Skip to content
Senior leaders reviewing a one‑page decision artefact comparing rules/process redesign with an AI pilot, with an acceptance ledger visible.

Insights

When not to use AI: a practical guide for senior transformation leaders

Antares Consultancy

AI is not always the answer. For transformation leaders charged with delivering regulated, mission‑critical or high‑value services the right technical choice is the one that reliably reduces risk and cost, and makes outcomes auditable and reversible.

Too often organisations choose AI because it looks modern, promises efficiency or because suppliers position it as the default. That choice can become expensive and fragile if the problem would be better solved by a simple rule, a redesign of work, or clearer ownership.

This note explains five practical situations where you should pause on AI, the indicators to look for, the alternatives that typically work better, and the procurement and acceptance actions senior leaders should require before any move towards AI.

1. When data is poor, sparse or sensitive

AI depends on data. If the available data are low quality, biased, highly sensitive or of uncertain provenance, an AI model will inherit those flaws and amplify them. That creates privacy, fairness and safety risks—and often costs more to remediate than building a non‑AI control.

Indicators: inconsistent records, large volumes of manual corrections, undocumented transformations in migration extracts, or data that cannot be processed without breaching statutory limits (for example where the Data (Use and Access) Act 2025 or the UK GDPR impose constraints).

Better alternatives: a simple deterministic rule‑set (if/then), a validated deterministic algorithm (for example a rules engine), or a process redesign that reduces reliance on contested fields. These options are cheaper to test, explain and audit, and are reversible.

Practical leader action: require a data manifest, lineage map and DPIA before any AI procurement. If the manifest shows substantial cleaning, missing fields or high human correction rates, insist on a non‑AI pilot first or a recorded human in‑the‑loop with traceability. The ICO’s guidance on automated decision‑making and the Data and AI ethics framework emphasise these controls for public bodies.

2. When decisions are high‑consequence or contestable

Where decisions materially affect people—eligibility for benefits, clinical triage, law enforcement, or significant financial outcomes—errors have real harms and regulatory scrutiny. AI’s probabilistic nature makes it harder to guarantee consistent, explainable outcomes at scale.

Indicators: decisions that require legal accountability, where explainability or meaningful human review is a statutory requirement, or where a wrong outcome creates downstream systemic risk.

Better alternatives: retain human judgement supported by decision aids, or codified checklists that force explicit rationale capture. If automation is needed, prefer white‑boxed deterministic algorithms with clear auditable traces and a defined human escalation path.

Practical leader action: treat high‑consequence use as an operational permit‑to‑operate decision. Demand an executable acceptance pack that demonstrates how decisions will be explained, contested and reversed, drawing on ICO and Government guidance about explaining AI decisions. Require staged pilots with staffed rehearsals and contractual reversibility tests before any wider deployment.

3. When no clear ownership or governance exists

AI projects thrive or fail depending on product ownership, operating model and the presence of accountable owners for data quality, model outcomes and lifecycle operations. Without clear ownership, AI produces brittle systems that become orphaned operational liabilities.

Indicators: no single accountable role for outcomes, split responsibilities between procurement, analytics and operations, or an absence of runbooks and defined early‑life support.

Better alternatives: postpone AI and invest in organisational clarity—define the owner, runbooks, acceptance tests and funded stabilisation. Often a process redesign or incremental automation with well‑defined SLOs delivers faster benefits with lower risk.

Practical leader action: before approving AI procurement, require a named service owner, staffed runbooks, a documented acceptance ledger linking benefits to testable outcomes, and a funded stabilisation window in the budget. The Green Book (2026) expects appraisal to be explicit about deliverability and operational controls; embed these into the business case.

4. When volume is low or cost‑benefit doesn’t stack up

AI is attractive for repetitive, high‑volume tasks. Where volumes are low, transactors are specialised, or exceptions dominate, the overhead of model training, monitoring, governance and retraining frequently outweighs the efficiency gains.

Indicators: few transactions per day/week, high exception rates, specialist knowledge required to adjudicate cases, or small user bases.

