Telecommunications
Network operations, service assurance, revenue assurance and customer operations, with the resilience obligations that come with critical infrastructure.
Telecommunications detail →Enterprise AI investment intelligence
Most AI proposals should not be approved as written.
DUN-AI is an assessment platform for organisations deciding where to invest in AI — and, just as often, where not to. It does not assume AI is the answer. Every use case is compared against thirteen alternatives, including doing nothing, redesigning the process, and using software you already own.
Built for telecommunications, banking, energy, government and healthcare, with regional content for the UK, the EU/EEA, the GCC and the US.
The decision usually skips the comparison
An AI business case is typically written after the technology has been chosen. The alternatives are named but not scored, the benefit is a single number with no stated assumptions, and the regulatory position is a paragraph written by someone who is not accountable for it.
No counterfactual
If the alternatives were never scored on the same basis, the business case is not a comparison. It is a justification.
False precision
A benefit stated to the nearest pound, derived from an assumption nobody has evidenced, reads as a measurement rather than an estimate.
Reasoning not recorded
Twelve months later, the question is why the decision was made. If the inputs and the logic were not captured, that question has no answer.
Fifteen possible outcomes. Six of them involve no new AI.
An assessment resolves to exactly one outcome from a fixed set. The set is closed, so a result cannot be softened into something more agreeable, and recommending against AI is a first-class result rather than a failure to conclude.
| Outcome | What it means |
|---|---|
| AI investment approved | The case holds, the alternatives were weaker, and no gate is triggered. |
| Approved with conditions | Approved, but specific controls or evidence must be in place first. |
| Conditional pilot recommended | The uncertainty is best resolved by a bounded, measurable pilot. |
| Restricted pilot only | A pilot may proceed under explicit constraints on scope and autonomy. |
| Buy an existing AI solution | The capability is commoditised; building it would not be justified. |
| Build a custom AI solution | The requirement is genuinely specific and the organisation can sustain it. |
| Use a hybrid human and AI model | Value comes from assisting people rather than replacing the decision. |
| Use existing platform capability No new AI | Software already licensed can do this. No new investment is warranted. |
| Use conventional automation No new AI | The task is deterministic. Rules or scripting are cheaper and auditable. |
| Retain the human-led process No new AI | The process should stay with people, at least for now. |
| Process redesign first No new AI | Automating the current process would encode its problems. Fix it first. |
| Data foundation required first No new AI | The data does not yet support the outcome. Address that before investing. |
| Defer pending regulatory clarity No new AI | The obligations are unsettled enough that committing now carries real risk. |
| Reject investment No new AI | The case does not stand, or a blocking gate applies. |
| Further evidence required | Too little is known to conclude responsibly. The report says exactly what is missing. |
The verdict is calculated, not written
Scores, gates, the counterfactual comparison and the confidence calculation are deterministic. The same answers produce the same verdict every time, and the path from input to conclusion is recorded step by step.
Establish the problem before the solution
The assessment opens on the business problem, the current process, what success would look like and how it would be measured. Technology questions come later.
Ask only what applies
Questions are selected by industry, region, deployment context and the answers already given. Each one explains why it is being asked and what it affects.
Score eleven dimensions
Problem clarity, data readiness, technical feasibility, organisational readiness, regulatory exposure, value credibility and five more — each with a stated rationale and its own confidence, never a single opaque number.
Evaluate the gates
Legal, safety, ethical, data and operational gates are checked independently of the scores. A blocking gate overrides the result regardless of how well it scored.
Compare thirteen alternatives
AI is one option among thirteen. Each is scored on capability fit, cost, time to value, risk and reversibility, and the report explains why each was not selected.
Model the value as a range
Costs and benefits are expressed as ranges with the assumptions listed, marked as evidenced or estimated, and tested for sensitivity. Never a single figure.
Calculate confidence, then conclude
Confidence reflects how much was answered, how much was evidenced and how consistent the answers were. Thin evidence produces a request for evidence, not a confident verdict.
Hard gates outrank scores
A use case that scores well on every dimension is still refused if it crosses a blocking gate — an unlawful processing basis, a prohibited practice, safety-critical autonomy without human control, or an unacceptable failure consequence. An attractive score cannot buy its way past a gate, and the report says which gate closed the decision.
Where a language model is used, and where it is not
DUN-AI can use a language model to draft narrative summaries and suggest clearer phrasing. It is never permitted to determine an outcome. The separation is structural rather than procedural: the decision engines cannot reach the AI gateway at all, and a test reads the module imports to prove it.
Every report states whether a language model influenced the verdict. The answer is always no, because the code makes any other answer impossible.
Read the responsible AI position →Never determined by a model
- The final verdict
- Any financial calculation
- Regulatory classification
- Whether a hard blocker applies
- Whether a risk is acceptable
May be model-assisted
- Narrative phrasing in a report
- Summarising evidence the user supplied
- Suggesting a clearer problem statement
Five industries, assessed on their own terms
Compare all five →Each industry brings its own questions, gates, risks and value patterns. A regional overlay then adds the obligations that apply where the organisation operates.
Banking
Credit, financial crime, conduct and operational resilience, where model governance and explainability decide what is deployable.
Banking detail →Energy
Grid, generation and asset management, where safety-critical control and operational technology boundaries constrain what autonomy is permissible.
Energy detail →Government
Public service delivery and casework, where transparency, contestability and the duty to explain a decision to a citizen are non-negotiable.
Government detail →Healthcare
Clinical and operational settings, where anything touching a clinical decision brings device regulation and clinical safety obligations with it.
Healthcare detail →What DUN-AI does not do
A tool that claims to remove judgement from an investment decision is overselling itself. These limits are published because a buyer should know them before a procurement conversation, not after.
- It does not give legal, regulatory, financial or clinical advice.
- It does not replace a lawyer, a data protection officer, a clinical safety officer or an accountable executive.
- It does not promise savings, returns, compliance or delivery success.
- It does not audit your systems, inspect your data or verify what you tell it.
- It does not convert between currencies, because an invented exchange rate would make a comparison look sounder than it is.
- It does not hide a weak evidence base behind a confident-looking number.
Start with the question, not the technology
An assessment begins by establishing what problem is being solved and what would count as success. If that cannot be answered, no technology decision should be made yet — and DUN-AI will say so.