1Salesforce is an Oracle database with a few thousand buttons stuck to it
You closed the deal on Tuesday. It is Friday afternoon, the forecast call is on Monday, and as far as your company is concerned the deal has not happened, because you have not told the machine. So you open the opportunity, and the machine asks for a stage, a close date, a probability, a forecast category, a next step, a competitor and a reason, with thirty more fields below the fold that somebody added in 2019. You knew all of it on Tuesday, and so did the customer. The one party to the transaction that did not know was the database, and the database is the one that sends your company a bill.
Salesforce is an Oracle database with a few thousand buttons stuck to it. I wrote that sentence in September 2023 at the top of a short post, and the first half of it was a statement of fact. In 2013 Salesforce signed a nine-year agreement to keep running its products on Oracle’s database, and according to reports it would not confirm, it spent the years that followed on an internal project to leave, code-named Sayonara.1 The best-known application company of its generation was the captive of its database vendor and named the escape plan accordingly, which tells you what a software company thinks of captivity when it is the one inside.
Let me take it one step further. Every enterprise application is a database with a badly designed interface on top of it, and the design is bad for a reason that has nothing to do with the talent of the designers. The reason is how the product is bought. A committee writes down every feature any department has ever asked for, puts the list in a spreadsheet, sends it to five vendors and scores the replies, and the vendor with the most rows ticked wins. Each row becomes a button. The people on the committee will never press the buttons, and the people who will press them were not on the committee. A few thousand buttons is what twenty years of winning that contest looks like: the fossil record of procurement, one layer to a sales cycle, with no one paid to take a layer away.
What the vendor owns in this arrangement is smaller than it looks. The rows in the database belong to the customer, at least on paper, and a row is a row in anybody’s database. The vendor owns the path to the rows: the screens, the sequence of clicks, and the trained hands of ten thousand employees who know where the buttons are. Certifications, careers and consultancies have grown up around that knowledge, and none of them has an interest in there being fewer buttons. The interface is the brand, the brand is the switching cost, and the switching cost is the business.
2No exit, no voice, and the industry calls it retention
A company with forty employees sells packaging to a multinational, and one morning a letter arrives: from next quarter, invoices will be accepted through the buyer’s procurement network and no other way. The company registers, learns the portal, retypes its invoices into somebody else’s form, and then receives a bill from the network for the privilege of being paid. It did not choose the software, it cannot leave the software without leaving its biggest customer, and there is no one at the network whose job depends on whether it is happy. About ten years before that post I had written that Ariba does not have customers, it has prisoners, and became briefly infamous in procurement circles for it.2 The line was about those supplier fees, and it was as true of the employee on the buyer’s side of the portal, who had not chosen the software either.
A product improves through two channels. People leave it and the maker notices the revenue going, or people complain to someone who has to listen and the maker fixes the complaint to keep them. Enterprise software has closed both channels to the people who use it. The user cannot leave, because the choice was never hers, and she cannot be heard, because the vendor’s customer is the committee and the committee does not use the product. The buyer, who could leave and would be heard, feels none of what the user feels. Albert Hirschman named the two channels exit and voice in 1970, in a book about what happens to the quality of a railway, a school or a state when one or both of them is blocked.3
Hirschman also explains the spreadsheets. His case was a Nigerian railway that stayed bad while lorries took its freight, and his explanation was that the customers who cared most about service were the first to switch to the lorries, so the railway lost its sharpest complaints along with their business. Every company has that leak. The people who care most about the work are the first to tunnel out of the official system into a spreadsheet and an email thread, the vendor never hears from them again, and the system is left to the users least likely to object.
The third word in Hirschman’s title is loyalty, and the industry has a measure of it, net revenue retention, the share of last year’s revenue that came back this year with the price rises added. Vendors report it to investors as proof of love, when what it measures is the height of the wall, and a number above a hundred per cent means the wall went up. Enterprise software sucks because it can: the people who suffer it cannot leave, and the people who could leave do not suffer it.
3Ten thousand Clippys, and a protocol so they can talk
In the month I wrote the post, Salesforce announced Einstein Copilot, Microsoft set the date on which Microsoft 365 Copilot would go on sale, and SAP launched an assistant called Joule.4 By the end of the year there was hardly an enterprise vendor without one, and each of them lived inside its maker’s application.
