Over sixteen years of selling technology to Saudi government entities, I have watched the same meeting play out many times. A vendor opens with the model: its benchmark scores, its parameters, its demo. The room is impressed. Six months later nothing has moved. The demo was never the problem. The problem is that nobody in the room could say which decision the model would change.
The question that unlocks budget
Senior sponsors in authorities and ministries are not buying AI. They are buying a better decision on something they are already accountable for: which contracts to renew, which complaints to escalate, which inspections to prioritise, which payments to release. When the conversation starts there, three things happen. The sponsor recognises their own problem. The finance team can see what a better decision is worth. And the technical team gets a scope that fits on one page.
So my first question in a discovery session is always the same: what is the decision you want to make better, and how will you know it got better? If the answer is vague, we are not ready to talk about models.
Work backwards from the decision
Once the decision is named, everything else follows in order. What evidence would a good decision-maker want in front of them? Which of that evidence exists today, in which system, and who owns it? Only then do we shape a use case, and only after that do we pick the technology. Sometimes the right answer is a large language model. Often it is a rules engine, a better dashboard, or a cleaner workflow with one model call in the middle.
Working this way protects the client from over-engineering. It also protects the vendor. A use case anchored on a decision has a natural acceptance test: did the decision get faster, cheaper or more accurate? That test becomes the milestone, and the milestone becomes the payment.
What changes in the proposal
A decision-first proposal reads differently. The executive summary names the decision and the sponsor, not the platform. The architecture section is shorter because it only covers what the use case needs. The commercial section ties each phase to an outcome the sponsor already cares about. In my experience these proposals win more often, and, more importantly, they survive the procurement process because every stakeholder can see their own interest in them.
Model-first pitch
- Impressive demo
- Vague owner
- No acceptance test
- Stalls at budget
Decision-first pitch
- Named decision
- Named sponsor
- Agreed success metric
- Funded phase one
Three habits worth keeping
First, write the decision on the first slide, in the sponsor's language, before any logo or diagram. Second, insist on a named owner for the success metric on the client side; a use case with no owner will not reach production. Third, keep the technology choice open until the data walk-through is done. Committing to a model before you have seen the data is the fastest way to a demo that never ships.
A discovery session, step by step
A good discovery session takes about ninety minutes and needs four people in the room: the sponsor who owns the decision, someone who does the work every day, someone who owns the data, and someone from finance or strategy who can put a value on improvement. Without the person who does the work, you design for an imagined process. Without the data owner, you promise what you cannot deliver.
We spend the first half hour on the decision itself. How often is it made? Who makes it today, with what information, and how long does it take? What does a bad decision cost, in money, time or public trust? The answers are usually more specific than people expect, and they give us the baseline every later claim will be measured against.
The second half hour is a data walk-through. We open the actual systems, not slides about them, and look at what is captured, how clean it is, and who can approve its use. This is where most AI ideas either become real or quietly drop away, and it is far cheaper to learn that in a meeting room than six weeks into a build.
The last half hour is for ranking. We list the candidate use cases on one page, score each on value and feasibility, and agree which one goes first. The output of the session is a single page that the sponsor can forward to leadership without editing.
When the sponsor cannot name the decision
Sometimes the sponsor has been asked to do something with AI and genuinely does not know where to start. That is fine, and it is common. In those cases I ask them to walk me through their last three escalations, the issues that reached their desk because nobody below could resolve them. Somewhere in those three stories there is almost always a decision that is slow, inconsistent or made on incomplete information. That decision is the first use case.
None of this is new to good consultants. What is new is how tempting it has become to lead with the model, because the models are genuinely impressive. The organisations that get value from AI in the next few years will be the ones whose partners resisted that temptation and started with the decision.
