The knowledge leg of enterprise AI, why execution without expertise fails

I have spent the last two years inside real enterprises, shipping agents that had to work on real cases under real deadlines. I have seen every way these projects fail, and I have watched almost every failure trace back to the same missing piece.

It is not the model. It is not the integration. It is not the change program.

It is that the company cannot tell you how the work actually gets done.

That gap has a name now. Every working enterprise agent stands on two legs: a knowledge leg and an action leg. Both have to be solid. The market keeps building the action leg. Almost nobody is building the knowledge leg. And the numbers are starting to catch up with that choice.

Two legs, one standing agent

Here is the frame in one paragraph.

The action leg is what most people mean when they say "agentic AI." It is the orchestration, the tool calls, the multi-agent hand-offs, the connectors into SAP and Salesforce and the ticketing system. It is the part of the stack that everyone is racing to build, because it is visible, demoable, and technically fun.

The knowledge leg is different. It is the operating knowledge that the agent stands on. Not the model's training data. Not the documents in the wiki. The tacit, situational, judgment-heavy expertise that lives in a company's best people, the risk analyst who knows which case gets approved and which does not, the planner who knows which supplier to trust when the system says everything is fine, the underwriter who knows which flag actually matters.

When both legs are solid, the agent stands. It handles the easy cases, and it handles the hard ones. It escalates for the right reasons. It stops when it should stop.

When only the action leg exists, you get a fluent, well-integrated system that fails on the cases that matter most. It handles the demo. It stumbles on the real work.

That mismatch is what is showing up now in the field data.

The numbers nobody wanted to see

For a while, the enterprise AI story ran on optimism. Now the field data is coming in.

MIT NANDA's research puts it starkly: 95% of enterprise AI investments show no measurable business impact. That is not a typo. It is not a bad-quarter finding. It is a structural signal about what is happening inside real organizations that have deployed real AI.

The RAND Corporation reached a similar conclusion from a different angle, roughly 80% of AI projects fail, more than twice the rate of comparable non-AI IT projects. McKinsey's State of AI research shows the same pattern: heavy AI adoption, high spend, and most organizations still without measurable EBIT impact at the enterprise level.

And now Gartner has added the agentic-specific version of the same story: a projected 40% of agentic AI projects will be cancelled by 2027, driven by unclear business value, escalating costs, and inadequate risk controls.

These numbers get read as "the technology is not ready." That is the wrong reading. The technology is ready. The models are strong enough. The tooling is good enough. The integrations are doable.

What is missing is the substrate the agents read from. The knowledge leg. The projects that fail are not failing because the action layer is weak. They are failing because the action layer has nothing solid underneath it.

Why action-only agents break on real cases

I have watched this pattern in project after project, across insurance, manufacturing, ecommerce, maritime and consulting. The shape is always the same.

The pilot begins with excitement. A team stands up a workflow, a claim triage flow, a supplier decision flow, a customer response flow. The connectors work. The agents call the right APIs. On the demo cases, the output looks impressive.

Then a real case hits the desk. Not a policy question with a documented answer. A situation the official process does not cover. An edge case. A conflict between two systems. A judgment call the operator has made a hundred times but has never written down.

The agent has no substrate for this. It reads from documents that describe the official process, not the real one. It has no access to the reasoning behind past decisions. It cannot tell which internal field is fresh and which is two years stale. It does not know that this supplier's quote should be discounted 15% because they always overstate, or that this claim category has a hidden approval loop that runs through a specific person.

So it does what a fluent agent with no operating knowledge always does. It guesses. Sometimes it guesses in a way that sounds plausible. Sometimes it guesses in a way that is quietly wrong. Sometimes it stalls and hands the case back with a generic response.

The team stops trusting it. Usage drops. The pilot slips into "we are still evaluating." Six months later it is cancelled.

That is the failure mode Gartner is pricing into its 2027 projection. And it is not going to be fixed by a better orchestration framework. The gap is upstream of orchestration.

What the knowledge leg actually is

The knowledge leg is not documents. This is where most enterprise AI programs go wrong.

Documents are the official version of the work. They describe what the company says should happen. Roughly 80% of how work actually gets done in large organizations is undocumented (Tallyfy). Research from Panopto shows that 42% of the expertise essential to a company's operations lives only in employees' heads. Knowledge workers, per McKinsey, still spend roughly 19% of their time searching for information they know exists somewhere but cannot find.