The placement is the tell. To do most office work, a model that reads and writes language needs access to the records and permission to act on them, and neither of those is a screen. An assistant that lived outside the applications would make the screens optional, and the screen is what the vendor owns, so each vendor built the assistant as a feature of the screen and gave the user a reason to stay where the buttons are. Take the screen away, and an employee who asks for last quarter’s pipeline has no more idea whose database answered than you have of whose cable carried your last phone call. A vendor the user never sees is a utility, and the brand of a utility is the bill.
The short-term result is a paperclip in every application, the world of ten thousand Clippys. Microsoft shipped the original in Office 97, an animated paperclip that saw you type “Dear” and offered to help with your letter; it was switched off by default in 2001 and removed in 2007.5 Clippy knew what you were typing and had no idea what you were trying to do, and its descendants inherit the defect for a structural reason: each of them knows one application, and work does not happen in one application. I put it as a joke in 2023. I could not wait, I wrote, to explain to my Workday copilot that it had to talk to the Microsoft Office copilot to update a commission plan in Salesforce based on a spreadsheet my boss had just sent, and what could possibly go wrong.
Nine months later SAP and Microsoft announced what they called a bi-directional integration between Joule and Copilot, so that one assistant could pass a question to the other. In April 2025 Google published a protocol that lets agents from different vendors find each other and hand each other tasks, with Salesforce, SAP, ServiceNow and Workday among its launch partners, and by that summer it belonged to the Linux Foundation, with Microsoft on the founding list.6 Every company in my joke has signed up to a standard for assistants talking to assistants. I wrote it as a joke, and the industry wrote it as a specification.
The specification does what it says and leaves the defect where it was. A commission plan is a promise a company makes to its salespeople. Part of it is recorded in the HR system, part of it in the sales system, and the live version is in a spreadsheet the boss sent, which is to say in the tunnel. Three assistants that each know a third of the promise can exchange messages about it, which is a meeting, and a meeting between parties who share no understanding of the promise is how commission plans went wrong before there was any software at all.7 The paperclip lives in the application and the work lives between the applications, and no protocol moves the one to where the other is.
4What is defined by its interface ends up in the drawer
Most homes have the drawer. In it are a satnav with a suction cup, a compact camera, an MP3 player, a dictaphone, a pocket radio and a knot of chargers that fit none of them. Each was a complete product, with a company behind it, a shelf in the shop and a magazine that reviewed it, and each was defined by its physical interface. The satnav was a map database with a screen and a suction cup stuck to it, the camera was a sensor with a shutter button, and the player was a hard disk with a wheel.
The function survived in every case, which is the part the makers missed. People take more photographs than they ever have and buy almost no cameras: the industry shipped a hundred and twenty-one million of them in 2010 and under eight million in 2023.8 People use maps more than they ever have, and on the day in October 2009 that Google announced free turn-by-turn navigation on a phone, Garmin lost a sixth of its value and TomTom a fifth. What died was the object wrapped around the function. A product defined by its interface dies when a more general interface arrives, and the data behind it goes up in value, because a map on every phone is worth more than a map in a device on a windscreen. The newspaper and the paperback were interfaces too, made of paper, and neither came through unaltered.
Customer relationship management, human capital management, enterprise resource planning and the rest are the satnav, the camera and the player of office work. Each is a function, keeping track of customers, paying people, paying suppliers, wrapped in an interface and sold as an object, with a brand, a conference and certified experts of its own. The general interface has arrived, and it is language. Companies will go on paying suppliers and forecasting sales, so the functions will survive and the records will be worth more than they were, and the objects are on their way to the drawer.
Each side of the argument has a warning in the camera numbers. The warning for me is that the camera makers had the best year in their history in 2010, three years after the iPhone went on sale, so a record year says little about which side of the peak a company is on, and less about how far away the peak is. The warning for the vendors is in who survived: the makers still in business went where the interface is the craft, to the professional camera whose dials are an extension of the photographer’s hands. Some enterprise screens will survive for that reason, the trader’s terminal, the designer’s canvas, the radiologist’s viewer, anywhere the screen is the instrument of a skill. The form is no such place, because nobody’s craft is in the dropdown.
