When a spreadsheet stops being the right tool
Most of them never should be replaced. The signals that say this one has to be, what the error research actually found, and the cheaper answers to try first.
By Sohail Chowdhary
A spreadsheet stops being the right tool when the process needs a rule the file cannot enforce, when more than a handful of people edit it at once, when nobody can say who changed a number and when, or when only one person could rebuild it. Nearly nine in ten production workbooks contain at least one error, but the typical formula is right about 98 to 99 times out of 100, so the risk is concentrated in a few bad files rather than spread evenly. Most teams should keep the spreadsheet and add a named owner and a backup. The ones that should not can usually try a database tool at £10 to £45 a user a month before commissioning anything, and a bespoke system starts at £10,000.
The short answer
A spreadsheet stops being the right tool when the process it runs needs a rule the file cannot enforce, and not before. Everything else, the size of the file, the number of tabs, how old it is, how embarrassing it looks, is noise. Most of the workbooks people ask us about should be kept.
This guide is written by a company that sells the expensive answer, so it is worth saying at the top which way we are pushing. We turn away work that a spreadsheet is already doing well, because a build that replaces something that worked is the kind a client resents a year later.
The four signals below are the ones we use ourselves, and most teams who read them find they have none of them.
There is a second thing worth separating. This is not the own-or-rent question, which starts from paying a vendor and asks whether to keep paying. This one starts somewhere different: you built the thing yourself, it worked, and it has quietly become the thing everything depends on.
When a spreadsheet is still the right tool
A spreadsheet is the right tool when one person owns the number, the shape of the problem is still changing, and being wrong is recoverable. That covers most of what most companies do in Excel, and it is why the tool has outlived every product built to replace it.
The honest position
If the file is a model rather than a record, if it is rebuilt each time rather than accumulated, if one person could recreate it in a morning, or if a mistake shows up before it matters, keep it. Give it an owner and a backup and stop reading. Nothing in the rest of this guide applies to you.The professional body agrees, and says so before it says anything about how to build one well. The Institute of Chartered Accountants in England and Wales makes it the fifth of its twenty principles: “Before starting, satisfy yourself that a spreadsheet is the appropriate tool for the job. Spreadsheets are not the answer to every problem.”1
The alternatives it names are a word processor, a database and an existing package, in that order. Bespoke software is not on the list, which is fair.
The signals, and what each one costs
Four signals matter, and each has a cheap fix that should be tried before anything is commissioned. The expensive answer is only right when the cheap fix has been tried and did not hold.
- The process needs a rule the file cannot enforceNothing over £4,000 goes out without a second approval. A job cannot be marked complete before the certificate is uploaded. A spreadsheet can hold a rule as a note or a conditional format, but it cannot refuse. Cheap fix first: a locked column and a named checker. If people route around it, the rule is real and the file is the wrong home for it.
- More than a handful of people edit it at onceMicrosoft’s own documentation for co-authoring is plain about what happens: “the last change that is saved ... is the one that wins”, and its suggested control is to give people separate areas or sheets3. A shared mailbox alongside it has a documented ceiling of 25 users, cannot encrypt what it sends, and cannot stop anyone deleting a message4. Cheap fix first: split the file by team, or move the queue into a tool with rows and owners.
- Nobody can say who changed a number, and whenThe question is not whether the file has version history. It is whether anyone can answer, in front of a customer or an auditor, who changed a figure and on what basis. Cheap fix first: turn on version history and stop emailing copies. If the answer still takes an afternoon, the process has outgrown the tool.
- One person could rebuild it, and nobody else couldThis is the one that usually decides it, and it is a headcount rather than a feature. The Enron corpus, the largest set of real corporate spreadsheets anyone has been able to study, shows how narrow the craft is in practice: about 75% of the spreadsheets used only the top fifteen functions, and only 134 of Excel’s 300-plus functions appeared at all9. The file is rarely complicated. The knowledge of why it is shaped that way is what leaves. Cheap fix first: have somebody else rebuild the critical part from the documentation. What they cannot rebuild is the risk.
