What you should take away
- Per-resolution vendor pricing can be excellent at low volume and punitive at scale, model the crossover explicitly.
- Buying does not remove model risk, evidence or complaints obligations; it relocates some of the work and adds oversight of a third party.
- The expensive part of building is rarely the agent; it is integration, evaluation and the permanent operating team.
- Most institutions end up hybrid: bought platform, built orchestration and controls.
Build-versus-buy papers in financial services usually compare the wrong pair of numbers: an annual licence fee against a one-off delivery estimate. Both are incomplete. The licence excludes your integration, evaluation and oversight; the build estimate excludes the operating team that will still be there in year three.
The honest comparison table
| Cost line | Buy | Build |
|---|---|---|
| Model inference | Usually inside vendor pricing, opaque | Direct, visible, negotiable |
| Retrieval and data | Partly included; your data preparation remains yours | Yours entirely |
| Integration to core systems | Yours | Yours |
| Evaluation and quality assurance | Yours to own, whatever the vendor provides | Yours |
| Model risk documentation | Yours, plus third-party oversight | Yours |
| Ongoing engineering and operations | Lower, not zero | Permanent team |
| Exit and portability | Contractual risk | Your code, your risk |
Both columns carry your obligations. Outsourcing an activity does not outsource accountability for it.
Where the crossover actually sits
Per-outcome vendor pricing is attractive until volume scales, at which point a fixed-cost build amortises better. The crossover is a function of volume, the vendor's unit price, and how much of the build's cost is genuinely fixed. Model it rather than debating it.
Illustrative annual cost at a £0.90 per-resolution vendor price
Illustrative arithmetic on a fixed-plus-variable build shape. Use your own vendor quote and your own fixed-cost estimate; the point is that the answer is volume-dependent.
Interactive model
Your in-house cost per outcome, at your volume
Set volume and your in-house cost lines to produce the number you should hold a vendor quote against.
Total cost per successful outcome
£1.20
AI-attempted workflow cost
£6.03m
Successful outcomes per year
5,040,000
Modelled annual operating difference
£16.37m
Where the cost sits
- Human escalationCalculated£4.48m74%Calculated
- Implementation, annualisedCalculated£500k8%Calculated
- Model inferenceCalculated£420k7%Calculated
- Retrieval & dataCalculated£180k3%Calculated
- Failures & retriesCalculated£180k3%Calculated
- Platform, evaluation & monitoringCalculated£150k2%Calculated
- Tools & APIsCalculated£120k2%Calculated
Largest modelled component: Human escalation (£4.48m). Outcome measured per successful autonomous resolution.
Illustrative assumptions. Compare the resulting cost per outcome with a vendor's per-resolution price on the same definition of resolution, vendors rarely use yours.
Non-financial factors that legitimately override the arithmetic
- Time to evidence: a bought platform may let you measure real resolution rates months earlier, which is worth real money in decision quality.
- Concentration and exit risk: third-party dependency for a client-facing regulated process attracts supervisory attention.
- Control over evaluation: if you cannot run your own eval set against the vendor's system, you cannot govern it.
- Talent reality: a build you cannot staff for three years is not cheaper; it is a different kind of expensive.
Apply this to your own workload
The figures above are illustrative assumptions. A BillingEngine assessment replaces them with your company's numbers and shows which assumptions decide the answer.