There’s a common narrative circulating about AI and consulting, and it gets the facts right while drawing the wrong conclusion. The narrative says generative AI is coming for consultants, because a model can now do in minutes the analysis that used to take a junior associate hours. Consulting firms are hiring accordingly, adding AI specialists faster than entry-level analysts and rolling out new AI features almost as fast as they can build them.
The work AI is absorbing was never the hard part of consulting, and it was never really what clients paid for. Understanding why requires looking at what consulting promised in the first place.
Consulting sold synthesis. A firm arrived, collected a company’s scattered data, and turned it into recommendations a leadership team could act on. The synthesis looked like the product. It was closer to the last 10 percent of the work.
Large language models are strong at the last 10 percent. Feed a model clean, well-labeled inputs (data) and a clear question, and it returns a sharp first-pass answer. But it has no way to judge whether those inputs were trustworthy in the first place. The scarce resource was rarely the analysis itself. Most of what looked like analysis was reconstruction, the tedious work of figuring out which system to believe before any analysis could start.
A confident recommendation required working out what the client’s data meant, and which system to trust when two sources disagreed. That reconstruction, done by hand, is what consumed the time and the budget. AI skips straight to the analysis, without doing the reconstruction first.
Models have commoditized the analysis. They don’t arrive knowing the business, and they can’t reconstruct that context on their own. Clients feel this gap. In a recent survey of 3,887 companies, Source Global Research found that 70 percent wouldn’t trust a report produced with AI, and skepticism ran higher among those who had used the tools, not lower. That inversion points to a larger truth. Put an AI agent inside a company, and it runs into an infrastructure problem rather than a reasoning one.
The data-sharing arrangements between internal systems were never built with an autonomous reader in mind. A model can produce clean, defensible analysis and still be wrong, because the problem sits underneath the reasoning. The model doesn’t know that a CRM system’s definition of an active customer differs from the billing system, or that several sources hold conflicting customer data. No prompt or larger model fixes that because the missing information doesn’t live in the model.
It lives in the company, as definitions and rules nobody has written down in one trustworthy place. The pattern repeats across companies. The model usually does its job. What fails is the quiet assumption that everyone agrees on what a customer is, when the systems clearly don’t. Two teams can pull the same metric and get two different answers. One system shows a customer as opted in to marketing; another quietly moved them to opted out after an unlogged complaint, and each team is correct against its own source.
The industry is putting a name to the gap. Gartner calls it a context layer and argues that agents need to understand the meaning behind data at each step of a workflow, warning that organizations that neglect semantics get inaccurate and inefficient agents. The label is new. The problem is the one consultants were always paid to solve by hand, one engagement at a time.
The obvious next move is to package that context and hand it to the agent directly. That instinct is understandable. Tooling for bundling reusable instructions into something an agent can run has genuinely gotten better, and packaging expertise as shareable code is a real capability now. But a shared instruction bundle carries portable knowledge about how to do a task in general. It doesn’t carry a company’s reconciled account of what its own data means. A team can install someone else’s playbook in seconds. Installing the answer to which of a company’s systems is right about a customer takes longer, and it’s the work most companies still haven’t done.
Timing matter because the bottleneck has moved. McKinsey reports that only 7 percent of companies have fully scaled AI, identifying data readiness, the work of connecting structured and unstructured data into a governed, reusable foundation, as the reason. Most companies haven’t built the shared, reconciled context that sits underneath a good model. Record quality matters more with AI in the loop because an agent acting on a flawed definition repeats the mistake at scale.
The practical work is crucial. A company needs to identify which system is the source of truth for each entity, such as a customer or an account, so differences can be reconciled. That reconciled context then needs to be something an agent can query directly, under governance and compliance controls that determine who and what may use it.
The challenge generalizes beyond any single question a company asks. Nobody needs a one-time answer to where the next store should go, or which audience to target this quarter. They need to ask that kind of question again and again, with the same reliability each time. That demands more than reconciled context. It demands the data science to turn that context into a repeatable process, and a way to measure how the process performs against the operational decisions it feeds.
This comes down to ownership. It sits between the teams that each run a system, invisible until an agent acts on the disagreement at scale. The business, not any one system’s team, must own the definitions, not just the systems that store them. That ownership is the next real advantage in AI. Everyone will have access to a comparable model. Almost no one will have reconciled their business into an account their agents can trust.
AI has made the analysis cheap and left the context exactly where it was. Companies that treat that context as infrastructure they own, rather than something they rent from an outside firm each time, will compound an advantage the model alone can’t provide.
