The Lean Tech Edition

Efficiency

The Software You Are Paying For Twice

Measured software prices have been flat for twenty-nine years. The bills have not. The gap between those two facts is where an owner-operator finds real money.

There is a particular kind of line item that survives every round of cost-cutting because nobody can quite remember what it does. It is nine dollars, or forty-nine, or two hundred and ninety a year, and it arrives on the card statement under a name invented by a marketing department in 2019. Cancelling it feels reckless. Keeping it feels like a small recurring defeat. Most owner-operated businesses carry several, and the honest reason they survive is not that anyone judged them worth the money. It is that nobody has sat down with the statement, a notepad and two uninterrupted hours.

Start with a fact that cuts against the general mood. Business software has not become dramatically more expensive. The Bureau of Labor Statistics publishes a producer price index for software publishers, which tracks what the industry actually charges for its output. Its base is December 1997, set at 100. In September 2026 it stood at 102.1. Across twenty-nine years, measured prices for the industry's output rose about two percent in nominal terms, while general prices roughly doubled. Over the most recent decade — September 2016 to September 2026 — the index rose 18.2 percent, against 39.4 percent for consumer prices. Priced against everything else a business buys, software got cheaper.

• 102.1 Producer price index, software publishers, September 2026, against a December 1997 base of 100 — source
• +18.2% The same index over the decade to September 2026 — source
• +39.4% Consumer price index, all items, over the same decade — source
• 77% Of small employer firms reporting rising costs, tariff costs or both as a financial challenge — source

That index deserves its caveats, because a flat line does not mean nobody raised a price. Producer price indexes measure prices received by domestic producers, weighted toward large contracts rather than the small-business till, and they are quality-adjusted — part of the flatness is that a dollar now buys considerably more software than it did, not that invoices held still. The series is also volatile and unseasonally adjusted, dipping below 93 in early 2024, so no single month should carry an argument. The index for data processing and hosting services tells a parallel story: up 27.4 percent since its December 2000 base, up 2.8 percent over the last twelve months. Nothing here resembles the doubling owners describe.

What is doing the work is count. Unit prices held roughly steady while units multiplied: more tools, more seats, more modules, more years of accumulated renewals nobody cancelled. Per-user pricing is the mechanism that turns a stable price into a rising bill, because roles change and provisioned seats do not. The subscription model supplies the second mechanism: a decision made once, in a hurry, in a month now three years gone, continues to be honoured by a card on file. Neither involves a price increase. Both produce a bill that grows while the index stays flat.

Rising costs of goods, services, and/or wages was the most common financial challenge reported in the prior 12 months.Federal Reserve Banks, 2026 Report on Employer Firms, Small Business Credit Survey

There are four distinct ways a small business pays twice, and they call for different remedies. The first is overlap: two tools whose feature sets have grown into each other, so the scheduling app and the customer system now both send appointment reminders. The second is seats still provisioned for people who left or roles that were consolidated. The third is migration residue — a month or a year in which the old plan and the new one were both live, and only the new one was ever cancelled from memory. The fourth is bundle blindness, the most expensive of the four, where a business buys a standalone product already included in a productivity suite it pays for monthly. Suites have spent a decade absorbing adjacent categories, so a business that adopted one in 2019 and bought point solutions in 2021 and 2023 is very often paying twice through no carelessness at all. The suite grew after the choices were made.

Now the famous number, because it will be quoted at you. Software-management vendors circulate a figure for the share of purchased licences that go unused, and the versions in circulation disagree wildly. Productiv put it at 53 percent in 2023, from what it described as hundreds of active instances of its own platform; that report is no longer on the publisher's own site and survives in web archives. Zylo published 51 percent in 2024 and currently says 46 percent — a company whose average customer, by its own account, spends $55.7 million a year on software. Nexthink said 49.96 percent, but of installed desktop software, measured deliberately during the first months of its own deployment to represent a lack of visibility. And the roughly 20-to-30 percent figure most often repeated comes from Flexera, which did not measure anything: it asked 503 IT asset managers at organisations of a thousand employees or more to estimate their own waste. Flexera has not published that number since 2024.

