How to Review a Financial Workbook in Seven Steps

A workbook is reviewed in seven steps. The reviewer reads the cover and the output, lists what the answer rests on, builds one total by a second route, runs the formula check along that list, traces each number back to its document, runs one case whose answer is already held, then writes findings by kind. Each step narrows the next, so the cheapest checks run first.

One uncomfortable calculation shapes every step below, so the arithmetic comes first. The file took nine hours to build. The reviewer has forty minutes. Nobody verifies a nine hour file in forty minutes and nobody ever has, so the honest question is not how to check everything. The question is which forty minutes to spend. A reviewA separate reading of a file somebody else built, looking for what the person who built it can no longer see. is an allocation decision wearing the costume of a checking exercise, and the seven steps below make that allocation.

Order is doing more work here than any single check. The early activities change what the later ones look at, so the same seven in a different sequence produce a different review. Reading the output shows which numbers left the file. Listing what those numbers rest on shows which cells to open. Done before anything else, the two of them leave the remaining thirty two minutes landing on the part of the file that carries the conclusion. Done last, they leave the hour spent on whatever the reviewer happened to meet first.

A mechanic looking over a second hand scooter does not strip the engine. The mechanic starts it, listens, looks at the tyres, checks the chassis number against the papers, and only then opens the one thing that sounded wrong. Three minutes of listening decides where the next forty go. A workbook review has that same shape: a short pass that costs almost nothing, and which earns its place by showing where to spend the expensive part.

One assignment runs through all seven steps below. Inside Kavery Capital Services Private Limited, the Kavery research desk was asked a question about Meenakshi Tubes Private Limited. Sharada Iyer built the file over nine hours. Prakash Nadar reviewed it in forty minutes. Latha Menon decided on the strength of what came out.

The seven steps never depend on what Meenakshi Tubes Private Limited actually makes. Substitute a bond, a fund, a warehouse or a single machine and not one of the seven steps changes. The steps check a container. The finance inside the container is a separate subject.

What Are the Seven Steps, and Why Is the Order Doing Most of the Work?

The whole sequence sits in one list, and each step is then walked through in turn. The reviewer reads the cover and the output, lists what the answer rests on, builds one total a second way, runs the formula check along that list, traces each number back to the document it came from, runs one case whose answer is already held, and writes the findings by kind. Seven steps, forty minutes, in that order and no other.

The rule setting the order is that a cheap step which narrows the next one always runs before an expensive step which does not. Step one costs three minutes and establishes what the file claims. Step two costs five and identifies which cells carry the claim. Between them they cost eight minutes out of forty, and everything after them is aimed rather than swept. The two most expensive steps, at eight and nine minutes, run fourth and fifth precisely because by then they know where to point.

Notice what is not in the list. There is no step for reading the file end to end, no step for opening every formula, and no step for tidying. Reading end to end, opening every formula and tidying are all real activities, and not one of them survives a forty minute budget. A procedure that would take four hours is not a procedure. Four hours is a wish.

Step One: What Is Opened First in a File Never Seen Before?

The cover first, then the output, and nothing else. Three minutes. Nothing is being checked yet, and the urge to start checking is worth resisting. The first three minutes buy a statement of what the file claims, and the rest of the review has something to check that claim against.

Nothing is verified in step one, and that is why it goes first rather than despite it. A step that costs three minutes and answers what am I looking at earns its place by making every later step cheaper. On this assignment the cover said who built it, who reviewed the last version, the date, and one line saying the file was built to answer one question and should not be reused for another. The output carried two figures that had left the file and appeared in a document Latha Menon read: an operating margin of 10.8 per cent for the current year against 12.0 per cent for the prior year, and a scenario margin of 5.0 per cent.

The numbers that left the file go down on paper before a single formula is opened. The list of numbers that left the file is what is actually under review. Everything else in the file is machinery serving those figures, and machinery that no published number depends on is not what the forty minutes are for.

