Part 2 ended with a company somewhere trying to fold back in what its people built with agents, and the question of whether it would ban it first. This part is about doing it well. It’s the one I wrote for people who decide what software a company buys, so it has more numbers and fewer bathtubs.
The spreadsheet came with a bill
In part 2 I called VisiCalc my favourite row in the table: people carried Apple IIs into the office, did their own sums without the data-processing department, and spreadsheets became the backbone of finance from the outside in. What I left out is the bill.
In 2012 JPMorgan’s Chief Investment Office lost at least $6.2 billion on the trades that became known as the London Whale (US Senate, 2013). The bank’s own task-force report found that the new risk model ran “through a series of Excel spreadsheets, which had to be completed manually, by a process of copying and pasting data from one spreadsheet to another”. One step divided by a sum where the modeller meant to divide by an average, which likely muted the volatility by a factor of two (JPMorgan Management Task Force, January 2013).
In April 2013, Thomas Herndon at UMass Amherst tried to reproduce a famous paper by Carmen Reinhart and Kenneth Rogoff, the one politicians had quoted as evidence that public debt above 90% of GDP chokes growth. With Michael Ash and Robert Pollin he found that a coding error in the working spreadsheet left out five countries entirely: Australia, Austria, Belgium, Canada and Denmark. To be fair to the spreadsheet, that error was the smallest part of the gap; most of it came from data left out and an unusual weighting. Corrected, average growth for countries with debt above 90% of GDP came out at 2.2%, not -0.1% (Herndon, Ash and Pollin, April 2013).
Nobody took the spreadsheets away after either of those (I can’t imagine the finance department that would’ve let them). Building around IT has always cost something, and I think it was still worth channelling instead of banning.
What I built
At a company I work with, there was no solution for translation management. What was on the table instead was a requirements-engineering process measured in months, and architecture working groups debating which translation software to buy. So I built one, on top of the existing code repositories and running on the company’s own infrastructure.
I did the same with a CRM sync, and with a trial setup that connects the CRM to the customer portals. The trial setup replaced the tutorials and training sessions on how to set up a customer trial.
All three were built with Claude and OpenSpec, which means the spec came first and the code followed. The spec is plain text that anyone can read and argue with before there’s any code, and that’s most of what a months-long requirements process is for (writing it down is now most of the work of building it). Here the arguing happened next to something that already ran.
So when I say shadow IT in this post, I’m including myself.
I’m not unusual
Microsoft and LinkedIn’s 2024 Work Trend Index surveyed 31,000 knowledge workers in 31 markets, and 78% of the ones using AI were bringing their own tools to work. I mentioned that number in part 2; it’s two years old now, and I’d bet it hasn’t gone down.
A year later MIT’s NANDA project published The GenAI Divide (MIT’s own link now redirects; that’s a mirror). “While only 40% of companies say they purchased an official LLM subscription, workers from over 90% of the companies we surveyed reported regular use of personal AI tools for work tasks.” In the same report, most AI pilots were stuck “with no measurable P&L impact”, and 95% of organisations were getting zero return. The sample’s small (interviews with 52 organisations, 153 senior leaders surveyed, a review of 300-odd public AI initiatives), and the authors call their own figures directional, so I wouldn’t put the decimals in a board deck. In their sample, though, the unofficial track was more than twice as widespread as the official one.
The risk, taken seriously
If you’re the CISO, you’ve been waiting for this section. In March 2023, within weeks of being allowed to use ChatGPT, engineers in Samsung’s chip division pasted in the source code of two internal programs and the transcript of a meeting (Economist Korea, 30 March 2023). In May, Samsung told staff in one of its biggest divisions to stop using generative AI tools for now (Bloomberg, 2 May 2023). Data leaving the building, decisions nobody can trace and no audit trail are all real, and some of what people build at home will be wrong in the same way the Reinhart-Rogoff spreadsheet was wrong.
My own mistake was a different one, and I think it’s the more common one. I set some of those tools up on a personal domain of mine. If I left tomorrow, the company would be running something that hangs off an address I own. The risk I created was about ownership and continuity.
That’s the trouble with a ban. It doesn’t take away the reason people reached for the tool, so the use moves to phones and home laptops, where nobody can see it. Seeing it is the control that works, and a ban gives that up first.
Why the seat breaks
Most business software is priced per seat. A seat prices access to a screen: one person, one login, one more licence. That works as long as the work happens on that screen.
