For thirty years, the buy-vs-build decision had a stable answer: buy the commodity, build the differentiator. Salesforce instead of a homegrown CRM. Stripe instead of a payment switch. Postgres instead of a custom store. The reasoning was always the same — building cost real engineer-years, and engineer-years were the scarce resource. So you spent them only on the things that gave you a moat.
Then AI coding assistants happened. Cursor, Claude Code, Copilot, and the new wave of agentic coding tools compressed the cost of building software by something like 5-10x for a competent senior engineer, and 20-50x for a junior with the right scaffolding. The dam broke. Suddenly the same engineer who would have spent six months on a CRM integration could ship the same thing in three weeks.
Predictably, an entire generation of engineering leaders concluded the obvious thing: we should build more. And just as predictably, most of them are wrong about which things to build.
What actually changed (and what didn't)
AI coding assistants made three things dramatically cheaper:
- Greenfield code — typing the first version of something that didn't exist. The blank file is no longer scary; it's a 90-second prompt.
- Boilerplate and glue — the 60% of any codebase that's just plumbing. AI tools eat this for breakfast.
- Translation between known patterns — porting a Rails app to Next.js, converting a REST API to GraphQL, wiring up a new ORM. If a pattern exists in the training data, the assistant will reproduce it competently.
But three other things did not get cheaper, and most of them are exactly where the long-term cost of software lives:
- Maintenance — keeping a system running for five years through migrations, security patches, dependency churn, and shifting requirements. AI assistants help during the diff, but the operational ownership is unchanged. Someone is still on the pager.
- Integration drift — when your Salesforce schema changes, when the upstream API you depend on adds a required field, when the auth provider rotates a contract. The AI can help you fix it, but only after a human notices it broke.
- Domain knowledge — knowing that this customer's invoices use European date formats, that this regulator audits the third Tuesday of every month, that this SKU schema was a hack from 2019 nobody dared touch. AI assistants do not have access to your tribal knowledge until you painstakingly write it down.
The first list is the cost-of-build. The second list is the cost-of-ownership. AI assistants slashed the first one. The second one is unchanged, and for most internal systems, the second one is 80% of the lifetime cost.
The new math
Here is the buy-vs-build calculation as of May 2026:
Old build cost (pre-2024): 6 months × 2 engineers × fully-loaded $300K/year ≈ $300K to ship V1, then ~$200K/year in maintenance.
New build cost (with AI assistants): 6 weeks × 1.5 engineers ≈ $50K to ship V1, then still ~$200K/year in maintenance.
Buy cost (any decent SaaS): $50K-$150K/year, all-in, with vendor-side maintenance.
The build V1 cost dropped from $300K to $50K — a 6x compression. The buy running cost is unchanged. So if you naively look at year-one numbers, building looks dramatically more attractive than it did in 2023. That's the seductive mistake.
The five-year picture flips it back:
- Build: $50K + (4 × $200K) = $850K
- Buy: 5 × $100K = $500K
The buy bucket still wins on five-year lifetime cost for any system that's even slightly complex, because the operational cost dominates the build cost as soon as the V1 ships. AI assistants didn't change that — they only made V1 cheaper.
When AI assistants tilt the math toward build
That said, there are categories where the new math genuinely changes the answer. Three of them:
1. The "small-T tools" category
Internal CLIs, dashboards, one-off scripts, migration tools, deployment helpers, observability glue. These used to be expensive to build and impossible to buy (no vendor sells a script that processes your specific CSV format). They were the dark matter of engineering org backlogs — never quite worth doing, never quite going away.
AI assistants killed that backlog. A junior engineer with a coding assistant ships ten of these in a week. Build is the right answer here, and the answer wasn't even available three years ago.
2. The "aggressively customised commodity" category
A generic CRM is commodity; your CRM, with seven custom objects and a fourteen-step lead-routing workflow, is not. In 2023, you'd customise the SaaS and accept the constraints. With AI assistants, building a thin custom CRM on top of a primitive (Postgres + a UI library + an event bus) is now genuinely tractable for a team of two.
This is the most interesting category because the answer used to be obvious (buy and customise) and now it's a real coin-flip. The deciding question becomes: will the requirements stabilise, or will they keep mutating? If they keep mutating, custom wins. If they stabilise, the SaaS still wins on operational cost.
3. The "regulatory or data-locality" category
Workflows that can't leave your environment for compliance reasons used to be a build by force. AI assistants made them genuinely affordable, where before they were a tax. This isn't a math change so much as a relief: the thing you had to build is now actually buildable in a reasonable timeframe.
