Fast forward five years and it is close to a guarantee that every manager has an AI helper in the loop on every meaningful decision they make. Gartner already predicts that half of business decisions will be augmented or automated by AI agents for decision intelligence by 2027,* and I'd call that conservative. The question worth arguing about is what has to exist underneath before any of it works.
The market is converging on some version of a context graph: a durable record of how decisions actually get made in a business, what inputs mattered, which exceptions were granted, and what happened next. The definitions still vary depending on who is writing about it that week. But without something like it there is no encoded judgment, only a model reasoning from scratch every time and hoping it lands in the same place twice. Encoding an organisation's judgment is not a job you hand to a model. It is a data problem, and it has to be built. Claude Code + SQL this is not.
Most teams are coming at this from the qualitative side. PlayerZero, Scribe and others ingest communications, tickets and documents, then work backwards to reverse-engineer how people decide. Over a long enough horizon that will work, but it is slow to reach critical mass and needs a human in the loop throughout, because an inferred decision rule has nothing solid to be validated against.
Decisions driven by quantitative data are far easier to codify. The inputs are already structured, the signals are measurable, and the judgment can be made explicit in a day or two: you sit down with a COO, an analyst or a high-performing regional manager and walk through the calls they make against real, live data. Twenty to twenty-five investigations cover a large share of the value that leaks out of a business through trends nobody caught in time.
Because nothing about the existing process has to change, this moves at a pace enterprise AI usually doesn't. Pilot in four to five weeks, production in ten to twelve, with no forward-deployed engineers. We augment the decisions people are already making rather than replacing the way they make them. Multi-location businesses are the sharpest version of this: the metrics are identical at every site, so judgment encoded once scales across the whole network. Every manager in the chain can get individualized briefings telling them what's happening, why, what needs attention, and a recommended action plan, closing the loop between analytics and actions.
The largest pawn chain in the United States has been running this in production since January, built from around twelve hours with their operations team. It has taken close to a week out of how long it takes them to work out which stores need attention. The biggest difference has been to their newer managers: the ones who found a dashboard hard to interpret now get the same read on a store that their most experienced operators would have given them.
The results shouldn't surprise anyone. Brynjolfsson, Li and Raymond found the same thing in customer support a few years ago: a 14% average productivity gain from AI assistance, rising to 34% for novice and lower-skilled workers, with almost no effect on the experts. In this case there are also immediate benefits to the bottom line - when adverse trends get caught early they can be corrected before the P&L ever sees them. The rollout reaches well over a thousand locations in Q4.
I believe two things about 2027. The first is that context graphs become an acknowledged requirement for scaling judgment with AI, so we'll be hearing a lot more about them. The second is that finance departments run out of patience with ambitious AI pilots that never quite reach production - and there are a lot of them out there. Those factors will favor redeploying funding, and talent to faster moving projects that can be stood up and deliver value in a quarter. If any of this sounds familiar, we'd love to talk.
AI decision management is the practice of encoding how decisions get made in a business (the inputs that mattered, the exceptions that were granted, the outcomes that followed) so that AI can augment those decisions consistently at scale. Without encoded judgment, an AI system reasons from scratch every time. With it, the system compounds institutional knowledge rather than replacing it.
A context graph is a durable record of how decisions get made in an organization: the inputs that mattered, the exceptions that were granted, and the outcomes that followed. It gives AI systems something solid to validate against rather than inferring decision rules from behavior alone. Without a context graph, an AI system has no encoded judgment; just a model reasoning from scratch every time.
Quantitative data is already structured, the signals are measurable, and the judgment can be made explicit quickly. You sit down with a strong operator and walk through the calls they make against live data. Qualitative approaches (ingesting communications, tickets, documents) work eventually but are slow to reach critical mass and produce inferred rules with nothing solid to validate against.
Pilot in four to five weeks, production in ten to twelve, with no forward-deployed engineers. Because the existing decision-making process doesn't have to change (only augmented) there is no change management battle to fight before value starts appearing.
Research by Brynjolfsson, Li and Raymond found a 14% average productivity gain from AI assistance, rising to 34% for novice and lower-skilled workers, with almost no effect on experts. The same pattern holds in operational settings: the managers who previously found dashboards hard to interpret get the same read on a location that the best operators would have given them. Augmentation helps the bottom before it helps the top.
Neil Shepherd is co-founder of Scoop Analytics, which builds AI-powered automated data investigation for multi-location businesses.