Two things that are commonly given as signals and are not. The size of the file: a modern Excel worksheet holds 1,048,576 rows by 16,384 columns2, and almost nobody reaches that. And the age of the file: a workbook that has run a process correctly for six years is evidence for the workbook, not against it.
What the error research actually says
About nine in ten production spreadsheets contain at least one error, but the typical formula is right 98 to 99 times out of 100, and the risk sits in a small number of very bad files rather than spread across all of them. That is a more useful sentence than the one everybody quotes, and unlike that one it survives its own sources.
The figure in circulation is “90% of spreadsheets contain errors”. The ICAEW prints it in its own principles document, without a citation1. Where it comes from is a summary by Raymond Panko of six field studies, 85 spreadsheets in total, which found errors in 94% of them, alongside laboratory work giving an average cell error rate of 3.9%6.
Three researchers then reviewed the same literature and found it too thin to carry the weight put on it. Of seven sources, three were unpublished, and most gave little detail on how an error was defined or found. The study accounting for 70% of the cell-error sample had not reported an error rate at all, but the proportion of cells an auditor raised an issue about on first review7.
Their conclusion is worth quoting because it is not the one an agency would choose: “the evidence behind the estimates in the literature does not provide sufficient grounds for a reliable, quantitative estimate”7. They add, fairly, that they do not think the true rate is negligible.
They then did the study properly. Fifty operational workbooks, collected from consultancies, a bank, a college, an energy company and a government agency, with no source contributing more than five. They audited 270,722 formulas8.
| Measure | Usually quoted | The 50-workbook audit |
|---|---|---|
| Workbooks with an error of any kind | 94% | 94% |
| Workbooks with a wrong result | not reported | 86% |
| Cell error rate, inclusive | 5.2% | 1.79% |
| Cell error rate, wrong results only | not reported | 0.87% |
| Workbooks with an error rate under 2% | not reported | 46% |
The spread is the point. Most audited workbooks were close to clean; one had a cell error rate of 28%. Hard-coding a number into a formula was the single commonest fault, at 37.7% of instances.
The conclusion a sceptical reader should take is not that spreadsheets are dangerous. It is that errors are normal in anything people build, which is Panko’s own framing, and that a handful of files carry most of the exposure. Finding which of yours those are is cheaper than replacing all of them.
Four famous failures, corrected
Every page on this subject tells the same four stories, and three of them are told wrongly. We read the primary documents. Here is what they say.
| The case | Usually said | What the primary source says |
|---|---|---|
| Public Health England, 2020 | Excel’s row limit lost 16,000 covid cases. | PHE’s own statement reports 15,841 cases affected between 25 September and 2 October and never mentions Excel10. The limit involved belonged to the legacy .xls format at 65,536 rows, not the current format’s 1,048,576. |
| JPMorgan, 2012 | A spreadsheet error cost the bank $6.2bn. | The bank’s own task force reports that the model ran on spreadsheets completed by copying and pasting, that its approval required automation which never happened and which nobody checked, and that one formula divided by a sum instead of an average, which “likely had the effect of muting volatility by a factor of two ... although it is unclear by exactly what amount”12. |
| Reinhart and Rogoff, 2013 | An Excel error shaped austerity policy. | The researchers who found it attribute 0.3 of the 2.3 percentage point gap to the spreadsheet and transcription errors. The rest came from which years were excluded and how countries were weighted, which are judgement calls no software would have caught13. |
| Gene names, 2016 and 2021 | Excel renames genes. | It does, and the count grew rather than fell: errors in 19.6% of affected papers in 2016 and 30.9% by 20211415. The naming authority eventually renamed the genes rather than wait for the tool16. |
The pattern in the first three is the same, and it is not about Excel. Each was a process nobody owned, running a number nobody checked. That failure mode moves with you into any tool.
