I was delighted to be going to a Gartner conference about decisions. I have been thinking and writing about decisions for the last two or three years, and here they were on the main stage. The funny thing is that somewhere in the two days I realised that the idea needs refining. When is a decision a decision, and when is it a choice? Once you look at the motivations behind a decision, it stops being a piece of analysis and starts to be, among other things, an expression of values. So this article has to start with a definition.

The conference was the Gartner Supply Chain Planning Summit in London this month. I run Convex Supply Chain, a consultancy that works on supply chain interventions. In previous years I was attached to a vendor’s or a consultancy’s branded booth, expected to lurk and pitch. This time I wandered in and out of sessions and met up with old friends. The conference app let me catch up, at a high level, on the tracks I had missed. It was two days well spent.

Below are the four themes that stood out for me, each viewed with the distinction between a choice and a decision in mind

The definition first

Here I use the two words in a specific way.

A decision is a selection that can be worked out once the objective and the constraints are agreed. Two competent planners with the same data reach the same answer. A system can make the decision, or check it.

A choice sets the objective. It says what the company wants when two or more things it wants are in conflict, which is why a strategy is a choice between at least two good options. Analysis does not settle a choice on its own, because the analysis needs the choice as an input. A person with the authority makes the choice, and is accountable for it.

The service level is a clear example. The target, say, 97% on time and in full for a product family, is a choice. It balances the cost of holding stock against the cost of a stock-out, and it ranks this family against the others, so somebody with executive authority has to say what the business chooses. The safety stock needed to deliver 97% is a decision. Where to hold that stock in a multi-echelon network is also a decision. Once the inventory target is chosen, the system nails it.  Herbert Simon drew a related line in 1960 between programmed and non-programmed decisions, the routine and the novel. 

Many planning problems happen because a choice is buried inside what is treated as a routine decision. Take the allocation of short supply. The weekly decision is how many units go to each customer, worked out from orders, stock on hand, incoming supply and customer priorities. If the priorities are agreed and written down, the allocation is a decision that the system makes and a planner checks. If the priorities are disputed, each weekly run demands a choice: whose revenue, or whose relationship, matters more? That choice is made in the meeting by whoever has the most influence, and it is probably made again the following week. The friction in these meetings comes from the choice. Few people argue about the maths.(Although traditionally data sources, quality, and provenance can be bones of contention).

A possible remedy is to separate decisions and choices, setting the priority rules once. Give them an executive owner and a review date, and write them down. The weekly allocation then goes back to being a decision. The choice does not go away. It is made by the person with the authority to own it, once, rather than week after week.

That is the definition. The four themes below test it against the conference’s themes. 

(The figures in this article come from surveys shown on slides. Some have small samples, and I have not seen the full research behind them. I would treat them as indications.)

1. The decision becomes the product

The opening keynote proposed a new objective for a planning function. The old measure was a better plan: a faster process, a more precise plan, more alignment. The new measure is a better decision, which the keynote described as the product of three things:

  1. How fast the decision is made
  2. How sound it is
  3. How far the organisation acts on it.

Three of the six takeaways that Gartner published after the event are about decisions, and several analysts asked companies to write down the decisions they have to make, each with an owner.

I welcome this. I argued in The Centaur’s Crossing that the decision is the deliverable of a planning process, and The Anatomy of Decisions describes how to find and record those decisions.

But I have two reservations.

The first is that the formula fits decisions but not choices. Speed, quality and execution can be measured against an agreed objective. For a choice, speed is not always a virtue, quality has no definition until the trade-off behind it is clear, and execution depends on whether the people who were overruled accept the authority of the person making the choice.

Two figures from the keynote are revealing on what this looks like in practice. In one Gartner survey, 67% of organisations said that they assess the quality of their decisions informally, rarely or never. In another, 73% of supply chain respondents said that other functions override or change their decisions. The quality of a decision can be measured against its objective. The quality of a choice cannot be measured until the priorities behind it are stated, so part of the 67% is a measurement gap and part of it is an unresolved governance question.

A decision overridden can be a correction: somebody had better information, or somebody was wrong. An override of a choice is another function asserting a different preference. The first is a data problem and the second is a question of authority. A company that records which kind of override it suffers from will know which problem to work on.

Before it buys another planning module, a company should look at its last planning cycle: which figures changed, who changed them, and whether the changes were executed. It should also ask which of those changes were decisions and which were choices. In my experience that takes at least one cycle to answer, and the answer is often uncomfortable. This is where the elephant in the room may raise its trunk.

A semiconductor company I talked to in London recognised this pattern. There is a lot of emotion behind the allocation of supply to customers, at all levels up to and including the CEO. That emotion is the trumpeting of a choice that has not been settled. Which customer the company will disappoint when supply is short is a choice, and it belongs to the executive who will own the commercial consequences. If the CEO makes this choice on emotion and gut feel, it sets an example that others will follow.

The second reservation is about the label. Gartner’s own Hype Cycle places “decision-centric planning” at the peak of expectations, with five to ten years before it becomes mainstream. The hype curve attracts products, and products can be built fast these days. I expect to see planning software sold as “decision-centric” in the next year, and I expect some buyers to believe that it will be a silver bullet. A tool can be good at decisions and blind to choices. Managing decisions is a tame problem, and settling choices is a wicked one, because the people involved do not agree on the objective itself. 

