Deep dive
From AI Insight to Meaningful Recommendations
The next best action needs to be relevant, feasible and appropriate for the person who will take it.
Identifying an operational condition is an important step. Knowing what to do about it is the step that changes the outcome.
Service leaders are under real pressure to make that step count. In a Gartner survey of 265 customer service and support leaders, 77% said they feel pressure from other senior executives to deploy AI, and 75% reported increased AI budgets compared with the previous year. Gartner identifies agent enablement as one of the four areas where AI delivers the most value in service and support, including real-time customer data insights and next-best-action recommendations.
Next-best-action is a well-established idea. Making it work in day-to-day operations depends on something more specific: recommendations that account for what the organisation can actually do, communicate what is known and what is not, and reach each person in a form that fits their responsibilities.
This article looks at how to get there.
What makes a recommendation meaningful
A meaningful recommendation passes three tests.
| Test | The question it answers | What it requires |
| Relevant | Does it address the condition that matters most to this person, right now? | Business context, priority and an understanding of the recipient’s role |
| Feasible | Can the organisation actually do it with the resources available? | Live capacity, skills, dependencies and timing |
| Appropriate | Is it permitted, and does it sit within this person’s authority? | Policies, approval rules and governance boundaries |
A recommendation can be analytically sound and still fail one of these tests. Suggesting extra staff when no appropriately skilled staff are available fails feasibility. Suggesting a change that needs executive approval to a team leader fails appropriateness. The three tests are what connect AI insight to real operational choices.
From a generic suggestion to a useful recommendation
Consider a workforce manager receiving an alert about deteriorating service levels in a billing queue.
| GENERIC SUGGESTION
“Increase staffing in the billing queue.” Directionally right, but it leaves the manager to work out whether it is possible, what it would cost elsewhere and how long it would help. |
CONTEXTUAL RECOMMENDATION
“Four agents in general enquiries are billing-skilled and that queue is running 18% below forecast demand. Moving them to billing for the next 45 minutes is expected to bring service level back within target without putting general enquiries at risk. After 11:30 a.m., general enquiries demand is forecast to rise, so plan to return them by then.” Requires: Workforce Manager approval · Reassess: 11:00 a.m. |
Illustrative example. All figures are examples, not customer results.
The contextual recommendation answers the questions a manager would ask before acting: are the right people available, what is the trade-off, and how long does the benefit last? Answering them requires data from several places at once: skills from workforce management, demand from the contact centre platform and forecasts from analytics.
Three principles for better recommendations
- Evidence before explanation
Recommendations should be grounded in the available facts. When the cause of a condition is uncertain, the recommendation should communicate what is known and what warrants further investigation, rather than presenting a plausible explanation as an established one.
- Options before automation
Intelligence should identify suitable interventions and their implications. Routine, low-risk actions may be candidates for automation where authorised. More consequential decisions follow defined approval processes. A simple tiered model helps make that explicit:
| Level | What the system does | Typical use | Human role |
| Inform | Surfaces the condition with context | Emerging trends, early warnings | Interpret and decide |
| Recommend | Proposes one or more options with trade-offs | Capacity changes, prioritisation | Choose and approve |
| Prepare | Recommends and drafts the action ready to execute | Roster adjustments, customer notifications | Review and confirm |
| Automate within guardrails | Executes a pre-approved, low-risk action | Routine routing or reminder rules | Set rules and monitor |
Most organisations will use several levels at once, matched to the risk and reversibility of each decision.
- Business relevance before technical sophistication
A statistically impressive anomaly may have little operational significance. A modest but persistent change affecting a critical service can justify immediate attention. Recommendations should be prioritised by business impact and the time remaining to act, not by how unusual the underlying data looks.
Communicating uncertainty honestly
One of the most important distinctions in AI-supported operations is between a plausible explanation and a validated conclusion. Higher demand and reduced capacity may be verified facts; their relative contribution to a service issue may still require analysis.