Every one of those figures is a vendor measuring its own customers — organisations that bought licence-management software because they already suspected they were wasting money, with software budgets in the tens of millions. No federal statistical agency measures software utilisation at all. The Bureau of Labor Statistics prices software; it does not count logins. The Census Bureau's capital expenditure data captures capitalised software — $135.0 billion for employer firms in 2022, of which sole proprietorships accounted for $3.2 billion — but a monthly subscription is an operating expense and appears in no federal table anywhere. That absence is the real story: there is no public number for what small businesses spend on software subscriptions, and therefore no credible average for how much of it is wasted. The only honest figure is the one your own export produces.

$135.0bn Capitalised computer software bought by US employer firms in 2022 — the last year federal statistics counted it. Subscriptions are operating expenses and appear in no federal table. — US Census Bureau, 2022 Annual Capital Expenditures Survey, Table 1a

That last instruction is the one that matters most, because an audit conducted in a cost-cutting mood will reliably destroy something load-bearing. The tool that looks idle may hold the appointment history, the tax records or the only copy of a customer list. Cancelling frequently means losing access to data that was never exported, which converts a small monthly saving into a problem with no price on it. Confirming that an export exists and opens is the single discipline separating an audit from an accident.

The deeper point is cadence rather than arithmetic. A software bill does not grow through bad decisions; it grows in the absence of a scheduled moment to revisit good ones. Prices are flat. Attention is the scarce input. A business that reviews its recurring charges once a year on a fixed date will not accumulate the residue that makes the annual figure alarming, and will have the useful side effect of knowing at any moment what it pays for and why. Put the date in the same calendar as the tax deadlines.

None of this is a technology project. It is bookkeeping applied to a category that has quietly escaped bookkeeping, and it works for the reason the price indexes reveal: the industry is not overcharging per unit. It is very good at selling another unit, and better still at never reminding you that it did. The money is not recovered by a negotiation. It is recovered by a list.

Efficiency

Automate the Thing Nobody Enjoys

The honest measurements of time saved by office technology are small, specific and conditional — which is exactly why the task you dread is the right place to start.

Every owner-operated business has a Sunday-evening task. It is the reconciliation, or the invoicing, or the rota, or the chasing of three customers who have not paid. It takes between forty minutes and two hours, it is not difficult, it produces nothing a customer would notice, and it sits at the end of the week like a toll. This is the task to automate, and the reason has little to do with efficiency in the abstract. The cost of the Sunday-evening task is not its duration. It is the way the whole weekend arranges itself around the knowledge that it is coming.

That routine office work can be mechanised is not speculative, and the evidence is unusually dry. The Bureau of Labor Statistics projects employment of bookkeeping, accounting and auditing clerks to fall six percent between 2025 and 2035, and says why in plain language: software innovations have automated many of their tasks, so the same amount of work can be done with fewer employees. Secretaries and administrative assistants, an occupation of 3.5 million people, are projected to decline two percent, which the Handbook attributes to digital tools letting staff prepare their own documents and to managers absorbing work that was once delegated. This is a federal statistical agency with nothing to sell describing the displacement of exactly the work that fills a small-business owner's evenings.

• $24.36 Median hourly wage, bookkeeping, accounting and auditing clerks, 2025 — source
• $46.89 Total employer cost per hour worked, private industry, June 2026 — source
• 1.4% Of total work hours: measured time saving reported by generative AI users — source
• 24.6% Of US businesses reported using AI in any business function, late 2026 — source

Software innovations have automated many of the tasks performed by bookkeeping, accounting, and auditing clerks.Bureau of Labor Statistics, Occupational Outlook Handbook

Having established the direction, the size of the prize needs deflating hard. The best-measured case is generative artificial intelligence, because economists asked a nationally representative sample rather than a vendor's customers. Alexander Bick, Adam Blandin and David Deming found that between one and five percent of all work hours were being assisted by the technology, with respondents reporting time savings equivalent to 1.4 percent of total work hours — about thirty-four minutes on a forty-hour week. Anders Humlum and Emilie Vestergaard, linking Danish survey responses to administrative records, could rule out effects on earnings and hours larger than two percent two years after the technology's launch. Workers reported gains; the payroll records did not show them.

The most important study for a reader of this magazine is the least encouraging. Nicholas Otis and colleagues ran a field experiment giving small-business owners access to an AI business mentor, and published it in Management Science. There was no detectable average effect on profits. Underneath the average, the effect split: roughly fifteen percent better for owners who were already performing well, roughly ten percent worse for those who were not. That is the single randomised trial on actual small-business owners, and its lesson is not that the tools do not work. It is that they amplify whatever process they are pointed at, which is why the state of the process matters more than the choice of tool.