It’s leaving the screen. When one person with a few agents does what used to take three, or the work happens through an API without anyone logging in, the seat count stops tracking the value, and the vendors already price for that. Intercom prices its Fin agent at $0.99 per outcome, which counts, for example, a conversation where the customer asks for no further help after Fin’s last answer (Intercom). Salesforce launched Agentforce in 2024 at $2 per conversation, and in May 2025 added Flex Credits at $0.10 per action (Salesforce, September 2024; May 2025). Neither price cares how many people logged in, and if the vendor prices the outcome, I’d want to buy the outcome too.
In December 2024 Satya Nadella said on the BG2 podcast that business applications are “essentially CRUD databases with a bunch of business logic”, that “the business logic is all going to these agents”, and that the agent era is “probably where they’ll all collapse” (BG2, 12 December 2024). He was talking about Microsoft’s own Dynamics at the time.
In late January and early February 2026 the market took his side for a week or so. Anthropic released a legal plug-in for Claude Cowork, its desktop agent (yes, the company whose model I write this series with), and software and services stocks lost $830 billion in market value in six trading days (Reuters, 4 February 2026). Marc Benioff, who sells more seats than almost anyone, answered on Salesforce’s earnings call on 25 February: “We have been through many SaaS apocalypses”, and “the apps aren’t gone away” (Salesforce Q4 FY26 call).
I’d guess they’re both right, about different parts of the stack. In my own case the CRM is still there, and what disappeared was the tutorial on how to click through it.
What to own instead
If the seat is the wrong thing to buy, the useful question is what you need to keep. I’d keep the data layer, the interfaces between tools and the right to leave.
By data layer I mean your customers, your contracts, your translations, in a form you can read without the vendor’s app. Interfaces are the glue between tools, and I’d build them on an open standard like the Model Context Protocol (more on that in part 2, and in how I moved the fourteen MCP servers behind my own setup), so the model on the other end can change. The right to leave sounds abstract until it costs money. After Broadcom closed its takeover of VMware in November 2023, it stopped selling perpetual licences and support renewals for them and moved customers onto subscriptions (VMware, January 2024). AT&T later sued, and one of the emails it filed in that case answers Broadcom’s offer with “This proposed annual increase of +1,050% in one year is extreme” (AT&T filing, 2024; The Register).
Channelling it, concretely
Channelling it takes four things, and they fit on one page.
- One owner. A named person, not a committee, who knows what’s been built and who to ask about it.
- Logged inputs and outputs. Which model was used, what data went in, what came out. That’s the audit trail the CISO wanted.
- A fixed monthly AI budget per team, instead of more seats. The team decides what to spend it on, and you see what they chose.
- A graduation rule. Once other people depend on a tool, it moves onto a company domain with a company owner. That’s the rule I needed for my own mistake, and I didn’t have it.
Part 4, next Tuesday, is about how you get to trust what these tools do.
Three questions for the next leadership meeting
- Which of our workflows exist only as someone’s prompt?
- What would it cost us to leave our biggest SaaS vendor?
- Who in this room built something with an agent this month?
Sources
- JPMorgan Chase, Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses, 16 January 2013 (copy held by the Yale Program on Financial Stability)
- US Senate Permanent Subcommittee on Investigations, JPMorgan Chase Whale Trades: A Case History of Derivatives Risks and Abuses, 15 March 2013
- Thomas Herndon, Michael Ash and Robert Pollin, Does High Public Debt Consistently Stifle Economic Growth? A Critique of Reinhart and Rogoff, PERI Working Paper 322, 15 April 2013
- OpenSpec, accessed November 2026
- Microsoft and LinkedIn, 2024 Work Trend Index: AI at work is here. Now comes the hard part, 8 May 2024
- Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025 (mirror; MIT’s own link now redirects)
- Economist Korea, Samsung ChatGPT leak report (in Korean), 30 March 2023
- Mark Gurman, Samsung bans staff’s AI use after spotting ChatGPT data leak, Bloomberg, 2 May 2023
- Intercom, Fin pricing, accessed November 2026
- Salesforce, Agentforce announcement, 12 September 2024, and flexible pricing, 15 May 2025
- BG2 Pod, Satya Nadella with Brad Gerstner and Bill Gurley, 12 December 2024 (from 46:44)
- Reuters, Global software stocks hit by Anthropic wake-up call on AI disruption, 4 February 2026
- Salesforce, Q4 FY26 earnings call transcript, 25 February 2026
- Anthropic, Introducing the Model Context Protocol, 25 November 2024
- VMware, End of availability of perpetual licensing and SaaS services, 22 January 2024
- AT&T v. Broadcom, court filing, 2024, and Simon Sharwood, AT&T claims VMware offered it a 1,050 percent price rise, The Register, 1 October 2024
MAKE YOUR CASE.
If this poked at something you're building: talk to me.