The gene-name case is the one worth keeping, because it is the cleanest evidence of a spreadsheet silently changing data that a technical, motivated audience could not stop it changing. Microsoft has since added a settings panel that turns off some conversions, though its own page on dates still says there is no way to turn that one off completely5.
What the tools are documented to do
Where a limit matters, it is worth quoting the vendor rather than the internet. These are from Microsoft’s own pages, read on 7 September 2026.
| What people ask | What Microsoft documents |
|---|---|
| How many rows? | 1,048,576 rows by 16,384 columns per worksheet, with 15-digit calculation precision2. |
| How many people can edit at once? | No maximum is published. Co-authoring requires the file in .xlsx, .xlsm or .xlsb on OneDrive or SharePoint Online, and “the last change that is saved ... is the one that wins”3. |
| How many people can share the inbox? | A shared mailbox “supports a maximum of 25 users”, stores 50 GB unlicensed, cannot encrypt what it sends and cannot stop users deleting messages4. |
The numbers that circulate for concurrent editors, usually ten recommended and ninety-nine maximum, are not on Microsoft’s documentation. The sentence that matters more is the one about the last save winning.
The cheaper answers, priced
Between a spreadsheet and a bespoke system there is a row of products that cost tens of pounds a month, and for a lot of processes one of them is the right answer. These are list prices from each vendor’s own page, read on 7 September 2026. Dollar prices are the vendor’s; £1 was about $1.35 at the European Central Bank’s reference rate27.
| Tool | List price | Where it stops |
|---|---|---|
| Airtable | $20 to $45 per user a month, billed annually21 | A hard record cap per base: 50,000 on Team, 125,000 on Business22. |
| Notion | $10 to $20 per member a month23 | Page history is 30 to 90 days, and database automations start at the Business tier. |
| Zapier | $19.99 to $69 a month, billed annually24 | The meter is per task, so cost tracks volume rather than headcount. It moves records between other people’s systems and holds no process of its own. |
| Microsoft Power Apps | £15.40 per user a month, paid yearly25 | The per-seat meter never stops, database capacity is charged by the gigabyte, and what you build does not leave the platform. |
| Retool | $10 to $50 per builder a month, plus per user26 | Two meters, and single sign-on and source control sit behind the unpriced enterprise tier. |
Smartsheet and monday.com are missing on purpose: neither page would render a price we could stand behind on the day, and a figure we cannot source does not go on this site. The three-year arithmetic for renting against owning is in the own-or-rent guide; this table is about where each tool stops, not what it totals.
Where the middle options break
The pattern is the same across all of them, and it is worth knowing before you migrate rather than after. Each of these tools is excellent up to a line it draws, and the line is rarely the one you were watching.
- A cap you meet suddenly. Airtable’s record limit is per base and hard22. A process that grows steadily hits it on an ordinary Tuesday.
- A meter that tracks the wrong thing. Zapier counts tasks, not people24. A busy month costs more than a quiet one, and nobody notices until the invoice.
- The rule you actually needed, one tier up. Governance features, single sign-on, audit trails, source control, sit in the tier with no published price on more than one of these products.
- Nothing to take with you. What you build in a low-code platform generally does not leave it. That is the trade, and it is a reasonable one as long as it is made deliberately.
None of that is an argument for building instead. It is an argument for knowing which line you will meet first, and how long it will take to get there.
Work out your own number
There is no credible independent measurement of what a spreadsheet-run process costs to keep running. The two figures in circulation were both commissioned by companies selling the replacement, so we are not going to quote them at you. Measure your own instead. It takes an afternoon.
- Count the copies. How many versions of the file exist across email, desktops and the shared drive? Anything above one is the process telling you something.
- Time one cycle. Have the person who runs it note the minutes, once, end to end, including the chasing. Multiply by how often it happens.
- Count the re-keying. Every place a number is typed into a second system is a place it can differ from the first.
- Count the people who could rebuild it. This is the number that usually decides. If it is one, the exposure is not the file.