Now the famous numbers, which are quoted at owners constantly and are worth tracing properly. The claim that workers lose two and a half hours a day searching for information descends from a single 2012 McKinsey Global Institute report, which said that the average interaction worker spends an estimated 28 percent of the week managing email and nearly 20 percent looking for internal information. Its own appendix explains the basis: estimates from a proprietary analyst report — Susan Feldman's Hidden cost of information work, published by International Data Corporation in May 2009 — combined with McKinsey's internal data and, in its own words, reports from social technology tool vendors. The underlying time-use data is therefore at least seventeen years old, was never published in a form anyone can inspect, and the figure most often repeated is a misreading: 8.8 hours a week of searching is 1.8 hours a day, while the 2.5-hour figure is the email line, not the searching line.

The companion claim — that automation saves the average small business some impressive number of hours a week — traces most often to a 2017 Smartsheet survey in which nearly sixty percent of respondents estimated they could save six or more hours a week if the repetitive parts of their jobs were automated. Three things about it: the saving was hypothetical rather than observed, the survey was commissioned by the vendor through an online panel, and sixty percent of the thousand respondents worked at organisations of a thousand employees or more. It is not a small-business study. Nor is the robotic-process-automation return-on-investment literature peer-reviewed: the widely circulated figures trace to a university outsourcing unit's reports and a teaching case, not to research.

34 minutes What 1.4 percent of a forty-hour week comes to — the measured average time saving reported by generative AI users — Bick, Blandin and Deming, The Rapid Adoption of Generative AI, published in Management Science

There is one properly measured administrative saving in genuinely small firms, and it is instructive about shape. Oyebola Okunogbe and Victor Pouliquen, studying a staged rollout of electronic tax filing to small and medium firms, found in the American Economic Journal that e-filing reduced the time firms spent on taxes by forty percent — on the order of five hours a month. Note what that was: one recurring, highly structured, externally mandated process, replaced end to end. Not a general productivity uplift. The savings that survive measurement are the ones attached to a specific repetitive process that was fully converted, which is precisely the opposite of buying a general-purpose tool and hoping.

What are other businesses actually doing with this? The Census Bureau's Business Trends and Outlook Survey, which reaches 200,000 businesses every two weeks, reported 24.6 percent of businesses using AI in any business function in late 2026, with 28.6 percent expecting to within six months. Read that number carefully, because Census changed the question wording twice — from use in producing goods or services to use in any business function — and warns that the series before and after are not comparable. Firms with one to four employees reported 24.3 percent, those with 250 or more 43.6 percent. Among firms using generative AI specifically, the tasks named most often were writing and editing documents and emails at 85.4 percent, searching for information at 49.9 percent, and information processing, paperwork or filing at 34.7 percent.

Choosing the candidate matters more than choosing the tool, and three shapes of work automate well because their output is checkable at a glance. Recurring data movement is the first — the same figures carried between systems on a schedule, where an error shows up as a mismatch. Scheduling and reminders are the second, because the failure mode reaches you before it reaches a customer. Document assembly is the third: quotes, invoices and standard letters built from a template and a short list of fields. Work whose errors are silent should be automated last, and only with a check in front of it. Census found that among firms using AI, 43.7 percent used it to supplement or enhance an employee's task while only 10.1 percent used it to perform a task previously done by a person — and 95.7 percent reported no change in total employment.

The characteristic failure is not technical. It is that the owner automates what is easiest to automate rather than what costs the most to keep, and ends up with a tidier version of work that was never the problem. The second failure is subtler: automation moves effort from doing to checking, and checking is harder to schedule and easier to skip. A process that runs unattended for nine months and then quietly stops is worse than one never automated, because the business has by then forgotten how to do it by hand. Both are avoided by the same discipline — pick the task you dread, and keep a human date with its output.

Thirty-four minutes is an honest number, and the literature offers nothing larger that survives contact with a business of four people. Aimed at a Sunday evening, on a process that has first been written down and then converted completely, it buys back something that does not price. Aimed at everything at once, on the strength of a vendor's survey, it buys a second job.

Efficiency

What Going Paperless Really Saved

An honest ledger: where the saving was real and enormous, where it was smaller than advertised, and where the numbers that sold the idea turned out to be invented.

The room that used to hold the filing cabinets is the most persuasive argument for going paperless and the least reliable one. The real ledger is stranger than either the brochures or the backlash suggest: office paper genuinely collapsed, the money saved came from a direction nobody predicted, the environmental case is weaker than the posters implied, and the statistics that sold the project appear to have been fabricated.

