Contents7
“Build or buy” is an old IT leadership question. AI has reopened it: if a coding assistant writes the code in a few hours, why pay a per-seat subscription for software that only half does what you want? The large consulting firms do not give the same answer, and their advice is not aimed at the same companies.
This article steps back from the debate. For the practical decision, tool by tool, our guide Business application: custom or off the shelf? gives the decision grid and the three-year comparison.
In short
Deloitte, which sells consulting, audit and AI services, advises private companies not to waste resources building what they can buy.
BCG, which sells AI consulting, wrote in September 2026 that companies should start small, build rather than buy, and be selective.
At EY, which sells consulting and AI services, 76% of the US executives surveyed say off-the-shelf software no longer meets their needs, but 72% say the software they built internally with AI is causing difficulties.
For a 50-person team, the real question is not “build or buy” but “who will run this software in three years”.
Two large firms, two opposite pieces of advice
The contradiction is striking when you read the two texts side by side.
Deloitte: buy
Guide for private companies, published in 2026
Do not waste resources building an AI capability you can buy
Logic: move fast, reduce risk
BCG: build
September 2026 text on its AI agents page
Start small, build rather than buy, be selective
Logic: keep control of what differentiates you
The Deloitte Private guide is written for companies that often have neither a development team nor the budget to build one. BCG’s text, published on its AI agents page, speaks to organisations that do, and for which software is a lever of differentiation. Both pieces of advice can be right, each for its own reader. Neither is about a 50-person team.
What executives of large companies say
EY, which sells consulting and AI services, surveyed 534 US executives at SVP level or above in 2026, in organisations already investing in AI. This is not an SME sample, and the base excludes companies that have not started.
say off-the-shelf software no longer meets their specific needs
have deployed or are testing internal software built with AI
say difficulties with that software are slowing their progress
The three figures tell a consistent story: the appetite to build is strong, the first results are harder than expected. The same mix of enthusiasm and caution appears elsewhere: BCG counts only 11% of large companies that rely mainly on in-house development for their AI, according to its “The Widening AI Value Gap” study of 1,250 executives, 90% of them at companies above one billion dollars (BCG, 2025).
For European companies, public statistics are unambiguous. According to INSEE, 80% of French companies with 10 or more people that use AI acquire it by buying ready-made software. Across the EU, 57.9% of companies that use AI buy ready-made software and 21.8% develop in-house, with the modes overlapping (Eurostat, isoc_eb_ai, 2024 data). Buying remains the norm, by far.
AI speeds up writing code, not the rest
The core argument for building is that AI has lowered the cost of development. That is true for part of the work, and less clear than often claimed.
- February 2023
In a lab trial on a single task, developers using an assistant write a web server 55.8% faster than the control group.
- July 2025
In a randomised trial by METR, 16 experienced developers take 19% longer with AI on their own projects, while believing they saved 20%.
- February 2026
METR measures again and finds completion time down 18% for the original developers and 4% for new ones, a signal it considers unreliable itself.
The honest reading: on an isolated, well-defined task, AI saves a lot; on a real project, with its history and constraints, the gain is uncertain and perception overstates it. Above all, writing code is only part of what software costs. There remain design, security, updates, backups, data migration, support, and the person who understands the code in three years. We developed this in Vibe coding: AI codes fast, the product stays human.
Key takeaway
Most of a software’s cost comes after go-live
AI cuts the time to write the first draft. It cuts neither maintenance nor responsibility for what runs in production.
What decides for a 50-person team
For a company of 20 to 500 people, we boil the decision down to four questions, in this order.
- Does this process set you apart?
Accounting, payroll and email are the same everywhere: buy. The process that makes clients choose you (a complex quote, site follow-up, a particular customer journey) deserves a closer look.
- Does an open tool cover 80% of the need?
If so, extending that tool almost always costs less than starting from scratch, and you keep its updates.
- Who will run the software in three years?
Without someone to maintain it, in-house or at a committed provider, built software becomes debt at its first incident.
- What does buying cost over three years?
Per-seat subscriptions climb with the team; compare over three years, not on the starting price. Our method is in Leaving SaaS: the real three-year cost.
The pricing model changes the calculation too. At EY, 82% of executives surveyed think per-seat pricing will become less relevant in their sector. If your vendors move to usage-based billing, bought software can become more or less expensive, and the comparison has to be redone.
What we think
The debate between Deloitte and BCG pits two extremes against each other. For most teams of 20 to 500 people, the right answer seems to us to lie in the middle: buy what is standard, extend an open tool for what matters, and only build from scratch as a last resort, on a narrow scope.
From the field
Extending rather than building, since early 2025
We have run Twenty, an open source CRM, in production since early 2025, notably as the CRM of Palmora Property, a real estate portal in Thailand. Rather than writing a CRM, we built a business application on top of it and published it on the Twenty store. We keep the tool’s updates, and we only maintain what is our own.
Our bias is known: we sell application development. That is precisely why we start every project with the question of buying. Software built for a need an off-the-shelf tool already covered is spending nobody will thank us for.
Where to start
List your software and, for each one, answer the four questions above. Most will land on “buy”. For the one or two processes left, compare extending an open tool with development, as a three-year cost. Our guide to app development cost helps you put numbers on it, and Low-code or code helps you choose the tooling. If development wins, our app development offer starts by scoping that narrow perimeter.
Frequently asked questions
Should you build your own software with AI in 2026?
Only for the process that sets you apart, and only if someone can maintain it. AI speeds up writing code, but studies remain divided on the real gain across a full project, and maintenance, security and support do not get cheaper. For the rest, buying remains the norm.
Why don’t Deloitte and BCG say the same thing?
Because they are not speaking to the same readers. Deloitte’s guide targets private companies, often without a development team; BCG’s text addresses organisations that have one and want to differentiate. Both firms sell AI consulting either way.
Is extending open source software building or buying?
It is a middle path: you start from an existing, maintained tool and only write what is specific to you. The cost is lower than full development and you keep the tool’s updates. The condition is choosing a tool whose licence and architecture allow such extensions.
Does AI reduce the cost of custom software?
It reduces the time to write part of the code. It does not reduce the cost of design, testing, security, hosting and maintenance, which often weigh more over three years. Always compare the total cost, not just the development price.
Sources
- Deloitte Private (new tab)deloitte.com
- AI agents page (new tab)bcg.com
- BCG, 2025 (new tab)media-publications.bcg.com
- INSEE (new tab)insee.fr
- Eurostat, isoc_eb_ai (new tab)ec.europa.eu
- 55.8% faster (new tab)arxiv.org
- 19% longer (new tab)arxiv.org
- Unreliable (new tab)metr.org
Jim · founder of Brain Plus AI
Also in web and mobile apps
Choose and price6 min
Business app: custom-built or off-the-shelf software?
Off-the-shelf software costs less as long as your way of working fits its boxes. Here are the signs that tip the balance towards custom, and the middle path that is often forgotten.
Stepping back22 min
Vibe coding: AI codes fast, the product stays human
AI has removed real technical barriers, but it has not removed what makes a product. What is true and what is false in "anyone can do anything with computers", with examples in both directions.
Choose and price6 min
Low-code or code: how far you can go without developers
Power Apps, AppSheet, Bubble, n8n: no-code tools let a business team quickly build useful tools, with limits on price, dependency and security. Here is where they excel, where they stop, and how to combine them with code.