Consulting Mental Math Practice: The Calculator Rule

One firm writes down that the case interview bans the calculator

The page is Bain's digital assessment page at bain.com. It states that pen, paper and a calculator are all allowed for the assessment, and that "calculators are allowed only for the digital assessment", with the same sentence going on to rule them out during the interviews that follow.

Three other published details from the same page matter for planning.

The assessment runs about 30 to 40 minutes. There are two possible ones, and the firm tells you which after you apply: "you will receive an email after submitting the application detailing whether you are required to complete Sova or Test Gorilla." Their rules differ. For Test Gorilla, "Complete the test in one sitting". For Sova, "You do not need to complete the Sova assessment in one sitting, but please ensure all sections are completed ahead of the deadline". Assuming the second rule while sitting the first is how an application gets lost.

And AI is out entirely for the assessment: "The use of AI tools are strictly prohibited; this includes recording, screenshots, or use of any software that transcribes or provides feedback during the assessment."

A second firm publishes a related boundary for a different exercise. BCG's Swiss page at careers.bcg.com describes a 30-minute cognitive test where "The majority of questions will be in multiple-choice format and must be answered without any tools or resources." That is Switzerland and a full-time consulting application, and the page says the process varies by role.

So one firm bans the calculator in the interview, and another bans tools in a screening test. Nobody else in our corpus writes either rule down. A page claiming a universal ban is generalizing from one source, and we name the source instead.

What each half of the process is therefore testing

Read the Bain boundary as a design decision and the two halves separate cleanly.

When execution is delegated to a machine, the only thing left to assess is judgment: which figure to divide by which, what base to use, whether the answer is plausible. When the machine is removed, execution comes back, and it now has to happen inside a conversation.

That second condition is the one candidates underestimate. The constraint in a case interview is rarely raw computation. It is computing while narrating, with someone watching, having just been asked a question.

Simon-Kucher is the one firm in our corpus that turns this into a preparation instruction, at simon-kucher.com. Among four published tips is simulating interview conditions, including time awareness. The same page lists solving quantitatively among five case performance factors, beside structuring information logically and interpreting and communicating results.

Oliver Wyman adds the physical half, in the criteria printed under a question of its Dairy Farm case at oliverwyman.com: it wants the data laid out so it can be computed quickly. Speed there is a function of how you arranged the page, before any arithmetic happened.

One mental math error in forty, and the arithmetic behind it

Here is the calculation that sets the target, and it is the one nobody publishes.

Suppose you are 95 percent accurate on any single operation. That sounds strong. A case calculation is a chain of four operations, and the chain is only right if every link is.

• Four operations at 95 percent: 0.95 squared is 0.9025, and 0.9025 squared is 0.8145. So the chain is right 81.5 percent of the time, and wrong 18.6 percent of the time.
• Four operations at 99 percent: 0.99 squared is 0.9801, and 0.9801 squared is 0.9606. So the chain is right 96.1 percent of the time, and wrong 3.9 percent.

Recomputed by hand on 30 September 2026. The jump from 95 to 99 percent per operation cuts the chance of a wrong chain from about one in five to about one in twenty-five.

Now convert that into a practice target. On a set of 40 operations:

• 6 wrong is 6 over 40, which is 15 percent.
• 2 wrong is 2 over 40, which is exactly 5 percent.
• 1 wrong is 1 over 40, which is 2.5 percent.

So the commonly repeated target of 5 percent is one error in twenty, and against a four-step chain it leaves an 18.6 percent failure rate. The useful target is one error in forty.

That number is ours. No firm publishes an accuracy threshold, and we are not going to attribute our training target to anyone else. What the firms do publish is the consequence: Accenture's workbook states that success depends in part on "your quantitative analysis skills", with no threshold attached.

Three minutes a calculation, against a published case interview clock

Accuracy has a twin constraint, and one firm publishes the clock to measure it against.

Accenture's case workbook at accenture.com states: "Case interviews typically last from 45 to 60 minutes, with a focus on the case itself lasting 30 to 45 minutes." The document carries a 2018 copyright and reports 435,000 people against 791,000 in the firm's September 2026 postings, so it is dated and we say so.

Take the lower bound of 30 minutes for the case portion and price a slow calculation.

• Six minutes on one calculation: 6 over 30 is 20 percent of the case portion.
• Against the 45-minute upper bound: 6 over 45 is 13.3 percent, rounded from 13.33.
• Three minutes each on three calculations: 9 over 30 is 30 percent of the case portion.

Recomputed by hand. Three minutes per calculation is the budget that works, because three calculations at three minutes leaves 70 percent of the case for structure, exhibits and the close.

The cost of exceeding that budget lands somewhere other than the arithmetic score. Those six minutes came out of the branch you never opened, and in our own structuring rubric Coverage carries 45 points. A slow calculation converts a math problem into a structure problem, which is why the time target belongs on a math page at all.

Four-week consulting mental math practice, 400 minutes

Four weeks, five sessions a week, twenty minutes a session. That is 400 minutes of work, or 6 hours 40 minutes, which is the smallest honest budget for moving an error rate.

Read the week-three target against the arithmetic above. Six of eight chains clean is 75 percent, which corresponds to roughly 93 percent per-operation accuracy, since 0.93 to the fourth is 0.748. That is the realistic midpoint, and week four is where the narration cost shows up.

Four rules make the protocol work, and each one is doing a specific job.

1. Score errors rather than time in weeks one and two. Speed without accuracy is a worse state than slow and right, because a wrong chain produces a confident wrong recommendation.
2. Count a chain as wrong if any link is wrong. This is the measure that matches a real case, where nobody grades your third operation in isolation.
3. Write every intermediate figure. Most errors under time are transcription errors, and an unwritten intermediate cannot be rechecked.
4. Say the unit and the base before each operation. Per household, per year, of revenue, of profit. This is the error class that survives correct arithmetic.

Narrating while computing, in three sentences

Week four exists because arithmetic accuracy measured in silence does not predict arithmetic accuracy measured out loud.

What narration has to contain is short. Three sentences per calculation.

Before. I am going to take fixed cost over unit contribution to get the breakeven volume. This tells the interviewer where you are going, and it lets them redirect you before you spend three minutes.

During. Forty million over ten cents, so four hundred million units. One line, said as you write it.

After. Four hundred million units against a market of three billion, so about 13 percent share, which is a lot to win in year one. This is the interpretation, and it is the part silence eats first.

The third sentence is the one the protocol is built to protect. Under time, a candidate who computed correctly and said only the number has done the work and left the point on the table.

Two published instructions bracket this behavior. Bain's case tips include thinking aloud even when unsure and pausing to gather your thoughts, which combine into announcing a pause and then taking it. And a named Kearney analyst publishes the closing instruction: never end on a number when working the quantitative part of a case.

Where the protocol meets our weights

These four lines are ours, taken from the grading code in src/lib/grading. No consulting firm publishes a marking scheme, in this corpus or anywhere we have looked.