The nearest thing to an independent measurement anyone has is the Enron corpus, where roughly a hundred emails a day involved a spreadsheet and 6% of those mentioned an error or a fault9. It measures how much a spreadsheet-run process circulates and gets corrected, which is not hours, but it is real and nobody sold it to you.
Who has to hold it, whichever way you go
The answer to a spreadsheet that has become a risk is a named person before it is a purchase. That holds whether you keep the file, move to a database tool or build something.
The ICAEW’s first principle is to decide what role spreadsheets play in the business and to manage the ones inside a key process differently from the ad hoc ones1.
The National Audit Office, auditing across government, reports that it regularly finds errors in departments’ models, including errors of £800 million and £45 million in two financial models in a single department’s 2020 to 2021 accounts, which had not been independently verified before the audit17.
The Treasury’s review after the West Coast franchise collapse recommended a single named senior responsible owner for every business-critical model18. None of those documents mentions Excel. All three are about ownership.
Two pieces of UK regulation are worth knowing about even though neither is likely to bind you directly. The Prudential Regulation Authority has written a definition of a model that a workbook computing a material number meets, and expects firms to keep a record of every one, though it applies to banks, building societies and PRA-designated investment firms rather than to a logistics company in Leeds19.
And from financial years beginning 1 January 2026, listed companies have to declare on the effectiveness of their material controls, with each board deciding what counts as material20. Neither is your obligation. Both may become your customer’s question.
One thing this guide will not claim: that building software removes the risk of a wrong number. It moves it. People write code too, and the error research that produced the spreadsheet figures found comparable base rates in programming6.
What a system can do that a file cannot is make a rule compulsory and keep a history nobody can quietly edit. If that is what your process needs, that is the reason, and it is the only one worth spending on.
If the signals are there
Bring the spreadsheet and the process it runs. The call is twenty minutes and free, and it ends in a recommendation, which is sometimes to keep the file and fix the ownership. If a build is right, it is scoped from a blueprint first and priced before you commit, from £10,000.
Questions people ask about spreadsheets
Sources
Every third-party figure above was read from the page cited, on the date shown. Vendors change prices without notice, so the date is part of the fact.
- 1ICAEW, Twenty principles for good spreadsheet practice, second edition. Principle 1 on managing key-process spreadsheets differently, Principle 5 on checking a spreadsheet is the right tool; the “approximately 90%” figure is printed without a citation. Seen 7 September 2026.
- 2Microsoft, Excel specifications and limits. 1,048,576 rows by 16,384 columns; 15-digit calculation precision. Seen 7 September 2026.
- 3Microsoft, Collaborate on Excel workbooks at the same time with co-authoring. No maximum number of co-authors is published; on a simultaneous edit the last save wins. Seen 7 September 2026.
- 4Microsoft Learn, About shared mailboxes in Microsoft 365. Maximum 25 users; 50 GB unlicensed; cannot encrypt outbound mail; deletions cannot be prevented. Seen 7 September 2026.
- 5Microsoft, Stop automatically changing numbers to dates. Automatic date conversion cannot be turned off completely; formatting as text or a leading apostrophe is the workaround. Seen 7 September 2026.
- 6Raymond R. Panko, What We Don’t Know About Spreadsheet Errors Today, EuSpRIG 2015. Errors in 94% of 85 spreadsheets across six field studies; average cell error rate 3.9% in laboratory work, against 1.9% to 3.7% base error rates found in code inspections. Seen 7 September 2026.
- 7Powell, Baker and Lawson, A critical review of the literature on spreadsheet errors, Decision Support Systems 46(1), 2008. 88 spreadsheets across seven sources, three unpublished; the evidence does not support a reliable quantitative estimate. Seen 7 September 2026.