2. AI: most of it is still in pilot

The numbers on AI were sobering for a technology conference. One analyst reported that 84% of supply chain organisations remain in pilot or limited deployment, and that 42% see zero, break-even or negative return on what they have spent. A consumer health company told the room that it had counted more than 50 digital programmes, with the spend tracked in detail, and that the value it could track and attribute to them was £0. I respect a speaker who puts that number on a slide. I have criticised Gartner events in the past for being too gung-ho. If they are showing figures this blunt, the situation on the ground must be dire.

The analysts did not blame the technology. One opened with self-driving cars, which now have a better safety record than human drivers and which few people say they would trust. Another said that automation is a governance challenge more than an AI challenge. A third said that the risk has moved from too little AI in a project to too much: more than the decision needs or the data can support.

This agrees with what I see. In AI Projects Deserve Project Management I argued that pilots stall because they are run without the discipline that we would apply to a new warehouse or an ERP release. Many have no link to a business decision and no agreed test for go or no-go.

A question should come before the pilot: is the objective agreed and written down? If it is not, the team is trying to automate a choice. The system will produce an answer, an executive will override it, and the project will join the 42%. The next question is whether the problem needs AI at all, or whether it needs plain automation.

I would add one point that was less visible in London. “Governance” risks becoming a committee. The practical form is a short answer, per decision, to the question of how much the system may decide alone. 

3. The human side ranks above the data

One survey result appeared in at least three presentations. Gartner asked 243 organisations what stops them from reaching their goals for autonomous planning:

  • Organisation structure: 45%
  • Analytics skills of stakeholders: 43%
  • Trust in algorithms: 39%
  • Data that is not available: 37%

The adoption figures tell the same story. In one survey, 62% said that their planning tool was deployed on time, and 45% said that people work around it. In another, 34% went back to their old tools after an automation project. One analyst made a useful separation between deployment, which is measured in technical milestones, and adoption, which is measured in behaviour.

This is where the value leaks out of a tech investment. When adoption falters, the company often blames the platform and switches vendor, and the cycle repeats. I continue to be amazed that vendors do not do more about this. The only reason I can think of is that they believe we are still in 1996, selling licences. Tell me I’m wrong!

I also think that “low adoption” is too coarse a diagnosis. A planner who overrides the system may lack trust, or may know something that the system does not, or may answer to a manager who is measured on a different number. The first two planners are overriding a decision. The third is making a choice that should have been made two levels up, because two executive priorities have not been reconciled. I think that is also why organisation structure tops the survey. Where governance is vague, the trade-offs get settled by default, in the planner’s overrides. An override is useful data, but few companies collect the detail. The digital fingerprints at the scene of the crime.

On the planner’s job, Gartner’s phrase is a move from operator to orchestrator. The planner sets what the system must do and inspects what it delivers. My own picture is the centaur, the human and machine partnership that Garry Kasparov described after Deep Blue, and I see the change as a promotion for the planner. The division of labour fits the distinction: the machine makes the decisions, and the human makes the choices.

The analysts were more cautious than I am on one point, and I think they are right: 45% of chief supply chain officers name the loss of domain expertise as a top concern. A planner can only overrule a recommendation if they still understand the supply chain behind it. A planner who has lost that understanding will not see when the system’s answer is right for the objective it was given and the objective is wrong. Do not automate a decision until you understand its drivers. How to pass that understanding to the next generation of planners, once the routine work that used to teach it is automated, is a question the conference did not answer. I don’t think we have understood this in society yet either.

4. What is left for the consultant?

One session put a hard question to my own trade: when answers are free, what are clients paying consultants for? The analyst’s argument was that AI compresses the base of the consulting pyramid, where junior staff gather data and build slides. She showed the market moving away from the sale of time and towards the sale of long-term capabilities, with the consultant accountable after go-live. Another analyst had a slide on build-it-yourself applications that included the line “Consultants may become competitors”.

I think the analysis is correct, and I do not find it comfortable. Much of what my industry has sold for 25 years was analysis and configuration, and both are getting cheaper. Analysis and configuration are decision-work, and decision-work is what the software now does.

The work that keeps its value is choice-work, and I see three kinds:

  1. Diagnosis. An outside reading of how an organisation makes its decisions today, and where the choices inside them are being made, before anybody proposes a design.
  2. Mediation. The work between people, where owners are agreed and trade-offs are forced into the open so that choices can be settled.
  3. Accountability. Staying long enough to see whether behaviour changed.

Answers to decisions are getting cheap. Choices are valuable, because somebody has to own the consequences. Perhaps a client should ask a consultant what they will be accountable for six months after go-live.

Afterword

Pope Leo XIV’s encyclical on AI, Magnifica Humanitas, published this May, opens with three questions: “Where are we going? Toward what goal do we wish to orient ourselves? What direction should we choose as a people and as a human community?” Each of them is a choice in the sense I have used here. A system cannot answer them from an agreed objective, because they are the questions that set the objective.

I can’t finish without adding that, while he accepts that the technology has value, the Pope sets out the risks that AI carries for a person. I take three from the book: easy results weaken judgement, output that looks objective goes unchallenged along with its blind spots, and the imitation of human contact erodes the real kind. Each has a small version in a planning function: a planner who stops checking, an answer that is trusted because it came from a system, and a trade-off by a model instead of by the people who are responsible for the outcome.

The argument of this article is a short one. A planning function produces decisions, and most of them can be automated once the choices behind them have been made. The choices do not go away when the software arrives, and they do not get easier. They need an owner, and they need to be written down. At the end of your next planning cycle, write down what was decided and by whom, and mark which of those were choices.

Santander, after dinner, back from London and about to catch the ferry home across the bay.

Santander, after dinner, back from London and about to catch the ferry home across the bay. My phone decided on the aperture and the shutter speed. I chose where to point it.