Step one is done when the reviewer can say out loud, without looking, what the file concluded and which numbers carry the conclusion. If not, the output gets read again rather than left behind. Two minutes more here is cheaper than twenty minutes of aimless checking later.

Step Two: What Does the Answer Actually Rest On?

Five minutes, a sheet of paper, and no formulas open. Each number written down in step one raises one question: what would it take to move it. Every input that would change the published figure if it were wrong goes on the list. The written list of inputs is the dependency listThe short written list of what the published answer actually rests on, made before any formula is opened., and it decides where the next twenty one minutes go.

Prakash Nadar wrote four lines. Source log item 3, the current year filed accounts, supplying revenue of Rs 21,20,00,000 and operating profit of Rs 2,30,00,000. Item 6, the prior year filed accounts, supplying Rs 18,40,00,000 and Rs 2,21,00,000. The quarterly break-up inside item 3. Item 5, the paperwork behind the December order of Rs 2,96,80,000. Four lines out of a source log of fourteen items, or 28.6 per cent of the log and rather less of the file.

Four things carried the whole answer out of a file holding 214 formulas, and until they are written down every one of those formulas looks equally worth opening. With the list in hand, the formulas that lie on the path from those four numbers to the published figures come to 26. Twenty six is 12.1 per cent of 214. The cut from 214 formulas to 26 is what the entire forty minute budget rests on.

The household version is a loan application. The bank is not checking the applicant's life. The bank is checking income, existing loan repayments, the value of what is pledged, and the record of paying on time. Four things. Everything else across the form's twelve printed sides exists to support one of those four, and an officer who spends the afternoon on the postal address has reviewed nothing.

Step two is done when the list fits on one side of a sheet and every line on it can be defended. A dependency list of twenty items is not a dependency list, it is the file rewritten, and it aims nothing.

Step Three: Can One Total Be Built a Second Way and Land on the Same Number?

Six minutes. Pick the single largest number the conclusion rests on and build it again from something the file did not use. Not the same arithmetic re-typed. A second routeAn independent way of arriving at the same total, using inputs the first route did not touch. means different inputs arriving at the same total. Agreement then means something.

On this assignment the largest number was revenue for the current year. The file reported Rs 21,20,00,000, taken straight from the filed accounts. The published figure never touched the quarterly break-up, and the second route came from there: Rs 4,80,00,000 for the first quarter, Rs 5,20,00,000 for the second, Rs 6,40,00,000 for the third and Rs 4,80,00,000 for the fourth. The four quarters add to Rs 21,20,00,000. The two totals agreed to the rupee.

Six minutes buys one route, so only one total gets this treatment, and pretending otherwise is how the budget breaks. Choose the number that the largest share of the conclusion hangs from. Here the choice was revenue. Both published margins divide by revenue, so a wrong revenue would move both figures Latha Menon read.

Step three is done when the two totals either agree or the gap has been written down in rupees. A gap that is not written down as a number, right then, becomes a vague memory of something looking odd by the time step seven arrives.

Step Four: Which Formulas Actually Get Opened?

Eight minutes, and the answer is the 26 on the dependency list rather than the 214 in the file. The formula check is the first expensive step, and it is fourth in the order for exactly that reason. Run before step two it is a sweep. Run after, it is a walk along a path somebody has already drawn.

There are two motions here and they cost very different amounts. The mechanical one is instant: the file is asked to show every formula with a digit typed inside it. The file returned 11 cells out of 214, or 5.1 per cent. The slow motion is reading, and reading is what gets rationed. Prakash Nadar read the 26 formulas that lie between the four listed numbers and the two published margins. Of the 11 cells with a number typed inside them, 2 sat on that path and were read carefully; the other 9 went straight onto the findings list without a second look.

Eleven cells, two of them on the path to the answer and nine of them not, is the whole reason the dependency list runs before the formula check rather than after. Both sets end up on the findings sheet. Only one set is worth eight minutes of reading, and without the list from step two there is no way to tell which.