Begin with the saving that is genuine and under-appreciated: paper itself got expensive. The Bureau of Labor Statistics producer price index for paper stood at 187.0 in September 2016 and 282.5 in September 2026 — a rise of 51.1 percent in a decade, against 39.4 percent for consumer prices, and up 10.1 percent in the last twelve months alone. Paper outpaced inflation, which means the sheet you did not buy saved a little more each year than the sheet before it. This is the exact inverse of what happened to software prices over the same decade, and it is the most defensible financial case for the switch: not that digital is cheap, but that paper stopped being.

• +51.1% Producer price index for paper, September 2016 to September 2026 — source
• −73% US production of printing and writing papers, 2000 to 2024, UN FAO data — source
• 1.2% Share of US commercial building electricity used by office equipment — printers, copiers, fax machines — source
• $60.6m Spent by the IRS on an in-house scanning system that scanned no returns before being halted — source

282.5 The producer price index for paper in September 2026, against 187.0 ten years earlier — Bureau of Labor Statistics, Producer Price Index by Commodity: Paper (WPU0913)

The volume collapse is not in dispute. United Nations Food and Agriculture Organization figures show US production of printing and writing papers falling from 23.0 million tonnes in 2000 to 6.3 million in 2024, a decline of 73 percent; newsprint fell 95 percent. The EPA's waste tables tell the same story from the other end: office-type papers, meaning copier and printer paper, peaked at 7.42 million tons generated in 2000 and were down to 3.97 million by 2018. American offices stopped consuming paper on a scale visible in national statistics. The paperless office was not a myth. It just took thirty years rather than five.

Postage is the second real saving, and for most small businesses the largest line recovered. First-Class Mail volume peaked in 2001 at 103.7 billion pieces and stood at 42.0 billion in the Postal Service's audited 2025 accounts, a fall of about fifty-nine percent. Total mail peaked in 2006 at 213.1 billion pieces and was 108.7 billion in 2025. Meanwhile a first-class stamp went from 34 cents in January 2001 to 82 cents in July 2026, a rise of 141 percent. Fewer letters, each costing far more: a compounding saving, and arithmetic rather than assertion. One footnote cuts against the usual story — the stamp price actually fell in April 2016, from 49 cents to 47.

Now the famous numbers, and this part is worth reading twice. Document-management marketing has for thirty years rested on a cluster of figures: that it costs $120 to find a misfiled document and $220 to reproduce a lost one; that 7.5 percent of all documents are lost for good; that a typical worker spends up to two hours a day looking for them. All are attributed to a 1998 study by the accountancy firm Coopers and Lybrand. In October 2010 the president of AIIM — the document-management industry's own trade association, a man who by his own admission had used the figures a million times in presentations — went looking for that study, could not find it, publicly called it an urban legend, and offered a year's free membership as a bounty to anyone who could produce the original.

I searched first my hard drive. No dice. Then the web. No dice. Many references to the 1998 Coopers and Lybrand report, no actual copy or link. Oops.John Mancini, then president of AIIM, October 2010, offering a bounty for the study behind the document-management industry's most-quoted statistics

The bounty was never claimed. Two years later a document-scanning company that had cited the figures in its own marketing asked whether the study had ever surfaced; the reply was two words and six full stops: still looking. That firm stopped using the statistics. Most of the industry did not, and the figures have since been laundered upward, with at least two peer-reviewed papers now citing, as their source for the 1998 study, the blog post explaining that it cannot be found. A further claim — that the average office worker uses 10,000 sheets a year — is routinely credited to the Environmental Protection Agency. The page most often blamed prints it with no attribution at all, having cited the EPA two paragraphs earlier for something different. The EPA counts tons in municipal solid waste, never sheets per worker.

Where going paperless saved less than claimed is the environment, and a correction is owed to a figure widely quoted approvingly. The EPA reports a 68.2 percent recycling rate for paper and paperboard, the highest of any material it tracks — but it is dominated by corrugated cardboard, recycled at 96.5 percent. It is not the office-paper rate. The EPA has not published one for office-type papers since 2005, when it was 62.1 percent; every year since shows a dash, the agency saying it lacks valid values to separate the sub-category out. Worse, the series stops at 2018 while page footers carry recent update dates that disguise it.

