Making decisions with AI support

0:00 Decision making with AI

Here's a question worth pausing on. If an AI assistant analyzes your data, generates options, scores them and recommends one, who made the decision? Legally, ethically and practically, you did. AI can widen your options, stress-test your reasoning and speed up analysis. It can't own accountability, fully understand your organization's context and values, or bear the consequences. In this lecture you'll learn a decision workflow with AI, the biases it can counter and amplify, five decision frameworks, an expected-value table with sensitivity, and how to document decisions so you can learn from them.

0:40 Why it matters

Why does this matter? Because the goal is augmented judgment: better decisions because AI helped you think, not decisions delegated to a model. Delegation fails in quiet ways. The model doesn't know the competitor cut is probably temporary. It doesn't know your brand promise. And it tends to agree with however you framed the question. Organizations that use AI well get faster and more thorough deliberation. Organizations that use it badly get confident mistakes with nobody accountable.

1:13 Decision workflow

Here's the workflow. Frame: the decision, options, criteria and constraints. Gather evidence: analyzes with verified numbers. Generate options: ask AI for alternatives you haven't considered. Stress-test: premortem, devil's advocate and assumption checks. Decide: weigh evidence and values, and record the reasoning. Review: after the outcome, compare with expectations and learn. Notice AI is strongest at generating and stress-testing, and people are essential at framing, deciding and owning the review.

1:43 Match rigor to stakes

Match rigor to stakes. Think of doors. Some decisions are two-way doors: reversible and low impact, like a subject line test. Walk through, measure, walk back if needed. Some are two-way but higher impact, like shifting budget between channels: verify the analysis, stress-test and pilot. And some are one-way doors: hard to reverse and high impact, like pricing changes or entering a new market. Those deserve full verification, multiple perspectives, experiments where possible, and documented reasoning.

2:16 Biases

Now biases. AI can help counter confirmation bias, by arguing the other side; narrow framing, by generating more options; and overconfidence, by listing failure modes. But AI can amplify biases too. Automation bias: trusting outputs because they look authoritative. Anchoring: the first AI suggestion becomes the default. And sycophancy: models may lean toward agreeing with the framing you give them. So ask neutrally, or explicitly ask the model to argue against your preferred option.

2:48 Five frameworks

Five frameworks worth knowing. Reversible versus irreversible, to set the rigor. Expected value, for options with estimable outcomes and probabilities. A weighted criteria matrix, when there are several options and criteria; AI can draft criteria and justified scores, but you set the weights, and a numeric total mustn't disguise subjective inputs. The premortem: imagine it's six months later and the decision failed badly, write the most plausible story of why, and list early warning signs. And a decision record with a review date.

3:24 Example 1: competitor price cut

First example, a business case. A regional e-commerce company faces a competitor's fifteen percent price cut in the UAE. Verified analysis shows most revenue comes from customers who rarely compare prices. The AI generates options beyond match or hold, including a delivery perk targeted at new customers. The premortem flags risk among price-sensitive new customers. The team holds price, adds the delivery perk for new customers, sets warning metrics and documents the decision. Three months later, retention held, and the record shows one assumption, delivery cost, was wrong: a lesson for next time.

4:04 Example 2: Dubai office or remote lead?

Second example. A Lahore software company considers opening a Dubai office, a one-way door, versus hiring a remote Gulf sales lead, a two-way door. The team uses AI to draft a premortem for each, and builds a simple expected-value table with explicit win-rate assumptions. The sensitivity analysis shows the office only pays off if win rates double, which nobody can support with evidence yet. They choose the reversible option, set a six-month review with success metrics, and record the decision.

4:39 Watch me do it, part 1

Watch me build an expected-value table. Three options for the price-cut decision: match price, hold with free delivery, or hold price. For each, a probability of keeping our target share, which is a judgment, and profit if we keep share and if we lose it, in thousands of dirhams per quarter, all illustrative. Expected profit is the probability times profit if kept, plus one minus the probability times profit if lost. The table ranks hold with free delivery first. But that ranking is only as good as the probabilities.

5:18 Watch me do it, part 2

So I run sensitivity. How low can the probability of keeping share fall before hold with free delivery stops beating matching the price? I loop over eighty, seventy, sixty and fifty percent. At sixty percent it still wins; at fifty it loses. Now the decision hinges on one question: are we confident at least three in five target customers stay? That tells me exactly what to monitor after launch, and what a quick test in one emirate could check first.

5:53 Premortem

Then the premortem, with AI. Imagine it's six months from now and our decision to hold price with free delivery failed badly. Write the most plausible story of why it failed, then list early warning signs we should monitor. The AI suggests the competitor's cut became permanent, delivery costs rose, and new customers churned after their first order. I turn the warning signs into three metrics with thresholds, and they go straight into the decision record.

6:26 The decision record

The decision record ties it together. Decision, date and owner. Evidence, with links to verified analysis. Alternatives considered. Key assumptions, including the tipping point. Review date and success metric. AI can draft it from your notes. When you review later, you can see whether a decision failed because of bad reasoning, bad information or bad luck, which are very different lessons. And for decisions affecting individuals, like hiring, credit or personalized pricing, take extra care: fairness, explainability and data protection rules apply.

7:02 Option generation

A word on option generation, because it's where AI adds the most value for the least risk. Give the model the decision, the options you already have, and the constraints, and ask for four more options with the main risk of each. For the price-cut case, it suggested bundling, a loyalty credit, targeting the perk at new customers only, and a time-limited match in one emirate. Two were weak, one was already on our list, and one became the chosen strategy. That's a good hit rate for thirty seconds of work.

7:42 Common mistakes

Common mistakes. Letting AI choose rather than advise. Asking leading questions that invite agreement. Treating a scoring matrix total as objective truth. Skipping sensitivity analysis on judgment-based numbers. And never reviewing decisions after the outcome, which wastes the most valuable learning data you have.

8:01 Recap and try this now

Recap. AI augments judgment; accountability stays with people. Frame, gather verified evidence, generate options, stress-test, decide and review. Match rigor to reversibility and impact. Use AI to counter confirmation bias and narrow framing, and guard against automation bias, anchoring and sycophancy. Use expected value with sensitivity, premortems and decision records. Try this now: take a decision your team faces, run the option-generation and premortem prompts, build a small expected-value table with a sensitivity check, and write a decision record with a review date.