"We already pay for ChatGPT — why would we build anything?" It is a completely fair question, and any AI development company that answers it with "because custom is better" is selling, not advising. Sometimes the subscription is genuinely the right answer. This guide gives you the honest dividing line, the real three-year numbers on both sides, and a test you can run this week to decide.
Short answer: Buy off-the-shelf when a person is doing general knowledge work in a chat window. Build custom when the work must run without a person present, must read and write to your systems, or follows a process specific to your company. Most companies that get this right end up doing both.
1. The Real Dividing Line (It Isn't Model Quality)
The most common misconception is that custom AI means a smarter model. It does not. Custom builds usually call the same frontier models you already have access to. What you are buying is not intelligence — it is plumbing, autonomy, and control.
An off-the-shelf assistant sits and waits for a human to paste in context, read the output, and act on it. That human is the integration layer. It works well, and it is cheap. The moment you want the work to happen without that person — at 2am, on every incoming ticket, across ten thousand records — the human-shaped gap in the middle becomes the problem, and no subscription tier fixes it.
2. Where Off-the-Shelf Genuinely Wins
We recommend this path more often than clients expect. Buy when:
- The task is general knowledge work. Drafting, summarising, rewriting, brainstorming, coding help, analysing a document someone pastes in.
- A human is always in the loop anyway. If someone reviews every output, you do not need engineered guardrails — you have a reviewer.
- You need value this week. Same-day deployment beats a two-month build when the need is immediate and modest.
- You are still discovering demand. Do not build for a workflow nobody has proven they want. Subscriptions are the cheapest possible demand experiment.
- The workflow is commoditised. Meeting notes, transcription, standard document Q&A. If a product does it well already, building it is a waste of capital.
3. Where Custom Development Wins
Build when one or more of these is true:
- It must run unattended. Every ticket triaged on arrival, every invoice reconciled overnight, every lead qualified before a human sees it.
- It needs your data reliably, with citations. Not a document someone remembered to paste, but live retrieval across your actual knowledge base with sources attached. That is a RAG system, and it is a build.
- It must write back into your systems. Updating the CRM, creating the ticket, issuing the refund. Reading is easy; taking actions safely is engineering.
- The workflow is your advantage. If your process is why customers choose you, encoding it into a generic tool everyone can buy gives that advantage away.
- You have hard compliance or residency requirements. On-premise processing, audit trails, or retrieval that strictly enforces internal permissions.
- Volume makes per-seat pricing absurd. Ten thousand documents a month processed by API costs a fraction of the equivalent human seat-hours.
4. The Three-Year Cost, Honestly
Year-one comparisons flatter subscriptions, because a build cost lands all at once and seat fees arrive quietly every month. Three years is the fair window. The key structural difference: subscription cost scales with headcount; custom cost scales with scope.
| Dimension | Off-the-Shelf | Custom Build |
|---|---|---|
| Upfront cost | None | $20,000 – $120,000+ |
| Recurring cost driver | Per user, per month | Usage & hosting |
| Cost as headcount doubles | Roughly doubles | Barely moves |
| Time to first value | Same day | 3 weeks – 4 months |
| Runs unattended | No | Yes, by design |
| Maintenance owner | Vendor | You or your partner |
| Switching risk | Price and terms can change | You own the code |
Enterprise AI seats commonly run $25–$60 per user per month. Run the arithmetic for your own headcount over 36 months before assuming the subscription is the frugal option — at a few hundred seats, the totals typically converge, and a custom system serving unlimited internal users on usage-based pricing starts to look very different.
The cost trap on both sides. Buyers underestimate subscriptions because seat counts creep — you approve 40 licences and audit 220 eighteen months later. Buyers underestimate custom builds because they budget the build and forget operations; expect year-one all-in to be 1.4×–1.8× the build quote. Our AI agent development cost guide breaks that multiplier down line by line.
5. The Decision Test
Take one specific workflow — not "AI for the company", one workflow — and ask three questions in order.
- Could a capable new employee do this with only a public web browser? No access to your CRM, database, or internal drives. If yes, an off-the-shelf tool can do it too. Buy.
- Does it need to happen when nobody is watching? On arrival, overnight, at volume, without a person triggering it. If yes, you need integration and autonomy. Build.
- If a competitor bought the identical tool tomorrow, would they get your result? If yes, it is not a differentiator — buy it cheaply and spend your build budget where the answer is no.
Most companies run this test and find three or four workflows in the "buy" column and one or two in the "build" column. That is the correct outcome, not a failure to commit.
6. The Sequence That Wastes the Least Money
The lowest-risk path is not choosing — it is ordering:
- Buy first, deliberately. Give staff an off-the-shelf assistant for a quarter. It costs little and generates something more valuable than any consultant's opinion: evidence of what people actually use it for.
- Watch for the workarounds. The workflow worth building is the one where people are pasting the same data in every day, or copying output into another system by hand. That copy-paste is the integration you should be paying an engineer to remove.
- Build one thing, narrowly. Pick the highest-volume workaround and automate exactly that. Prove it, then extend — the second build is far cheaper because the plumbing exists.
- Keep both. There is no point migrating general knowledge work into a custom system. Let the subscription do what it is good at.
What we tell clients on the first call: if you have not yet given your team an off-the-shelf assistant, do that before commissioning anything from us. It is cheaper than our discovery phase and it will tell you which workflow to build. We would rather start a project with that evidence in hand than guess with you.
The Bottom Line
Off-the-shelf AI makes your people faster. Custom AI removes the need for a person in that loop at all. Those are different purchases solving different problems, and the question was never which is better — it is which one this specific workflow needs.
Ready to work out what a build would actually cost? Start with our AI agent development cost breakdown or the RAG system pricing guide. If you are further back than that, the AI readiness checklist tells you what to prepare first, and how to choose an AI development company covers vetting a partner.