The cost that moved rather than vanished is storage, and the honest version is more modest than the headlines. Lawrence Berkeley National Laboratory, in its 2025 update for the Department of Energy, put US data centre consumption at 192 terawatt-hours in 2024 — 4.7 percent of national electricity — with a reference case of 11.8 percent by 2030. That is a real resource cost, moved from your filing room to a building in another state and re-billed monthly. But set against it what paperless eliminates. On federal buildings data, office equipment — printers, copiers, fax machines — is 4 of the 227 billion kilowatt-hours an American office building uses, and 14 of 1,196 billion across all commercial buildings. Just over one percent. The trade is real, and a knockout in neither direction.

The rebound effect needs the same treatment, because the received wisdom is out of date. Abigail Sellen and Richard Harper argued at book length in 2001 that paper-displacing technology paradoxically increased paper use: email arrived, and people printed it to read. They were right, and the paradox held for roughly a decade. Then it broke. Michael Briscoe, revisiting the question with FAO data for a peer-reviewed journal in 2022, found US paper consumption peaking around 1999 at some 95 million tonnes and falling to 63 million by 2020 — the lowest since 1983. His explanation is the smartphone: once a document is legible in your hand, the reason to print it goes away. One global caveat keeps the ledger honest, though. While American offices cut printing paper by three-quarters, world production of paper of all kinds rose thirty percent, because packaging grew 84 percent and tissue 93. The office gave paper back and the parcel took it.

The clearest warning about conversion costs comes from the largest paperless project in the country. The Treasury Inspector General reported in February 2026 that the Internal Revenue Service had spent nearly $61 million developing an in-house scanning system halted before testing, having scanned no returns at all. Paper returns, by the agency's own estimate, cost forty-three times more to process than electronic ones and account for 72 percent of individual processing costs while being six percent of filings, so the prize is real. But processing time for paper employment-tax forms went from 25 days in 2024 to 45 in 2025 to 72 in 2026, and auditors found the agency counting 751,000 scanned paper returns inside its published electronic-filing figures. Converting a backlog is harder and slower than any business case assumes.

The last entry is behavioural. Paper imposed a discipline through its bulk: a cabinet filled, and somebody had to decide what to discard. Digital storage removes the prompt, so retention becomes infinite and the subscription grows at the rate of accumulation rather than need. The businesses that genuinely saved money are, almost without exception, those that also set a deletion schedule. Without one the cabinet was not eliminated; it was rented. Otherwise the ledger reads: the office-paper saving real and far larger than the sceptics allow, the postage saving real and compounding, the environmental saving smaller than the posters claimed, the energy cost moved rather than removed and worth about one percent either way, the rebound effect real and then reversed, and the four figures that sold the whole idea untraceable to any document anyone has ever seen.

Efficiency

Cheaper Is Not Always Less

The price on the plan page is one term in a longer sum. The others are your hours at loaded cost, the price of getting your data back out, and the chance the vendor does not outlast you.

The nine-dollar plan is a genuine bargain roughly half the time, and the trouble is that the half is not randomly distributed. It is reliably a bargain for commodity work with a clear exit and reliably a disaster for anything holding the records the business runs on. The price on the plan page is a real number and the smallest term in the sum that matters. Three other terms are always present, never advertised, and entirely estimable in advance by an owner willing to spend twenty minutes on arithmetic before spending a year on consequences.

The first missing term is your own time, and it is usually larger than the whole price difference. The Bureau of Labor Statistics reports total employer compensation for private industry workers at $46.89 per hour worked in June 2026 — wages of $32.82 and benefits of $14.07, so benefits are thirty percent of the total and an hour of labour costs about 1.43 times its wage. A cheaper tool that takes six extra hours to configure and three to learn has consumed something on the order of $420 of loaded time before it processes a single transaction. If the saving against the alternative is fifteen dollars a month, the cheaper option is behind for the first two and a half years.

• 50.2% Of establishments opened in the year to March 2015, the share still operating five years later — source
• 34.6% The share of that same cohort still operating after ten years — source
• $30.24 Median hourly wage, computer support specialists, 2025 — source
• $64.44 Median hourly wage, software developers, quality assurance analysts and testers, 2025 — source

That second pair is the term people forget entirely: the price of help when something cheap breaks. Cheaper tools come with thinner support, which transfers the work of diagnosis to you or to somebody you pay by the hour. The 2025 median wage for computer support specialists was $30.24 an hour; for software developers, quality assurance analysts and testers it was $64.44. Those are medians for employees, so what a small business pays a contractor for a short engagement will be a multiple of them. A single afternoon of outside help can exceed a year of the price difference, and needing it is not a remote possibility. It is the ordinary consequence of choosing a product whose business model excludes hand-holding.

