The schedule below is a planning framework, not a guarantee that a pilot can be completed in 90 days. Procurement, integrations, data approvals, language testing, and operational readiness may take longer. Adjust the phases to your organization’s risk, resources, and decision process.

Before day one: select a bounded use case

Choose one repeated customer-service task where the inputs, expected outcome, and safe boundaries can be described. Examples may include drafting an agent summary, retrieving an approved knowledge answer, or routing a clearly defined request. Avoid starting with a broad goal such as “automate the contact centre.”

  • Name the business owner and accountable operational lead.
  • Write the customer problem and the intended improvement in one sentence.
  • Define what the AI may access, suggest, or do—and what requires a human.
  • Identify language and channel requirements, including relevant Arabic varieties.
  • Establish a baseline from the current process before changing it.

Days 1–30: prepare the workflow and safeguards

Map the interaction from customer entry to resolution. Confirm the approved knowledge sources, system integrations, user permissions, data flows, retention settings, and human escalation path. Test representative cases offline, including confusing and out-of-scope requests.

Train participating employees on system limits, how to verify suggestions, how to correct records, and how to report incidents. Agree on pilot entry criteria and stop conditions before any customer-facing launch.

Days 31–60: run a limited, supervised trial

Start with a small, prepared team or a controlled subset of interactions. Keep human review in place for generated answers or actions with meaningful customer impact. Sample outputs regularly and investigate errors rather than hiding them in an average score.

  • Track whether the AI used current and approved information.
  • Record misunderstandings, failed integrations, unsafe suggestions, and unnecessary escalations.
  • Gather agent feedback about usefulness, correction time, and workflow friction.
  • Compare a suitable set of outcomes with the baseline or a carefully selected comparison group.

Days 61–90: evaluate, adjust, and decide

Review the evidence with service operations, technology, privacy, security, and relevant business owners. Separate observed outcomes from assumptions. Update the workflow or narrow the scope if the system is helpful only for a subset of requests. Re-test after material changes.

Possible decisions are to stop, continue a controlled pilot, revise the use case, or expand gradually. Expansion should require evidence that service quality and safeguards remain acceptable, not simply that the technology is available.

Choose measures that reflect customer outcomes

  • Resolution: Was the issue resolved correctly, and did the customer need to contact the organization again?
  • Quality: Were answers accurate, understandable, and appropriate for the language and channel?
  • Escalation: Did the system route uncertainty and sensitive requests to the right person?
  • Customer effort: Did customers repeat information, abandon the interaction, or report confusion?
  • Employee impact: Did agent workload improve, or was time shifted to verification and correction?
  • Operations: What are the ongoing integration, review, support, and maintenance costs?

Define each metric, its data source, and acceptable threshold before the pilot. Watch for trade-offs: shorter handling time could coincide with more repeat contacts, and a higher automation rate could coincide with weaker resolution quality.

Set clear stop criteria

Pause or restrict the pilot if it exposes data, provides materially misleading guidance, mishandles a high-impact request, repeatedly fails language quality criteria, or cannot reliably hand off to a person. Assign someone to make the pause decision and a process to investigate before restarting.

NIST’s Generative AI Profile is a voluntary general risk-management resource, not GCC law or a guarantee of safe operation. This 90-day structure is a planning aid; it is not a promise of cost savings, service improvement, or deployment readiness.

Important information

This article is general information, not legal, security, privacy, financial, or procurement advice. Review the actual system and use case with qualified advisers appropriate to your organization.

Back to all NBS insights