Kiseki Labs Request a briefing
Business

Build vs Buy AI Is Settled. Governing Shadow AI Isn't

03.06.2026 Updated 06.10.2026 9 min read Eddie Forson

76% of AI use cases are now bought, and much of that buying started with staff, not IT. Why governance is the real decision, plus a 30-day shadow AI audit.

Neon visualisation of a glowing central AI hub linked to official dashboards, surrounded by dozens of dim, unsanctioned AI chat windows half-hidden in the shadows between data towers

Only 40% of the companies MIT surveyed for its 2025 GenAI Divide research had bought an official LLM subscription. Yet workers at over 90% of them were using personal AI tools for work. So who made the build versus buy decision?

Menlo Ventures' numbers suggest the debate is over: 76% of enterprise AI use cases are now bought rather than built. We think that settles the wrong question. A large share of the buying started before anyone asked, on personal accounts and credit cards. The decision still open is how to govern what's already in use.

Build vs buy AI: what the numbers show

Menlo has tracked the split for three years, and since 2024 the direction is clear.

YearWhat Menlo measuredBoughtBuilt in-house
2023Enterprises buying third-party generative AI software80%Not reported
2024AI solutions53%47%
2025AI use cases76%24%

The unit changes between editions, so read it as a direction, not a precise series. 2024 looks like the high-water mark for building. Menlo is an investor in Anthropic and in several of the product-led tools it cites, so weigh its enthusiasm for them accordingly.

The more revealing numbers are about how the buying happens. In the 2025 report, 27% of AI application spend comes through product-led growth, where individuals adopt a tool before anyone signs a contract, against 7% in traditional software. Add shadow adoption, meaning staff paying for tools like ChatGPT Plus on personal cards, and Menlo estimates product-led tools may account for close to 40% of application AI spend.

Cursor reached $200 million in revenue before it hired a single enterprise sales rep. And once an organisation commits to exploring an AI product, 47% of those deals reach production, against 25% for traditional SaaS. In AI, the contract often arrives after the usage, sometimes long after.

What is shadow AI?

Shadow AI is any AI tool, model or feature used for work without the organisation's approval or oversight. The obvious cases are personal ChatGPT, Claude or Gemini accounts. The less obvious ones matter just as much: free-tier tools with training rights buried in their terms, browser extensions that read every page, AI features quietly switched on inside software you already pay for, coding assistants and command-line agents connected to your repositories, and API keys paid for on someone's own card.

Most writing on shadow AI treats it purely as a security problem. It is one. It's also the best demand signal most leadership teams have. Every unsanctioned tool is an employee telling you, with their own money, where AI saves them time. MIT's researchers made the same point: "Forward-thinking organizations are beginning to bridge this gap by learning from shadow usage and analyzing which personal tools deliver value before procuring enterprise alternatives."

Martin Casado of a16z has watched the top-down version play out from the boardroom. "What does the board say? We need more AI. And what does the CEO say? Oh, okay, I'll get like a consultant to do more AI. And then they have some centralized project that nobody knows how it works… those things will fail." Meanwhile, as he put it, "everybody's using ChatGPT very effectively." We're consultants, and we agree with him. A programme that starts by stamping out the usage that works throws away its best evidence.

Build a little, buy a lot, govern all of it

In practice most organisations are doing three jobs at once.

When it's the right callWhat it costs you
BuildThe workflow or data is an advantage a competitor can't buy, such as proprietary pricing or underwriting logicMaintenance, evaluation and every model upgrade, for as long as it runs
BuyThe need is near-term and productivity-shaped, such as coding assistants or meeting notesData terms, lock-in and usage-based bills to watch
GovernAlways, starting with what's already in useA few focused weeks, then a light ongoing process

Our rule on building is deliberately strict. A lot of "we'll build it ourselves" instinct is engineering pride, and it's expensive pride. MIT found that external partnerships with learning-capable, customised tools reached deployment about 67% of the time, against about 33% for internal builds. The researchers are careful to say that doesn't prove causation, and the figures come from interviews, not market data. Anything that fails the advantage test belongs in the buy column, and the engineering time it frees up is often the bigger return.

Building still pays in the right places, and AI coding tools have made it far cheaper. Shawn Wang (swyx) of Latent Space reckons he could replace a SaaS product his company pays about $200,000 a year for with a custom build costing around $2,000.

But AI has made building cheaper without making owning any cheaper. A system you build has to keep up with the models underneath it. Menlo noted that Claude Sonnet 4.5's release prompted Cognition to rebuild Devin entirely, and that Manus has reconstructed its agent framework four times. If you build, budget for the rebuilds. Budget too for an evaluation suite, a fixed set of real cases with known right answers, so you can tell whether each rebuild actually improved anything.

Where the risk goes when you buy AI

Buying moves the risk somewhere else: from "can we build this?" to "can we integrate, secure and govern this?"

Data leaks quietly through tools nobody approved. It's a governance problem that looks like a productivity win, and it tends to surface at the worst moment, in a security review or a customer's due-diligence questionnaire. At a London Software Guild meetup in June, Daniel Trugman of Requesty, which sells software that sits between a company's apps and its AI providers, described a retroactive scan of traffic through Requesty's own gateway. It found "a lot of full names, a lot of email addresses" heading to model providers, along with access tokens, Stripe's among them.