Better alternatives: manual processing, specialist adjudicators, or light automation (macros, RPA or rules engines). These approaches keep cost and complexity low and are straightforward to scale later if volume grows.

Practical leader action: require a simple cost‑benefit threshold in procurement: estimate total cost of ownership for model lifecycle (data, retraining, monitoring, compliance) and compare it to alternatives. If the projected annual run rate for AI operations exceeds simple automation or human processing with coaching, choose the latter.

5. When AI adds cost, opacity or regulatory uncertainty without improving outcomes

Not every problem benefits from probabilistic inference. If AI introduces external dependencies (proprietary models, opaque supplier stacks), additional cyber and data‑protection controls, or uncertainty about long‑term portability, it can reduce resilience and increase exit costs.

Indicators: reliance on black‑box models, unclear model provenance, tight supplier lock‑in, or when regulatory frameworks create material obligations (audits, logging, DPIAs) that the organisation is not equipped to manage.

Better alternatives: prefer open, auditable algorithms, a rules‑first approach, or contract clauses that mandate model portability and reversibility. If AI is still attractive, insist on portability tests, escrow for models and data, and independent verification during acceptance.

Practical leader action: during procurement demand model portability clauses, reversibility tests and a verified acceptance ledger. Where suppliers claim compliance, require reproducible tests and a supplier‑witnessed recovery rehearsal as part of acceptance evidence.

When not to use AI: a practical guide for senior transformation leaders: editorial image for 5. When AI adds cost, opacity or regulatory uncertainty without improving outcomes

Practical checklist for leaders deciding whether to use AI

1) Data readiness: demand a manifest, lineage map and DPIA; if data are poor or sensitive prefer non‑AI controls.

2) Consequence test: if harm is material, mandate human‑in‑the‑loop, explainability and staffed rehearsals.

3) Ownership and runbooks: require a named service owner, early‑life staffing and a funded stabilisation window; embed these in the business case per Green Book expectations.

4) Volume and economics: compare total cost of ownership for an AI solution versus rules/automation/manual processing before procurement.

5) Procurement and acceptance: require reversibility clauses, model‑and‑data escrow (where appropriate), independent verification tests and an executable acceptance pack before any go‑live decision.

How to present a conservative, evidence‑led option to a board

Boards and SROs need decision artefacts that translate technical risk into conditional approvals. Present the alternatives as testable options: a non‑AI pilot with measurable outcomes; a set of deterministic rules and their expected impact; or a process redesign with a simple automation pilot. For each option provide the acceptance ledger: dataset manifest, test cases, owners, runbook and staged reversal plan.

This approach aligns with the Green Book requirement that appraisal addresses deliverability and operational risk, and with ICO expectations on deploying automation safely. It also makes the decision reversible if a future evidence base supports AI adoption.

Choose the tool that reduces risk and increases certainty. Sometimes the best use of technology is to avoid complexity until you can manage it.

Antares recommended actions

These are Antares's recommended first actions for organisations turning the issues in this article into practical governance and delivery.

  1. Require a data manifest, DPIA and lineage map before any AI procurement.
  2. Treat high‑consequence decisions as operational permits‑to‑operate with staffed rehearsals and explainability tests.
  3. Prefer deterministic rules, simple automation or process redesign when data are poor, volumes low, or ownership weak.
  4. Embed named owners, runbooks and funded stabilisation into the business case as the Green Book expects.
  5. Demand contractual reversibility, model/data escrow and independent verification as part of acceptance evidence.

Our insights are provided for general information only and reflect the position at the date of publication. They do not constitute legal, financial, regulatory, cybersecurity or other advice tailored to your circumstances and should not be relied upon as a substitute for appropriate professional advice.

While we take reasonable care over our content, we do not guarantee that it is complete, accurate or current. Reading our insights does not create a client relationship with Antares Consultancy. To the fullest extent permitted by law, we accept no liability for decisions made or losses arising from reliance on this content. External links are provided for convenience and do not imply endorsement.

If you'd like to discuss your own transformation, we'd be pleased to start the conversation.