- 8Powell, Baker and Lawson, Errors in Operational Spreadsheets, Journal of Organizational and End User Computing 21(3), 2009. 50 workbooks, 270,722 formulas audited; 94% with an error, 86% with a wrong result, cell error rate 1.79% inclusive and 0.87% for wrong results. Seen 7 September 2026.
- 9Hermans and Murphy-Hill, Enron’s Spreadsheets and Related Emails, ICSE 2015. 15,770 spreadsheets released; about 100 spreadsheet emails a day, 6% mentioning an error or fault; 75% of spreadsheets used only the top fifteen functions. Seen 7 September 2026.
- 10Public Health England, statement on delayed reporting of COVID-19 cases, 4 October 2020. 15,841 cases affected between 25 September and 2 October; the statement does not mention Excel. Seen 7 September 2026.
- 11Fetzer and Graeber, Measuring the scientific effectiveness of contact tracing, PNAS 118(33), 2021. Properly traced cases associated with 63% fewer subsequent infections and 66% fewer deaths over the following six weeks. Seen 7 September 2026.
- 12Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses, 16 January 2013. The formula divided by a sum instead of an average, “likely” muting volatility by a factor of two, amount unclear; the required automation never happened and nobody followed up. Seen 7 September 2026.
- 13Herndon, Ash and Pollin, A Critique of Reinhart and Rogoff, PERI Working Paper 322, 15 April 2013. The spreadsheet and transcription errors account for 0.4 of the gap; excluded years and country weighting account for almost 2 percentage points. Seen 7 September 2026.
- 14Ziemann, Eren and El-Osta, Gene name errors are widespread in the scientific literature, Genome Biology 17:177, 2016. Gene name errors in 19.6% of published articles with supplementary Excel gene lists. Seen 7 September 2026.
- 15Abeysooriya and others, Gene name errors: Lessons not learned, PLOS Computational Biology, 30 July 2021. Errors in 30.9% of articles with supplementary Excel gene lists, up from 19.6%. Seen 7 September 2026.
- 16HUGO Gene Nomenclature Committee, New Guidelines 2020. Symbols that affect data handling were changed; MARCH1 became MARCHF1. Seen 7 September 2026.
- 17National Audit Office, Financial modelling in government, HC 1015, 27 January 2022. Errors of £800 million and £45 million found in two models in one department’s 2020-21 accounts, which had not been independently verified. Seen 7 September 2026.
- 18HM Treasury, Review of quality assurance of Government analytical models, March 2013. Commissioned after the Intercity West Coast franchise competition; recommends a single senior responsible owner for each business-critical model. Seen 7 September 2026.
- 19Prudential Regulation Authority, SS1/23 Model risk management principles for banks, May 2023. Applies to UK banks, building societies and PRA-designated investment firms with internal model approval; the word spreadsheet does not appear. Seen 7 September 2026.
- 20Financial Reporting Council, UK Corporate Governance Code 2024. Provision 29 applies to financial years beginning on or after 1 January 2026; what counts as a material control is for each board to determine. Seen 7 September 2026.
- 21Airtable pricing. Team $20 and Business $45 per user a month, billed annually. Seen 7 September 2026.
- 22Airtable, plans and limits. 50,000 records per base on Team, 125,000 on Business. Seen 7 September 2026.
- 23Notion pricing. Plus $10 and Business $20 per member a month. Seen 7 September 2026.
- 24Zapier pricing. Professional $19.99 and Team $69 a month on annual billing; a task is counted whenever Zapier completes a unit of work. Seen 7 September 2026.
- 25Microsoft Power Apps pricing. Power Apps Premium £15.40 per user a month paid yearly; Microsoft notes prices are for marketing purposes and may not reflect actual list price. Seen 7 September 2026.
- 26Retool pricing. Team builder $10 and Business builder $50 a month on annual billing, plus per-user charges. Seen 7 September 2026.
- 27European Central Bank, euro foreign exchange reference rates, 3 September 2026. USD 1.1615 and GBP 0.86055 to the euro, which makes £1 about $1.35. Seen 4 September 2026.