Expect roughly one in twenty. Somebody arriving at this step expecting zero will treat 11 as an accusation, an argument will start, and the check quietly stops being run on the next file. Somebody expecting sixty will conclude the whole file is untrustworthy when it is fine. Eleven in 214 is an ordinary careful file on an ordinary week.

Step four is done when every formula on the path has been read once and the cells with numbers typed inside them are listed. Not when the file is clean. A clean file is step seven, and somebody else's work.

Step Five: Does Each Number Still Point at the Document It Came From?

Nine minutes, the longest step in the review, and the only one that spends its time outside the file. Take each item on the dependency list, open the actual document behind it, and follow the published figure back to see which item it consumed. Four items, roughly two minutes each, with the balance spent on the one that did not behave.

Three of the four traced clean. The fourth did not. The current year operating margin should consume item 3 twice: once for the operating profit on top and once for the revenue underneath. The cell consumed item 3 once and item 6, the prior year filed accounts, once. Rs 2,30,00,000 was being divided by Rs 18,40,00,000 rather than by Rs 21,20,00,000, and that produces exactly 12.5 per cent.

Look at what 12.5 per cent does. The prior year margin was 12.0 per cent. A current year figure of 12.5 reads as a small improvement, and improvement was the opposite of what the finished work concluded. The true figure of 10.8 per cent sits below the internal floor of 11 per cent that the whole assignment was measured against. One cell, one wrong item number, and the conclusion flips from a fall through the floor to a modest rise.

Nothing inside the file was capable of finding this mismatch, and the step that looks outside is worth nine of the forty minutes for that alone. The cell was a well formed division. The division pointed at a real cell holding a correctly filed number. Every label in the neighbourhood was honest. Revenue itself was right, so the reconciliation in step three passed. Revenue was simply not the number this cell reached for. Only walking out to the documents and asking which one this figure consumed turns up a mismatch.

Step five is done when every item on the dependency list has been opened as a document and matched to the figures that consumed it. Not when the numbers look familiar. Familiar is what a wrong number looks like by the fourth reading.

Step Six: Does the File Return an Answer Already Held?

Four minutes. One input whose correct output is already known gets set, and what comes back is examined. The setting is a known caseAn input setting whose correct answer is already held, so the file either returns it or it does not., and it is the cheapest test on the list because it checks not the arithmetic but whether the machine as a whole behaves.

The file carries one input for the share of revenue treated as non-repeating, used to build the scenario. Set that input to zero and the scenario collapses into the year as reported, so the file must return the filed margin of 10.8 per cent. It did. Set it to 14 per cent, the share the December order came to, and the file returned a scenario revenue of Rs 18,23,20,000, an operating profit of Rs 91,96,000 and a margin of 5.0 per cent, matching the figure on the output tab.

A file that fails the known case is wrong everywhere rather than in one place, and a file that passes it says nothing about whether the right numbers went in. Passing means the structure holds together. Passing says nothing about whether the inputs are the ones the documents support. Step five asked that question, and step six sits after it rather than before for exactly that reason.

Step six is done when the case returns the figure held, or when what it returned instead has been written down. Adjusting the input until it works is not a test, it is fitting.

Step Seven: How Do the Findings Get Written Down?

Five minutes, and this is where a good review most often throws away its value. At this point the reviewer holds a pile of cell references, and the pile has to become a small number of findings. One findingOne thing that is wrong, written so that the person who built the file can act on it without asking a question. is one thing wrong, written so the builder can fix it without coming back to ask what was meant.

Write them by kindGrouping the findings by the class of problem each belongs to, rather than by where in the file each instance sits. rather than by location. Fourteen cells were wrong on this file. Written by location that is fourteen lines and fourteen separate decisions for Sharada Iyer. The fourteen cells belong to three classes, so written by kind they come to three lines: 11 with a number typed inside the formula, 2 reading the prior year where the label says current year, and 1 input cell carrying no source log item number. Eleven plus two plus one is fourteen, with nothing left over.