The second missing term is vendor survival, and here a small business should extend itself the courtesy of remembering that its supplier is also a small business. The Bureau of Labor Statistics tracks establishments from birth. Of those that opened in the year to March 2015, 79.6 percent were still operating after one year, 50.2 percent after five, and 34.6 percent after ten. Those are rates for all private establishments rather than software companies specifically, and they are the right order of magnitude for the question that matters: will the cheap tool holding your customer list still exist when you need it in 2032? About a third of the businesses founded alongside it will. Price is a signal here, because a vendor charging too little to be sustainable is telling you something about its own survival odds.

$16.6bn Losses reported to the FBI Internet Crime Complaint Center in 2024, across 859,532 complaints — a 33 percent rise on 2023 — FBI Internet Crime Complaint Center, 2024 IC3 Annual Report

The third missing term is the exit, and this one economists have studied properly. Switching costs and lock-in are a developed field: Joseph Farrell and Paul Klemperer's survey in the Handbook of Industrial Organization sets out how costs incurred after a customer commits — retraining, data conversion, lost history, contractual friction — shift bargaining power to the incumbent and allow prices to rise later without losing the customer. The mechanism requires no bad faith. The cost of leaving simply grows with every month of accumulated records, and a vendor that priced aggressively to acquire you has every reason to recover the difference once leaving has become expensive. A low entry price and a high exit price are not in tension. They are the same strategy.

This number does not include estimates of lost business, time, wages, files, or equipment, or any third-party remediation services acquired by an entity.FBI Internet Crime Complaint Center, 2024 IC3 Annual Report, on reported ransomware losses

The fourth term is tail risk, and it needs care because this is the part of the market most thickly populated with invented numbers. The claim that sixty percent of small businesses close within six months of a cyberattack is repeated constantly and has no study behind it that anyone has produced. The FBI's Internet Crime Complaint Center, which publishes the only federal tally of reported cyber-crime losses, reports complaints and dollars: 859,532 complaints and $16.6 billion in 2024, with business email compromise alone accounting for $2.77 billion. It does not report business closures, because it does not collect them — and the establishment survival data above shows about half of all businesses closing within five years for every reason combined, so a six-month closure rate attributed to a single cause could not be separated from the baseline even if someone had measured it.

Two neighbouring figures deserve the same treatment. The widely quoted average cost of a data breach comes from an annual study published by a technology vendor, using its own cost-attribution method on a sample that excludes the smallest incidents — which is the category a small business is most likely to suffer. Cost-of-downtime figures, quoted per hour or per minute, originate with vendors selling resilience, generally without a published methodology or an inspectable sample. The mechanism, though, is real and measurable in your own business: take the hours you cannot trade without the system, multiply by average revenue per trading hour, add the loaded cost of staff idle or working around the failure, and add what you would spend on emergency help. That number is specific to you, defensible, and usually sobering.

None of which argues for expensive software. Cheap is straightforwardly correct wherever the work is commodity, the data small or easily re-created, and leaving costs an afternoon: file storage, video calls, scheduling, note-taking, password management, simple websites. In those categories the market is competitive, the formats interchangeable, and a premium buys very little. There is decent evidence that the cheap end genuinely helps small firms in particular — Timothy DeStefano, Richard Kneller and Jonathan Timmis, in the Review of Economics and Statistics, found that cloud computing raised revenue, employment and productivity among younger and smaller firms while showing no such effect for incumbents. An owner who moves all of that to the cheapest credible option and reviews it annually will save real money with almost no exposure, and should.

Cheap is wrong in one identifiable place: the system of record. Whatever holds the customer list, the transaction history, the payroll or the regulated documents is not a tool but the memory of the business, and the four-term sum there is dominated by exit cost and tail risk rather than price. That is the one place to buy the boring, established, slightly overpriced option, to pay annually without resentment, and to run an export every quarter whether or not anything seems wrong. Everywhere else, be ruthless.

The discipline is neither frugality nor caution. It is insisting that a saving be stated as a complete sum before it is called a saving — price, plus your hours at what they actually cost, plus the price of leaving, plus what happens on the worst day. Most of the time that sum still favours the cheaper option, which is exactly why it is worth doing. It tells you which half you are in.