5The seat was the shadow of the screen
Until 2024 Salesforce sold its applications the way the industry always had, by the user per month. In the autumn of that year the company put a second unit beside the first, two dollars per conversation with its new agent, and in May 2025 a third, ten cents per action. Later in 2025 it was selling the agent by the user per month again, and by the start of 2026 it was reporting a fourth unit of its own invention, the agentic work unit, to its investors.9
The seat, a named person in front of a screen, has been the unit of sale in enterprise software for as long as licences have been counted, and it made sense while the screen was the product, because a screen needs a person in front of it. The seat is the shadow of the screen. When work is done by language over records there is no one in the chair to count, and a vendor that tries four units in eighteen months is a vendor that no longer knows what it sells. The return to the seat was the customers’ doing as much as the vendor’s, since a finance director can budget for heads and cannot budget for conversations, and that preference will slow the change more than any vendor can.
In 2023 I wrote that GPUs, models and even distribution would not be a moat in this market. The first two are simple, because a moat has to be something the customer cannot rent, and compute and models are both rented by the token from several suppliers at prices that keep falling. Distribution is the interesting one, since distribution was the moat of the application era: the field sales force, the integrator’s bench and the annual conference, all of it built to sell seats, one application at a time, to committees.
That machine depends on the customer agreeing to see itself as the vendors’ categories cut it, a sales department that needs a sales application, an HR department that needs an HR application and a finance department that needs three. A company is people organised around outcomes before it is a set of departments, and the commission plan, the supplier relationship and the customer’s problem each cross all of those departments on an ordinary day.10 Once the application stops being the unit, what is sold is the outcome, and distribution passes to whoever stands closest to the work, which is seldom the vendor with the biggest booth. Start-ups come in through that gap, and so do the consultants. A moat dug around the seat protects an empty chair.
6The box did not make the ship faster
On 26 April 1956 a crane at Port Newark lifted fifty-eight aluminium truck bodies onto the deck of a converted tanker, the Ideal X, which sailed for Houston with them. The man who had arranged it, Malcom McLean, was a road haulier from North Carolina who had bought a shipping line the year before, and an official of the longshoremen’s union, asked what he thought of the ship, said he would like to sink it. He had understood what he was looking at. Loading loose cargo onto a ship cost five dollars and eighty-three cents a ton that year, and loading the Ideal X cost under sixteen cents.11
The ship was no faster than any other tanker, and the cargo was what it had always been. What the box changed was the seam. Before it, every handover, lorry to shed, shed to quay, quay to hold, meant gangs of men lifting each crate, sack and barrel by hand for days; the Ideal X was loaded in under eight hours, and its cargo was packed once at the factory and not touched again until it arrived. The cost of a large system collects at its seams, the places where one party hands work to another, and whoever makes a seam cheap changes the economics of everything on either side of it. Economists call it transaction cost, and Ronald Coase argued in 1937 that it is the reason firms exist and the force that decides where their edges are.12
An enterprise is full of seams, and the labour at them is break-bulk. Between the customer’s email and the order system, between the supplier’s invoice and the buyer’s ledger, between the boss’s spreadsheet and the commission plan, somebody reads a document in one shape and retypes it in another. For forty years the industry’s answer has been to standardise the cargo: electronic data interchange in the 1970s, XML in the 1990s, e-invoicing networks and then e-invoicing laws, all of them asking the sender to produce the receiver’s format at the source. I spent fifteen years at Tradeshift on that answer, in a business whose unit of time is the supplier onboarding campaign, measured in quarters. The analysts who count invoices put the electronic share of business-to-business invoicing at twenty-nine per cent in 2026, and name governments, which have started to require it by law, as the strongest driver.13
McLean did the opposite, and that is why it worked: he left the cargo alone and standardised the box. A model that reads is the first box computing has had. It takes the invoice as a PDF, the order as an email and the commission plan as a spreadsheet, and carries what they mean across the seam without asking the sender to change anything. That is what I meant in 2023 by calling the large language model the container of computing. Software people hear the word and think of Kubernetes, which is Greek for helmsman and is named after the man who steers a ship full of the boxes I mean.
