What does custom AI software actually cost?

The useful answer is not a universal price. It is a way to separate a contained experiment from a dependable system that can safely participate in your business.

Start with the job, not the model

“Build us an AI agent” is not yet a project definition. A useful brief identifies a job: review an incoming document, prepare a recommendation, answer from controlled knowledge, or move a case through a workflow. Cost follows the number of systems touched, the consequence of mistakes and the amount of ambiguity the system must handle.

A narrow assistant working from approved documents can be comparatively contained. An agent that writes to core systems, communicates with customers or makes consequential recommendations requires more engineering around the model.

Four bands of investment

1. Discovery and feasibility

This tests whether the proposed workflow is technically and commercially credible. It maps data, risks, edge cases and a measurable definition of success. The output may be a prototype, but its real value is evidence for a build decision.

2. A contained workflow pilot

A pilot handles one clear workflow for a small user group, usually with a human approving important outputs. It proves whether the system saves time or improves quality in real conditions.

3. A production system

Production adds identity, permissions, integrations, monitoring, recovery, evaluation and a usable interface. These are not peripheral costs. They are what turn an impressive demonstration into software a team can depend on.

4. A business-critical platform

When AI sits inside a core operation, expect deeper security, auditability, data governance, service levels and ongoing optimisation. The model may remain a small part of the total system.

The factors that change the estimate

• Data readiness: clean, accessible information reduces discovery and integration work.
• Autonomy: suggesting is cheaper and safer than acting without approval.
• Integration depth: each source or destination introduces permissions, failure states and maintenance.
• Evaluation: subjective tasks need carefully designed examples and review criteria.
• Operational risk: regulated or high-consequence decisions need stronger controls.

How to buy responsibly

Fund the uncertainty in stages. First prove that the job can be performed to an agreed standard. Then prove that users obtain real value. Only then harden and scale the system. Ask every proposal to distinguish discovery, product engineering, model usage and continuing operations.

The right first question is therefore not “What does AI cost?” It is “What valuable job can we define tightly enough to test?” That question usually produces a smaller, safer and more commercially honest starting point.