross error: the decimal point in the wrong place, the quote that omitted MEP entirely, the FF&E budget that cannot furnish the number of rooms.

The figure shows how to think about the bands: broad, illustrative ranges that separate a basic fit-out from a bespoke one. These bands are for reasoning about magnitude, not for quoting - real rates swing with city, date, spec and market, so any band Claude offers must be replaced with figures you have verified locally. The value is in the ratio and the smell test: if a proposed number sits three bands away from where the scope suggests it should, something is wrong, and you have found it early. Ask Claude to state the assumptions behind any band it gives you, so you can see exactly what would have to be true for the number to hold.

The second use turns Claude into a structuring engine for a rough estimate - the kind you need at concept stage, long before a QS is engaged, to test whether the brief and the budget can coexist. Claude is genuinely good at building the skeleton: prompt it to lay out an order-of-magnitude estimate as a line-item structure and it will produce something clean.

Build me an order-of-magnitude cost estimate STRUCTURE for
a 3-storey office fit-out, as a table. List the elements
(demolition, partitions, ceilings, flooring, MEP, joinery,
FF&E, prelims, contingency, fees). Leave all rates blank
for me to fill with verified local figures. Once I enter
quantities and rates, compute each line and the total, and
SHOW YOUR WORKING so I can check it.

The hand-off is the same as in the last lesson: Claude supplies the structure and the arithmetic; you supply the rates, verified from actual quotes, a real rate library, or a QS. Build it as an Artifact so it becomes a working sheet. But now comes a discipline particular to numbers: Claude can get the arithmetic wrong even when the rates are right. Language models are not calculators; they can slip on a multiplication or misplace a subtotal, and they do it as confidently as everything else. So there are two independent failure modes to guard against - a wrong rate (your responsibility to verify) and wrong arithmetic (Claude's slip to catch). Always make it show its working, spot-check lines by hand, and reconcile totals two ways where you can - a unit rate on floor area cross-checked against the sum of the elements. When the two methods disagree, you have found an error before it reached the client. (Module 8.3 shows how a real spreadsheet or a small tool removes the arithmetic risk entirely.)

A quick before-and-after shows why the cross-check earns its keep. Ask carelessly - "roughly cost this 900 sq m office fit-out" - and Claude returns a single confident total, say a bit over three crore, with an invented blended rate baked invisibly inside it; you have no idea which line is soft. Ask deliberately - the structured skeleton above, with your verified rates entered element by element - and the same job resolves into checkable parts, so that when the total lands you can see it is the joinery and MEP driving it, and you can interrogate exactly those. The second version takes ten more minutes and is worth a hundred times the first, because it is inspectable. That inspectability is the entire difference between a number you can defend and a number you merely hope is right.

How a rough estimate is presented is as professional a matter as how it is calculated. A number that leaves your desk without its caveats will be remembered as a promise, so a Claude-assisted estimate must never be dressed up as a quote. Label it for exactly what it is: order-of-magnitude, at concept stage, plus or minus a stated range, with the assumptions and exclusions written down beside it. Claude is, usefully, excellent at drafting precisely that scaffolding - the assumptions register, the exclusions list, the caveat paragraph - which protects both you and the client and is too often skipped when a bare number is rushed out.

Build in what a rough estimate always needs: a contingency appropriate to how early and how uncertain the design is, and an allowance for the time between estimate and construction. Ask Claude to remind you what an estimate at this stage should include and exclude, and to flag the big-ticket items most likely to move - it is a good checklist partner even where it cannot supply the figures.

And know the boundary where Claude stops. For anything that carries real consequence - a figure a client will commit to, a tender sum, a loan application, a fee based on cost - you need a quantity surveyor or a proper cost plan, not a language model's arithmetic. Claude sets up the thinking and catches the howlers; the QS carries the number that carries weight, and the accountability that goes with it. Used inside those limits - sanity-checking magnitude, structuring the maths, drafting the caveats - Claude genuinely speeds up the early cost conversation. Pushed past them, it invents money you will have to answer for. Record your assumptions either way; a rough number with its reasoning attached is defensible, a bare number is a hostage.