Scaling AI with Confidence

Start with the business problem, not the technology

The first question is not, “Where can we use AI?”

It is:

Which business problems are important enough to deserve capital and management attention?

For each opportunity, the case keeps five questions separate:

1. Value — Is the outcome material enough to justify investment?
2. AI suitability — Does AI add capability beyond simpler automation, analytics or process redesign?
3. Readiness — Can the organisation execute this use case safely, reliably and repeatedly now?
4. Consequence — What happens if the AI is wrong, over-trusted, biased, unavailable or misused?
5. Evidence strength — How much is demonstrated, and how much remains assumption?

Keeping those questions separate matters. A use case can be commercially attractive but not ready. It can be technically impressive but strategically weak. It can also be valuable and feasible while carrying consequences that justify stronger assurance before scale.

Five opportunities, five different decisions

In the worked case, Fiction Group considers five opportunities.

1. Enterprise Knowledge Assistant — Pilot

The opportunity is to reduce avoidable search and reconciliation effort while improving access to current, authorised information.

The likely constraint is not model capability. It is knowledge governance: authoritative sources, permissions, version control, provenance and conflict handling.

2. Predictive Asset Intervention — Validate data first

The value hypothesis is to direct limited inspection and maintenance capacity towards assets with higher evidence-based intervention need.

The first investment is therefore not a model build. It is a data-suitability test: whether failures, interventions and outcomes are recorded consistently enough to support prediction.

3. AI-Assisted Field Risk Assessment — Governed pilot

The opportunity is to help competent operatives navigate controlled safety information and produce more relevant, timely site-specific assessments.

The boundary is explicit:

AI may retrieve, prompt, structure, suggest and cite. The competent operative observes, judges, escalates, stops work and decides whether work proceeds.

AI can support access to relevant information and human judgement. It does not confer competence or determine that work is safe to proceed.

4. Customer-Service AI Agent — Bounded experiment

Autonomy should be earned.

The case starts with recommendation and drafting rather than broad autonomous action.

Autonomy is treated as a permission to be earned through evidence. Movement to higher autonomy requires defined action limits, identity controls, logs, exception handling and evidence that risk remains within tolerance.

5. AI Recruitment Screening — Defer Wave 1

The business problem is legitimate, but other opportunities currently present a stronger value-to-assurance ratio.

For UK deployment, the Information Commissioner’s Office has highlighted both the efficiency benefits and the fairness, transparency and data-protection safeguards relevant to automated recruitment. For EU operations, recruitment and candidate-evaluation systems appear in Annex III of the EU AI Act and sit within its Article 6 high-risk classification framework where the relevant conditions apply.

The decision is therefore not “AI recruitment is unacceptable.” It is that this is not currently the best first-wave use of scarce investment and governance capacity.

Readiness is not one score. Find the constraint.

A single maturity percentage can conceal the one weakness that actually prevents scale.

Fiction Group therefore assesses readiness across strategy, data, infrastructure, governance, talent and culture, then asks:

What is the binding constraint for this use case?

For the knowledge assistant, it may be source ownership and permissions. For predictive asset intervention, historical data suitability. For a customer agent, authority limits and escalation. For field risk assessment, controlled safety knowledge and safe human-AI interaction.

The practical output of a readiness assessment is a sequence: identify the constraint, address it, then test whether another constraint becomes limiting.

Governance intensity should reflect consequence and autonomy

Fiction Group uses internal governance tiers to determine the level of approval, assurance, human oversight and monitoring required for each use case.

These are internal management controls, not legal classifications. Where external legal or regulatory requirements apply, those requirements take precedence.

The principle is straightforward: the greater the potential impact of error, and the more independently the system can act, the stronger the assurance should be.

The governance model is informed by recognised standards and frameworks including:

• ISO/IEC 42001:2023, which specifies requirements for establishing, implementing, maintaining and continually improving an AI management system;
• ISO/IEC 42005:2025, which provides guidance for AI system impact assessment;
• NIST AI RMF 1.0, a voluntary, non-sector-specific framework for managing AI risk through Govern, Map, Measure and Manage; and
• NIST AI 600-1, the Generative AI Profile and companion resource to the AI RMF.

These references inform management design; they do not replace applicable law or sector-specific obligations.

High-consequence AI requires meaningful human authority

This matters particularly in the field risk-assessment case.

