Build It, or Buy Seats?
Per-seat AI licensing is simple and predictable until your headcount grows, at which point a custom deployment that looked expensive starts looking cheap. This compares both over three years and tells you where the crossover sits for your organisation.
The crossover figure — how many users make building cheaper — is usually the number that settles the discussion internally.
Build vs Buy — Three-Year Comparison
A positive saving favours building. Near zero, buy — it is reversible.
Your negotiated AUD price, including any prerequisite licence.
Including permissions, evaluation and security review.
Tokens plus infrastructure — use the LLM cost calculator.
Positive favours building. Negative favours commercial licences.
Cost comparison only. It excludes change management, training, engineering opportunity cost and delivery risk — all of which favour buying. Data sovereignty and regulatory constraints may justify building regardless of the result.
Reading the Comparison
Cost is only one of three axes in this decision, and it is frequently not the deciding one. The calculator handles the money; the other two need judgement.
Per-seat scales linearly, custom does not
Licence cost rises in direct proportion to headcount forever. A custom build has a large fixed cost and a much smaller marginal cost per additional user, which is why the comparison flips at a certain scale rather than favouring one approach universally.
Buying is dramatically faster
Commercial licences can be deployed in days. A custom build takes months, and during those months you have paid for capability you do not yet have. For urgent needs, buying now and building later is frequently the right sequence rather than a compromise.
Some requirements are not about cost at all
Data sovereignty, sector-specific regulatory constraints, integration with systems no vendor supports, or genuine competitive differentiation can each justify building at any scale. If one of those applies, the calculator is informative rather than decisive.
What the Comparison Includes
Three years is the right horizon: long enough for a build to amortise, short enough that the technology landscape remains recognisable.
Commercial licence path
Users multiplied by monthly per-seat cost across thirty-six months. Simple, predictable, and it never stops.
Custom build path
One-off build cost plus monthly running costs across thirty-six months, plus ongoing maintenance expressed as a share of the build.
Difference and crossover
The three-year saving or premium, and the user count at which the two paths cost the same.
What is deliberately excluded
Internal change management, training and the opportunity cost of engineering time spent building rather than on something else.
When Each Approach Is Genuinely Right
The cost comparison rarely settles this alone. These are the situations where the answer is clear regardless of what the numbers say.
Buy when the need is broad and generic
If what you need is general assistance with documents, email and meetings across a whole organisation, commercial products already do this well and are improving continuously without any effort from you. Building a worse version of a well-funded commercial product is a common and expensive mistake.
- General productivity assistance across a broad user base
- Requirements that match what commercial products already do well
- No unusual data residency or regulatory constraint
- Need for capability in weeks rather than months
Build when the value is in your own data and processes
Custom deployments earn their cost when the application depends on your proprietary data, your specific workflows, or systems no commercial vendor integrates with. That is where a general product cannot reach and where the resulting capability is genuinely yours.
- The application must reason over proprietary internal data
- Deep integration with systems no vendor supports
- The workflow is specific to your organisation, not generic
- The capability itself is a competitive differentiator
Build when sovereignty or regulation requires it
Government, defence, health and parts of financial services frequently face requirements that commercial multi-tenant products cannot satisfy — data residency, isolation, auditability or specific certification. In these cases the decision is made by the constraint rather than by the spreadsheet.
- Data must remain in Australia or within a specific environment
- Sector obligations preclude multi-tenant processing
- Full auditability of model behaviour and data flows is required
- Contractual obligations to clients restrict where data may go
The hybrid path most organisations end up on
Commercial licences for broad general productivity, plus one or two custom applications where proprietary data creates real value. This is what most mature organisations arrive at, and choosing it deliberately is faster than arriving at it after two years of arguing.
- Commercial licences for general knowledge work across the organisation
- Custom builds for the two or three genuinely differentiated use cases
- Avoids building a worse version of a well-funded commercial product
- Avoids paying per seat for something only a small team needs
Next Steps
LLM Cost Calculator
Model the running cost of the custom path in detail, per query and per user.
Model running costs →Private LLM Readiness Assessment
Check whether your organisation has the foundations to build successfully.
Assess readiness →Microsoft Copilot Alternative
How a private deployment compares with the main commercial option in Australia.
See the comparison →Frequently Asked Questions
Use the actual quoted price for the specific product and tier you are considering, including any enterprise agreement discount you have negotiated, and remember to convert to Australian dollars if the list price is quoted in US dollars. Also check whether the AI capability requires an underlying licence you do not already hold, because that prerequisite is frequently omitted from the headline figure and can substantially change the comparison. Enterprise discounts at volume can be significant, so use your real negotiated number rather than list pricing.
Take whatever figure you first think of and interrogate the assumptions behind it. A production-grade retrieval application for an organisation includes document ingestion and processing, retrieval and re-ranking, the application interface, authentication and permissions, evaluation and testing, monitoring, and security review. Permissions are the component most consistently underestimated — ensuring users can only retrieve documents they are already entitled to see is often the hardest part of an enterprise deployment, and skipping it creates a serious information disclosure risk.
Fifteen to twenty-five per cent of the original build cost per year is a reasonable planning range for a production AI application, which is somewhat higher than conventional software. The additional burden comes from model deprecations forcing migration, prompt and retrieval behaviour drifting as your document corpus changes, and evaluation needing periodic re-running to confirm quality has not degraded. Teams that budget nothing for maintenance end up with an application that quietly degrades until someone declares the project a failure.
Not on its own. The crossover tells you the point at which the two paths cost the same over three years, which is genuinely useful for a budget discussion but ignores speed, risk and capability. Buying delivers value in days; building delivers it in months, and during those months you have the cost without the benefit. If your user count sits near the crossover, the non-financial factors should decide — and near the crossover, buying is usually the lower-risk choice because it is reversible.
Yes, and for most organisations this is the sensible sequence rather than a compromise. Commercial licences let people start working with the technology immediately, which surfaces the genuinely valuable use cases far more reliably than a planning exercise does. After six months of real usage you will know which two or three applications justify a custom build, and you will specify them far better than you could have upfront. The main thing to avoid is signing a long licence term that prevents you reducing seats once custom applications take over part of the workload.
Several things, all of which push in the same direction — against building. Internal change management and training are excluded, and these are significant for either path but larger for a custom build because there is no vendor-provided material. The opportunity cost of engineering time is excluded, which matters because engineers building an AI application are not building something else. Risk is excluded: a custom build might not work, whereas commercial licences demonstrably do. Treat a marginal calculated advantage for building as an argument for buying.
Want a Straight Answer?
Tell us your user count, your use case and your constraints. We will tell you honestly whether to build — and we frequently tell organisations to buy.