Optimization and Recommendations
From seeing to prescribing
The final BI layer: not just seeing or predicting, but prescribing — optimization finds the best resource allocation; recommendations target the next best action. Prescriptive output must stay constraint-honest and explainable.
What you will learn
- Frame an optimization as an objective with constraints
- Deliver next-best-action recommendations responsibly
- Keep prescriptive output explainable and constraint-honest
What is this?
Optimization and recommendations is the prescriptive frontier of BI: optimization solves for the best allocation under constraints (where do we put 40 SDRs to maximize pipeline?), and recommendations target next-best actions per record (which price, which promotion, which coverage). The discipline: the objective and constraints are explicit, results are explainable, and the machine proposes while a human disposes.
Why do you need it?
- Prescribing beats describing and predicting when resources are scarce.
- Clear objectives + constraints make 'why this allocation' defendable.
- Recommender outputs raise engagement pages from trivia to guidance.
- It rounds the BI arc: what happened → what next → what's best.
Mental model
A chef's mise en place for 200 covers
Optimization under constraint
The perfect menu dies if the kitchen has 8 burners and 5 hours. The chef optimizes: highest-yield dishes under capacity — same onion, best whites. Optimization is reserved for exactly that: yield, under explicit constraints, then the recommendation is served with its reasoning. Recommendations without constraints are menus for a kitchen that doesn't exist.
How it works
- Objective: the one thing to maximize/minimize (pipeline, margin, coverage).
- Constraints: capacity, budget, coverage, fairness — explicit, listable.
- Solve: an LP/optimizer sweeps feasible allocations under the constraints.
- Recommend: per-record next-best-action with its expected lift and its 'why'.
- Prescribe cautiously: propose + explain; the decision stays human's.
The prescriptive pipeline.
Examples
A capacity allocation framed correctly
OBJECTIVE Max pipeline coverage in 60 daysCONSTRAINTS - 6 SDRs, each 40 calls/day - region budgets fixed - fairness: no segment < 15% coverageSOLUTION Rebalance: 2 SDRs EU→US (US coverage 0.8→1.3) expected +4.2% pipeline, +0.9MRECOMMEND Per-account: next-best-sequence sheet with lift + reasoning per accountHUMAN GATE Approved by sales ops; SDRs keep veto.Objective, constraints, solver, recommendations with reasoning, and a human gate — prescriptions people can hold accountable.
In production
- SDR allocation optimized across regions under coverage + fairness constraints.
- Retention next-best-offer scores each account with expected lift.
- Routing optimization for a field team under travel + priority constraints.
Best practices
- Objective and constraints come from the business, written down, before solving.
- Recommendations carry their lift and their reasoning, not just an order.
- Publish the feasibility story — why this allocation beat the alternatives.
- Human gate at the end: the machine proposes, the team disposes.
Common mistakes
- Optimizing a mystery objective — 'the model says best' with no stated goal.
- Forgotten fairness constraints silently starving a segment.
- Recommendations without reasoning — the team tunes out a black box.
- Model output treated as command rather than the proposal in a review.
Debugging tips
- Allocation looks bizarre → absences: a hard constraint got mis-encoded; verify every line.
- Team ignores recommendations → no 'why' shipped; add lift + reasoning.
- Solver edge 'best' disagrees with judgment → objective wrong; re-derive the business goal.
Interview checkpoint
Quick quiz
A prescriptive recommendation should always include:
The optimizer's output is the final decision.