A builder fixing a class thinks once and a builder fixing cells thinks fourteen times, so three instructions beat fourteen, and the fourteenth edit is where the new mistake goes in. Each line names the kind, lists the cells it covers, and says what to do. Saying what to do is not politeness. A finding a builder cannot act on is a complaint.

How Long Should a Review Take, and What Buys That Time?

Forty minutes against nine hours of building. Work it both ways. The two denominators give different answers, and people quote whichever suits them. Nine hours is 540 minutes, so forty minutes of review is 7.4 per cent of the time spent building. Add the review to the build and the assignment took 580 minutes end to end, of which the review was 6.9 per cent. Call it about seven per cent either way, and hold that as the ordinary shape rather than a target.

Seven per cent sounds too small to be worth anything, and it is only enough because the seven steps are in the order they are in. An unordered forty minutes buys a quarter of the file and no conclusion. An ordered forty minutes buys the cover, the output, a dependency list, one reconciled total, twenty six read formulas, four traced documents, one known case and a findings sheet. Same clock. The order is the whole difference.

The number is worth holding because of what happens when work runs late. Review is the last thing in the sequence and the easiest thing to cut, and the file most likely to have something in it is the one built in a rush the night before. A desk that treats review as roughly seven per cent of the build has a number to defend when somebody asks for the file an hour early. A desk that treats it as whatever time is left has nothing to defend and loses the argument every time.

When Does the Review Stop?

The review stops when every line on the dependency list has been reached by at least one step that looked outside the file, and the known case returned what it should. On this assignment that condition was met at the thirty fifth minute, and the last five went on writing the findings.

The file will never be exhausted, so the review cannot wait for that, and waiting is how a review turns into a rebuild. A great deal of the file was never opened. Roughly 188 of the 214 formulas were never read by anybody other than Sharada Iyer, and that is the intended result rather than a confession. The 188 unread formulas do not carry the two figures Latha Menon read.

There are two honest ways to end early, and both get written down. If the known case does not return the figure held, the review stops and says so. Everything downstream of a structural fault is unreliable, and reading more of it wastes the remaining minutes. If a line on the dependency list cannot be traced to a document at all, the review stops on that line and records it as untraceable rather than nodding it through. Neither of those is a failed review. A review that ends with a written statement of what it could not reach is worth more than one that ends with a tick.

Who Actually Runs Seven Steps Like These, and What Do They Get for It?

A credit officer at a lender is the clearest case. A borrower sends a projection workbook supporting a request for a line. The officer has perhaps half an hour and no interest in the borrower's spreadsheet habits. So they read the cover and the requested amount, list what the repayment capacity rests on, rebuild one revenue figure from something the borrower did not use such as filed returns or bank statements, and set the growth input to zero to see whether the file still services the loan. The officer has run steps one, two, three and six, in order, on somebody else's file, in half an hour.

An analyst receiving a model from a colleague runs the same shape for a different reason. The question is not whether the model is tidy but whether the output can be put in front of somebody who will act on it. The dependency list is what lets the analyst say, in a meeting, that the conclusion rests on four things and that three of them were traced to documents that morning. A defensible answer is not one where everything was checked. A defensible answer is one where what was checked is written down and what was not is also written down.

A household does a smaller version and rarely calls it a review. The bank officer produces a repayment schedule on a laptop and turns the screen around. The useful three minutes are not spent on the schedule. The three minutes are spent asking which interest rate went in, over how many years, and whether the insurance premium is inside the instalment or on top. The questions are a dependency list, built out loud, on a file somebody else made. Four questions, and they beat reading every row of a two hundred row table.

In every one of those cases the reviewer is short of time, holds no authorship of the file, and gets their result from the order of the questions rather than from the volume of checking. That is what the seven steps are for, and it is why they survive being moved from a research desk to a loan counter without changing shape.

Kavery Capital Services Private Limited, the Kavery research desk, Meenakshi Tubes Private Limited, Sharada Iyer, Prakash Nadar and Latha Menon are invented.
Educational material. Not advice on any investment, tax, budget or market position.