GBCORP BLOG · AI
Explainable AI Costs in 2026: Assessment, Programme and Audit Trail
Explaining an AI decision after a regulator asks is expensive. Designing the system to explain itself is not. This guide prices both routes and shows what a buyer should expect to pay for each stage in 2026.
Explainable AI means every automated decision can be traced to its inputs, its model version and a reason a person can read. For decisions that affect credit, hiring, insurance, health or access to services, that is becoming a legal expectation as well as a good practice: the EU AI Act phases in obligations for high-risk systems between 2025 and 2027, and similar expectations are spreading to other markets. The cost of meeting them depends mostly on one choice: building explanation in, or bolting it on.
What each stage costs in 2026
| Stage | What it covers | GBCORP |
|---|---|---|
| Readiness assessment | Inventory of models and decisions, risk tiering, gap analysis against the applicable rules | From $16,000 |
| Interpretability design | Glass-box models where possible; explanation methods where a black-box model is required | Part of the programme |
| Per-decision audit trail | Inputs, model version, confidence and explanation logged and searchable | Part of the programme |
| Bias and drift monitoring | Fairness metrics across groups, alerts when behaviour shifts | Part of the programme |
| Governance documentation | Model cards, data cards, risk assessments for compliance teams and auditors | Part of the programme |
| Full programme | All of the above for the models in scope | From $55,000 |
The readiness assessment is the sensible first purchase: it tells a business which of its models are high-risk, what is missing, and what the programme will cost before committing to it. The full programme price depends on how many models are in scope and how much of the logging and monitoring already exists.
Built in or bolted on
Retrofitting explanation onto a system already in production usually costs more than designing it in, for three reasons. The data needed to explain a past decision was often never logged, so it has to be reconstructed or the history accepted as unexplainable. The model may have been chosen for accuracy alone, and replacing it means retesting everything downstream. And the documentation has to be written by people reading code rather than by the people who made the design choices. A new system that plans for explanation from the start pays for logging and documentation once, as part of the build.
What drives the price up or down
| Factor | Lowers the cost | Raises the cost |
|---|---|---|
| Number of models in scope | One scoring model | A portfolio across departments |
| Risk tier | Minimal-risk uses needing disclosure | High-risk decisions needing the full apparatus |
| Existing logging | Inputs and versions already recorded | No decision history kept |
| Model type | Interpretable models | Large black-box models needing explanation layers |
| Review process | A human already reviews each decision | Fully automated decisions |
The cost of not explaining
The alternative is not free. Unexplainable decisions invite complaints that cannot be answered, audit findings that halt a product, and rework under deadline pressure. A business that cannot say why a customer was refused often has to reverse the decision anyway. Budgeting for explanation is budgeting to keep a model in use.
What the readiness assessment delivers
The assessment is a fixed piece of work with written outputs a business keeps. It lists every model and automated decision in use, including the ones inside bought software. It assigns each a risk tier with the reasoning written down. It records what is already logged, documented and reviewed, and what is missing. It sets out the changes each high-risk system needs, in order of urgency, with a cost for each. And it proposes where a person should review or override decisions. A compliance team can use the document on its own; the full programme is simply the most direct way to act on it.
Timeline and running costs
An assessment typically takes a few weeks, depending on how many systems are in scope and how quickly the people who own them are available. A programme then runs system by system, starting with the highest risk. After it, the running cost is modest but real: monitoring for bias and drift, retaining logs for the period the rules require, and updating documentation whenever a model changes. Budget for that as a line of its own, because explanation that stops being maintained stops being explanation.
Six questions to ask any provider
- Which of our models are high-risk, and why?
- Will you change the model, or add explanation around it?
- What exactly is logged for each decision, and for how long?
- How is bias measured, against which groups, and who is alerted?
- Which documents will our auditors receive?
- Where does a person review or override a decision?
Questions we hear first
Do all AI systems need explainability?
No. Minimal-risk uses mainly need disclosure. High-risk decisions about people need documentation, logging, oversight and explanation.
Does explainability reduce accuracy?
Sometimes slightly, often not at all. Interpretable models match black-box accuracy on many business problems, and explanation layers exist where they do not.
Can you explain a model you did not build?
Yes. The assessment maps it, and explanation, logging and monitoring are added around it where the model itself cannot change.
How long does a programme take?
The assessment takes a few weeks; the programme timeline depends on the number of models and the logging that exists.
Where to start
The GBCORP Explainable AI service starts with a readiness assessment from $16,000, with a full programme from $55,000. The GBCORP AI development and machine learning hub sets out the wider method, and the 2026 guide to explainable AI and the EU AI Act explains what the rules ask for.
GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Premium Business, and EV's & Solutions.