A well-designed recommendation makes those layers visible:
| Layer | Example wording |
| Observed facts | “Demand is 14% above the comparable period. Two scheduled agents are unavailable.” |
| Likely contributors | “This morning’s billing notice is a likely driver of the increase in handling time.” |
| Open questions | “It is not yet clear whether demand will ease after the 11:00 a.m. update.” |
| Expected effect | “Reallocating four agents is expected to restore service level within 30 minutes.” |
Different roles, different recommendations
The same emerging condition requires different information for different people. This is persona-based intelligence in practice: recommendations shaped to each person’s decision responsibility, rather than a single view for everyone.
Take a field service operation where open jobs are rising as two technicians go off shift early:
| Role | Decision responsibility | Recommendation they receive |
| Dispatch team leader | The next hour of job allocation | Reassign three same-day jobs to the nearest available technicians; offer two customers a later window |
| Field operations manager | Capacity across regions and shifts | Approve overtime for one crew or borrow capacity from the neighbouring region; impact on tomorrow’s schedule shown |
| Service executive | Customer commitments and cost | Same-day commitment at risk for 12 customers; intervention options and cost; result reported at end of day |
| Data and analytics team | Quality of the intelligence | Recommendation acceptance and accuracy for this condition type; data freshness issues flagged |
Illustrative example. All figures are examples, not customer results.
Each view reflects a distinct decision responsibility. When intelligence mirrors those responsibilities, more people can use AI productively without anyone receiving information or options outside their authority.
Five habits of recommendations people trust
A recommendation only creates value if someone acts on it. Teams that see high acceptance tend to share a handful of habits in how recommendations are presented.
- Show the working
Link every recommendation to the evidence behind it, one tap away. People accept advice more readily when they can check it, even if they rarely do.
- Offer a real alternative
Present a second option with its trade-off, such as “move four agents now” or “offer callbacks and reassess at 11:00”. A choice keeps the decision with the person who owns it and reveals which trade-offs they prefer.
- State the cost as well as the benefit
If moving agents puts another queue at modest risk, say so and quantify it. A recommendation that hides its cost loses credibility the first time that cost appears.
- Give it an expiry time
Operational recommendations go stale quickly. “Valid until 10:40” tells the recipient when the advice stops applying and prevents an out-of-date action being taken after conditions have changed.
- Learn from rejection
When a recommendation is declined, capture the reason with a one-tap choice (not feasible, wrong timing, already handled, disagree) and avoid repeating the same suggestion without new evidence. Forrester highlights the value of exactly this kind of loop in service operations, where “insights prompt upskilling and CSR coaching, new knowledge base content, and improved processes.”
Getting started
- Pick one recurring decision. Choose a decision that happens often and has a measurable outcome, such as rebalancing capacity in a priority queue.
- List the real options. Document the interventions the organisation can actually take, their constraints and who can approve them.
- Map the data each option needs. Identify which systems hold the skills, capacity, demand and policy data required to assess feasibility.
- Set the level of automation. Decide where each option sits on the inform–recommend–prepare–automate scale.
- Design by role and measure. Shape the recommendation for each role and track acceptance and outcomes from day one.
The Decision-Ready Alert Canvas (see below) includes a worksheet for defining actions, guardrails and owners for a single use case.
Where emite fits
Feasibility is the test most recommendations fail, and it depends on live operational data: who is rostered, what they are skilled to do and how demand is tracking elsewhere. emite Advanced iPaaS brings workforce, contact centre and operational data together so options can be checked against current capacity, and Advanced Visualisation gives each role its own view, from team leader to executive.
AskEmite offers suggested next steps grounded in that connected data. People can ask why an option was suggested or what else is possible, and decide for themselves, with consequential decisions staying with the people who own them. Your data answers back.
The key takeaway
AI insight earns its value when it becomes a recommendation someone can act on. That means grounding recommendations in evidence, matching automation to risk, prioritising by business relevance, being clear about uncertainty, and shaping each recommendation to the person who will act on it.
A recommendation is only as valuable as the action it makes possible.
References
- Gartner, Gartner Says the Most Valuable AI Use Cases for Customer Service and Support Fall into Four Areas, 8 October 2025 (survey of 265 service and support leaders, April–May 2025). Read the source
- Forrester, Think About The Agentic Shift With Your Next Customer Service Solution, Kate Leggett, March 2026. Read the source
Scenarios and figures marked illustrative are examples, not customer results. This article is not sponsored or endorsed by any organisation mentioned.
RELATED RESOURCES