The Health and Safety Executive’s June 2026 policy statement says that health and safety law applies regardless of the technology used, including workplace AI. HSE also states that AI uses affecting workplace health and safety require risk assessment and appropriate controls to reduce risk so far as is reasonably practicable, including consideration of cyber-security threats.

HSE’s 2026/27 programme also includes research into cognitive offloading and cognitive resilience in AI-augmented safety decision-making.

That does not prove a negative effect in the fictional case. It does mean that human reliance is a legitimate variable to test rather than assume away.

A high acceptance rate is therefore not automatically a success metric. It may indicate excellent recommendations — or over-reliance. Human challenge, escalation and independent judgement need to be evaluated alongside technical performance.

A pilot must prove more than technical performance

Fiction Group requires pilots to produce evidence across four dimensions:

• Technical: Does the system perform sufficiently well?
• Operational: Does the redesigned workflow improve?
• Human: Do people use, challenge and escalate appropriately?
• Risk: Are controls effective and residual risk acceptable?

The case also uses an internal evidence maturity scale for capital allocation:

• E0 — hypothesis;
• E1 — problem evidenced;
• E2 — feasibility evidenced;
• E3 — pilot result;
• E4 — repeatability across contexts;
• E5 — sustained production evidence.

The key point is:

Capital follows evidence. A successful pilot is evidence of a pilot result — not enterprise-scale evidence.

A successful pilot is not enterprise-scale evidence.

Separate savings from capacity value

AI business cases can become less credible when all time saved is presented as cash savings.

The worked case separates hard financial value, capacity value, operational value, risk/control value and strategic option value.

For the illustrative knowledge-assistant calculation, gross addressable time is converted into capacity value and then reduced by an explicit realisation assumption before recurring operating cost is considered.

The purpose is not to manufacture a precise ROI number. It is to make assumptions visible, challengeable and replaceable with measured organisational evidence.

What the Board should see

A board-level AI portfolio view should cover more than model performance.

It should show:

• value realised against the approved case;
• adoption inside the redesigned workflow;
• reliability and material errors;
• risk events, escalations and exceptions;
• governance ownership and overdue actions; and
• portfolio health — including pilots that have been redesigned or stopped.

A credible AI portfolio must be able to stop weak initiatives, not only launch new ones.

Adoption is not login count.

A technically strong tool can fail to create value if the surrounding workflow does not change. Fiction Group therefore measures whether work actually moves through the redesigned process - not simply whether employees opened the application.

The recommendation

The fictional Board is asked to approve:

• the first-wave portfolio and an evidence-led funding model;
• a federated operating model with central guardrails;
• an AI Governance & Assurance capability;
• an AI Decision Record and named accountable owner for material AI systems; and
• staged capital release only when value, reliability, adoption, residual risk and accountable ownership meet agreed thresholds.

Scale is a consequence of evidence — not the objective of every pilot.

That is the core of the case.

AI transformation is not simply about introducing more AI. It is about making better decisions about where AI belongs, what authority it should have, what must be proven, and when further investment is justified.

Download the full Executive Brief

A Worked Executive Case in AI Transformation, Governance and Enterprise Value

10-page executive brief · PDF · No sign-up required

About the author

Ajadi Ahmed

Chartered Safety & Health Practitioner (CMIOSH) - OSHCR Registered Consultant - BSc, PG Dip. Environment, Health & Safety

Founder, IntelliSafely and VeriPrem

Primary references

1. Health and Safety Executive (HSE), HSE’s regulatory approach to Artificial Intelligence (AI), 12 June 2026. https://press.hse.gov.uk/2026/06/12/hses-regulatory-approach-to-artificial-intelligence-ai/ (opens in a new tab)
2. Information Commissioner’s Office (ICO), Automated decisions can streamline the hiring process - with the right safeguards in place, 31 March 2026. https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/03/automated-decisions-can-streamline-the-hiring-process-with-the-right-safeguards-in-place/ (opens in a new tab)
3. Information Commissioner’s Office (ICO), Recruitment rewired, 2026. https://ico.org.uk/about-the-ico/what-we-do/recruitment-rewired/ (opens in a new tab)
4. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 6 and Annex III. https://eur-lex.europa.eu/eli/reg/2024/1689 (opens in a new tab)
5. ISO/IEC 42001:2023. https://www.iso.org/standard/42001 (opens in a new tab)
6. ISO/IEC 42005:2025. https://www.iso.org/standard/42005 (opens in a new tab)
7. NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf (opens in a new tab)
8. NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf (opens in a new tab)