When the math still says buy (loudly)
For everything else — and "everything else" is most of the software your company runs on — the answer is still buy. Specifically:
Anything with a marketplace listing. If three vendors sell a workflow with public benchmarks, the category is mature. Building it in-house is a roadmap tax with no upside. The 2026 AI marketplace and AI workflow marketplace ecosystems exist precisely to make this decision obvious.
Anything with a real network effect. Slack, GitHub, Linear, Stripe — the value isn't the code, it's the network. AI assistants don't help you build a network.
Anything that's regulated and standardised. Payroll, payments, identity. The vendors carry the compliance burden for you. That's worth a 10x premium.
Anything that requires 24/7 operational excellence you don't have. Auth, databases, email delivery, observability. You don't want your team woken up by these systems. You want a vendor's team woken up.
Anything you'd build "the same as everyone else." If your version would be 90% identical to the off-the-shelf version, you're building a worse version of someone else's product. Stop.
The "buy with marketplace, build the composition" pattern
The pattern that's working in 2026 is neither pure-buy nor pure-build. It's buy the units of capability from a marketplace, build the composition layer that's specific to your business.
Concretely:
- Buy invoice extraction from the workflow marketplace.
- Buy ticket triage from the workflow marketplace.
- Buy a Slack MCP server from the MCP marketplace.
- Buy a TTS voice from the voice marketplace.
- Build the orchestration that says: "when an invoice comes in, run extraction, route exceptions to a specific Slack channel based on amount, and have the voice agent call the supplier to confirm anomalies over $10K".
The composition layer is where your tribal knowledge lives. The unit-of-capability layer is where your tribal knowledge does not live. AI coding assistants are spectacular at helping you build the composition layer, because that layer is small, clearly-scoped, and changes slowly compared to the marketplace primitives underneath it.
This pattern compounds. As the marketplaces mature — AI workflow marketplace, MCP marketplace, avatar marketplace, voice marketplace, model marketplace — your composition layer gets thinner, your buy bucket gets thicker, and your team's leverage goes up.
The five questions to ask before you build anything in 2026
When an engineer on your team proposes building something instead of buying it, ask these:
- Does this exist in three or more marketplace listings? If yes, building is a roadmap tax. Buy.
- Will we maintain this for five years, or is it a one-shot? One-shots are now build territory. Five-year systems are still buy territory.
- Is the "moat" in the code or in the data? If the moat is data, buy the code that operates on it. If the moat is code, build it — but be honest, the moat is almost never the code.
- What's the operational cost? Add 5x your V1 estimate. Then compare. If V1 with AI assistants is $50K and operations is $250K over five years, you're comparing $300K to a $500K SaaS. That's closer than year-one math suggests.
- What happens when the AI assistants improve another 5x next year? Code you wrote in May 2026 with AI assistance will look like 2024-era code in 2027. Code you bought from a vendor will be quietly upgraded by their team. Long-term, vendor code converges on best practice; your code stays at the level it shipped at.
Question 5 is the one most leaders skip. AI coding assistants are a tide that lifts all boats — including the boats your vendors are sitting on. The compression you're feeling on internal builds is also compressing the cost structure of your SaaS providers. Some of that will show up as price drops. More of it will show up as products that get dramatically better without you doing anything.
The honest summary
AI coding assistants made the first version of internal software 5-10x cheaper. That's a real and large change. It moves the buy-vs-build line meaningfully — but mostly for small tools, aggressive customisation, and regulated workflows.
For the bulk of software a company runs on — CRM, support, payments, identity, observability, marketing automation, document workflows — the answer is still buy. The new wrinkle is what you buy. Increasingly, you're not buying full-stack SaaS; you're buying units of capability from a marketplace and composing them with a thin custom layer that AI assistants help you ship in days, not quarters.
That composition layer is the real differentiator. It's where your tribal knowledge lives, where your business logic lives, and where the AI assistant adds the most leverage. Spend your engineering capacity there. Buy everything else.
The companies that get this right in 2026 will look like they have a small team shipping enormous output. They don't have more engineers — they have a thicker buy bucket, a thinner build bucket, and a composition layer that AI assistants made tractable.
That's the new shape of the discipline.
Looking for the buy side? Browse the Swfte AI marketplace for workflows, MCP servers, avatars, voices, and models that install in minutes. Or read The AI Workflow Marketplace for the unit-economics breakdown.