The other failure is welding a critical workflow to a single model. Models are superseded fast: Menlo found that within a month of Claude 4's release, Claude Sonnet 3.5's share of Anthropic users fell from 83% to 16%. Sometimes they vanish outright, as Fable 5 did for every customer for nearly three weeks in June under a US government directive. And the lock-in deepens as systems learn your business. MIT warned in 2025 that in the following few quarters, "several enterprises will lock in vendor relationships that will be nearly impossible to unwind."

Buying is still usually the right call. Negotiate data and portability terms up front, avoid hard-coding any workflow to one model, and keep an evaluation suite that can tell you whether an alternative is good enough before you need one.

How to run a shadow AI audit in 30 days

You can't write an AI strategy on top of a stack you can't see. We apply the same rule to strategy that we apply to engineering: establish ground truth first. For AI adoption, ground truth is an AI tool register, a list of what people actually run and what data it touches, and a month is enough to build a first version. Name the owners at the start: a senior sponsor from the business, the CIO or CISO, and your data protection officer.

Week 1: shadow AI detection

No single source catches everything, so ask IT to combine several: AI subscriptions in expense and corporate card data; AI applications granted access to your Microsoft 365 or Google Workspace tenant through single sign-on or OAuth; web gateway or DNS logs showing traffic to model providers; browser extensions on managed devices; and the AI features switched on inside software you already license. Pay particular attention to coding assistants and command-line agents, which can read whole repositories, and to the connectors people have granted them. Inventory the API keys the company has issued, and run secret scanning across your repositories for the ones it hasn't.

Then ask people directly. A short, no-blame survey asking what they use and what it saves them will surface tools the logs miss, along with the business case for each. Frame it as learning, or you'll get silence.

Week 2: map what data goes where

Triage the top 20 or so tools by usage and data sensitivity instead of mapping everything at once. For each, record which classes of data pass through it (customer personal data, source code, financials, credentials), whether the provider can train on inputs, and how long it keeps them. For UK and EU organisations, add two questions: do we have a data processing agreement with this provider, and what's the transfer mechanism if data leaves the UK or EEA? A personal account gives the organisation no contract at all, so personal data should never go through one. Rotate any credential you find in a prompt.

The most useful question to ask of every tool is simple: can it reach a model provider without passing through anything we can inspect?

Week 3: decide to sanction, replace or retire

Every tool on the register gets one of three decisions. Sanction it by moving it onto enterprise terms with single sign-on and an agreed data policy. Replace it with an enterprise product that does the same job, using the shadow usage as proof of demand. Or retire it, and name the alternative at the same time, so usage moves somewhere visible instead of further underground. Procurement won't finish in a week, but the decisions can.

Week 4: build the governed path

The goal is to make the approved route faster than the shadow one. That means a published list of approved tools, a one-page usage policy organised by data class, and an intake route that approves a new tool in days. Add spend visibility by team and, ideally, a single inspection point for API traffic to model providers.

Our Opportunity Diagnostic covers the decision end of this work: two to three weeks, a prioritised and costed shortlist of use cases, a build versus buy call on each one, and risks and governance named up front. We'd start it from what your teams already use, because shadow usage is the best early evidence of where AI will pay.

Frequently asked questions

What is shadow AI?
Shadow AI is the use of AI tools, models or features for work without the organisation's approval or oversight. It includes personal chatbot accounts, free-tier tools, browser extensions, AI features inside existing software, coding agents and API keys paid for personally.

Is using ChatGPT at work shadow AI?
It depends on the account, not the product. A personal ChatGPT account used for work, outside any agreed data terms, is shadow AI. The same tool on an enterprise workspace with single sign-on and an agreed data policy is sanctioned AI.

Should we build or buy AI?
Buy by default for anything near-term and productivity-shaped. Build only where the workflow or data gives you an advantage a competitor can't purchase, and budget for maintaining it through model changes. Either way, govern what's already in use first.

How do you detect shadow AI?
Combine expense and card data, single sign-on and OAuth grants, web gateway or DNS logs, browser extension inventories, API key inventories, secret scanning and a no-blame staff survey. No single source catches everything.

Should we ban shadow AI?
A blanket ban tends to push usage out of sight and throws away the best evidence of where AI creates value. Restrictions can be right for specific data classes or tools. Sanction what works, replace what's risky, and make the approved route faster than the shadow one.

An executive briefing is a quick way to find out what's already running in your organisation, and what it tells you about demand.


This article first appeared on 3 June 2026 as an edition of AI for Decision Makers, Eddie Forson's LinkedIn newsletter, under the title Buy, Build, or Govern? The AI question most CTOs are getting backwards. We've since updated it with what happened next and corrected a few details.

Next

Facing something similar?

Thirty minutes on your situation, and a straight answer on whether the work is worth doing.

Not ready for a conversation? Take the AI readiness assessment: twenty questions, about six minutes, and you'll see where your organisation is stuck before you talk to anyone.