The cheap seam was the small effect, and the large one was that the work moved. London’s enclosed docks, once the busiest port in the world, closed one after another between 1967 and 1981, because the trade had gone downriver to Tilbury and up the coast to Felixstowe, where there was room for cranes and lorries, and because factories could sit an ocean away from their customers. One study puts the growth in trade between the industrialised countries that adopted the container at about seven hundred per cent over twenty years, more than the trade agreements of the period delivered.14 Enterprise applications are the old docks. They stand where they stand because handling was expensive and had to be concentrated, in the ERP team, the shared-service centre and the integrator’s bench, and when the seam is cheap the location stops mattering. Digitising a supply chain took years because ten thousand suppliers had to be talked into changing what they sent, and when none of them has to change I expect it to take weeks.
The box took ten years and a fight over standard sizes before it carried much of anything, and the old docks looked healthy for most of that decade.
A sealed box is also a box that goes unopened, so it moves mistakes and contraband at the price of cargo, and a model that misreads an invoice carries the misreading across the seam at full speed. Where exactness is owed, in the amount, the tax and the authority to pay, structure stays, which is what the e-invoicing laws are for.
And the dockers lost their trade. On the American West Coast the men doing the handling negotiated a share of what the box saved, in 1960, and kept their incomes while the work changed under them, and that is the chapter of the story the software industry has not read.15
A model does for meaning what the box did for cargo: it makes no application faster, it makes the seams between them cheap, and the seams are where the cost was.
7Tomorrow will not be like yesterday, and the walls are still standing
Three years is long enough to mark the post. The copilots came, one to an application; the paperclips needed to talk to each other and were given a protocol; and the unit of sale came loose. Fifteen months after I called the enterprise application a database with a bad interface, the chief executive of Microsoft described business applications as databases with business logic on top and said the logic was leaving for the agents.16
I also wrote “no more prisoners”, and the prisoners are still inside. Salesforce closed its 2026 financial year with revenue of forty-one and a half billion dollars, the most in its history, and another seventy-two billion under contract and not yet delivered, a figure the accounts call remaining performance obligation and a customer might call the sentence still to serve.17 Its shares fell anyway, along with the rest of the sector’s. In the six months to April 2026 the main fund tracking software shares lost more than a quarter of its value while the companies in it reported good earnings, because a share price is a forecast and revenue is a history, and the two had stopped agreeing. I read the revenue with the camera makers’ best year in mind, and I would not read the share prices as proof of anything.
What I underestimated is that the lock can move. In 2023 I treated the screen as the lock, and in mid-2025 Salesforce changed the terms of Slack’s programming interface so that other companies’ software could no longer index or store a customer’s messages. Salesforce described the change as data protection, and one of the companies shut out told its customers it would hamper their ability to use their data with the AI platform of their choice.18 By August 2026 the vendor was telling investors that its platform is a harness that makes the data, the workflows and the business logic inside Salesforce available to every agent, model and interface. That is the position I took in 2023, stated by the vendor, with a toll booth on it: the screen is conceded, and the keys are not.
What has changed for the prisoner is exit. For the first time the person doing the work can leave the screen without her company leaving the vendor, and Hirschman’s point about voice is that it is as loud as the exit behind it. The buyer’s part is to write the door into the contract: the right to read and write your records with whatever mind you choose, at the cost of carrying them, for as long as you are a customer and on the day you stop being one. A vendor confident in its product will sign that, and a vendor that refuses has told you what it thinks its product is.
I do not know how this plays out, and I would not trust anyone who says they do, least of all anyone with seats to sell. The bet I made in 2023 has seven of its ten years left: that we will look back on the enterprise applications of this era the way we look into the drawer, with some affection and no wish to go back. Beyond Work, the company I run, sits on the far side of that bet, and the reader should price that in.
The deal you closed on Tuesday was real on Tuesday. It was the database that needed Friday afternoon, a form and thirty fields, and the database was the one party to the transaction that could have been asked to do the work itself. Tomorrow will not be like yesterday, and for anyone who has spent a Friday afternoon telling a machine what they did on Tuesday, that is good news.19