Before Scaling AI: Five Human Checks for UAE Businesses

An illustrative hotel workflow shows how managers can assess responsibilities, staff readiness and service quality before expanding an AI trial.

By Mafalda Carvalho

Before approving a wider AI rollout, ask an employee to demonstrate an ordinary task, a mistake and an exception. A polished demonstration of the tool is useful. A demonstration of how the team will work with it tells a manager something different: whether people can complete the job when the answer is uncertain.

Consider a hypothetical UAE hotel trialling AI-assisted replies to routine guest enquiries. This is an illustrative workflow, not a client case or evidence of results. AI prepares a draft; an employee verifies it before release. The trial excludes autonomous booking changes and decisions about complaints or individual support needs.

The operational question is whether the hotel can maintain an accurate, useful service after counting the work of checking and correction. Five checks can help structure that decision.

1. Draw a boundary the team can use

The reservations manager should specify which enquiries enter the trial. A question about published breakfast hours may qualify. A request to guarantee an accessible room for a particular stay requires confirmation by the responsible team. A fluent answer must never turn an unverified possibility into a commitment.

The approved information should have an owner and a review date. The system may draft from that material, but current availability comes from the hotel’s authorised booking records. An outdated property description cannot establish that a room is available today.

The pilot should use an approved AI service with a clear boundary on data inputs. Staff should avoid entering unnecessary guest identifiers, booking references or sensitive details. They can check booking records separately without copying those records into the drafting tool.

Continue when employees consistently distinguish permitted enquiries from exceptions. Change the instructions when categories are ambiguous. Pause the affected use when the system repeatedly makes commitments outside its boundary. These are proposed management decisions, not legal thresholds.

2. Give the reviewer access and time

The reviewer should compare each factual commitment with its authoritative source. Breakfast times are checked against approved property information; a statement about a booking or availability is checked against the relevant current record. A source link that nobody opens does not establish accuracy.

For example, a guest asks, “What time is breakfast served?” The employee checks the applicable information and date before approving a reply. If the guest instead asks whether a room can accommodate a specific access requirement, the employee refers the request for individual confirmation rather than approving a generic description.

The supervisor owns reviewer allocation and cover. Review time belongs in the staffing plan and performance measures. Continue when checking is feasible and traceable. Change the workload or workflow when checks create a growing queue. Pause if employees are expected to approve statements they cannot verify.

3. Test judgement before widening access

Training should include deliberately imperfect drafts. Ask employees to spot an invented opening time, an unsupported room guarantee and a missing qualification. These are rehearsal materials, never messages sent to guests.

The training lead and reservations supervisor should assess whether employees can identify the problem and explain the next action. Attendance alone cannot show this. Where replies involve several languages, the hotel needs people capable of checking their meaning; polished wording does not establish that the commitment is accurate.

Continue when staff demonstrate the required judgement across the proposed enquiry types. Change training when errors cluster around one task. Pause access for an employee or use case that cannot yet be reviewed reliably, while providing support. Treating uncertainty as useful information makes this assessment more credible than rewarding confident approval.

4. Make escalation end in a decision

An escalation process should identify the person on duty who can resolve an exception and the approved response while the guest waits. Requests involving individual accessibility needs, complaints, sensitive circumstances and policy exceptions may need different responsible teams.

The duty manager should also own the fallback arrangement. If the tool or source system is unavailable, employees use the existing manual process. The trial should test that route, including a shift when its usual champion is absent.

Continue when exceptions reach someone authorised to act within the agreed service window. Change cover when requests repeatedly wait for an unavailable individual. Pause the affected workflow if an unresolved exception can pass through as an approved response. Logging a concern is only the beginning; someone must decide what happens next.

5. Compare the complete service

The operations manager should agree the comparison before reviewing results. Use comparable enquiries and record workload, complexity and staffing conditions. Keep initial training and setup costs visible, but separate from recurring handling time.

A simple pilot scorecard could compare the existing process with the trial using these measures:

• Total employee minutes per resolved enquiry, including preparation, review, correction and follow-up.
• Factual errors per 100 sampled replies, distinguishing incorrect details from unsupported commitments.
• Escalations per 100 eligible enquiries, with the reason and time to resolution.
• Staff use: the proportion of eligible enquiries handled through the trial, plus reasons for choosing the manual route.
• Service quality: response time, repeat contact about the same issue and complaints attributable to the reply.
• Coverage: the shifts, languages and enquiry types actually tested.

Apply the same assessment criteria to samples from both processes. Record errors caught before sending separately from errors that reach guests.

The denominator matters. A falling error count may simply reflect fewer enquiries. More escalations may indicate that staff are recognising exceptions correctly. Low use may reveal unsuitable inputs or insufficient review time. The scorecard should support investigation rather than turn every number into a target.

For illustration, a hotel might require every sampled factual commitment to be traceable to an approved current source, and pause an affected use after an unsupported availability guarantee. These are illustrative controls, not validated industry benchmarks. The hotel must set its own criteria according to consequences, evidence and applicable obligations.

Continue only when the overall result supports expansion under the tested conditions. Change a process that saves drafting time but increases correction or repeat contact. Pause when the evidence cannot establish whether service quality is being maintained.

The NIST AI Risk Management Framework 1.0 provides a general basis for assigning roles, understanding context, evaluating human oversight and monitoring risk. It is voluntary guidance, not a UAE compliance certificate or a substitute for relevant local requirements. The five checks and hotel scorecard above are my practical recommendations for this hypothetical workflow.

The decision at the end of a trial need not be a simple yes or no. A hotel may expand one routine category, keep another in testing and exclude requests requiring individual judgement. Recording those boundaries makes the next decision clearer.

For me, the useful question is whether an employee can explain what they may approve, what they must check and who will help when the answer is uncertain. That is something a business can test before it asks more people to rely on the system.

About the author

Mafalda Carvalho is a leadership practitioner, author and founder of Prestige Care Consulting, based in Greater Manchester. Her work focuses on people, culture and the practical implementation of organisational change.

Disclosure

The author is founder of Prestige Care Consulting, which offers leadership and organisational change services. This article contains a hypothetical example and does not report a client deployment.

Reference

Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, NIST AI 100-1.

Also by Mafalda Carvalho: Why Organisational Change Can Fail in UAE Companies: Seven Human Factors Leaders Overlook.

Featured photograph: Andri Wijayanto / Pexels. Author photograph: Courtesy of Mafalda Carvalho.