The knowledge leg is what fills that gap. It is operating knowledge: the decision logic, the exception handling, the trusted sources, the escalation triggers, and the situational reasoning that make a real operation function. It is what a senior employee has built up over years and can demonstrate on demand but cannot easily write down.

For an agent to stand on that knowledge, it has to be captured in a form the agent can actually read. Not a transcript. Not a nicer wiki page. Not a fine-tuned model that memorized last quarter's SOPs. A structured, machine-readable expert profile that carries the why behind the decisions, wired into the systems where the work happens, and kept current as the business changes.

That is what expertise mining produces, and it is the specific substrate the knowledge leg needs.

Why knowledge has to come before action

You could argue the order does not matter, build the action leg, then bolt on knowledge later. In practice, that order fails.

The reason is that the action leg is designed around the substrate underneath it. If you build orchestration on top of documents, your agents will make decisions the way documents describe them. If you build orchestration on top of retrieval from a wiki, your agents will handle the cases the wiki covers. The scope of the action layer is inherited from the scope of the knowledge layer.

If the knowledge layer is thin, the action layer is thin. You cannot make it deeper by rewriting prompts.

Building the knowledge leg first inverts this. You start with the operating logic of the actual work, the way a senior underwriter reasons about a marginal case, the way a planner decides whether to override the system, the way an ops lead knows which escalation matters. You structure it. You validate it. Then you build the action layer on top so the agent has something real to act on.

This is why the pilots that survive tend to look the same. They started with a small piece of captured expertise, they wrapped a narrow governed action layer around it, and they expanded outward. The pilots that fail almost always started at the action layer, assumed the knowledge would come from documents and RAG, and hit the substrate ceiling within weeks.

What "the knowledge leg first" looks like in practice

The knowledge leg is not a philosophy. It is a specific piece of work that produces a specific artifact.

At Verti, we sit with an expert in a normal conversation and pull out how they actually work, not what the SOP says, but what they do. We capture the decisions, the reasons behind them, the exceptions, the sources they trust, and the moments they stop and escalate. We turn that into a machine-readable expert profile that carries the operating logic in structured form.

That profile is what an agent reads from. It knows which system holds which field, whether the field is fresh, and what an agent is allowed to do with it: read it, draft with it, or stop and ask a human. The action layer sits on top of this substrate, not in place of it.

The pattern shows up in the deployments where this order was respected.

At a large multi-brand insurance group, the insurance renewal workflow moved from six minutes to one, an 83% cut in handling time. The risk team went from eight people to three on a related process, freeing about 833 labor hours per month per ten thousand renewals. The reason it worked is that the underwriting judgment was captured first. The action layer only executes what the captured expertise says is appropriate. When the case falls outside that boundary, it stops.

At a supplements manufacturer, candidate screening for a hundred applicants dropped from eight hours to thirty minutes. The trick was not a faster resume parser. It was capturing what a senior recruiter actually looks for, the signals that predict success in that specific role, at that specific company, in that specific market. Once that operating knowledge existed as a profile, the action layer could apply it consistently.

At a maritime chartering firm, an eight-person team now runs with one. At an ecommerce operation, live insight agents surface daily, weekly and monthly recommendations across sales, marketing and operations, because the operating knowledge of the experienced merchandisers was captured before the action layer was wired in.

None of these are demos. They are what happens when the knowledge leg is built first.

The cliff behind the numbers

There is a second reason the knowledge leg matters now, and it is worse than the pilot-failure story.

The people who hold the operating knowledge inside enterprises are moving. In the U.S. private sector, median employee tenure is 3.5 years (Bureau of Labor Statistics). In manufacturing, the Deloitte and Manufacturing Institute study projects 2.1 million unfilled roles by 2030, driven largely by retirement of experienced workers. When those people leave, the operating knowledge leaves with them, the decision reasoning, the exception handling, the vendor context, the "I always check this field first because the other one is usually stale" logic that keeps the operation running.

An enterprise that has not built its knowledge leg is exposed on two sides at once. Its AI pilots fail because there is no substrate for the agents to read from. And its human operations lose depth every quarter as senior people rotate out and take the operating knowledge with them.

Building the knowledge leg does not just make the agents work. It captures a category of asset the company has never actually owned, the operating expertise that lives in a small number of experienced people and has always been one departure away from disappearing.

That is why the case for building it first is not just an AI case. It is an operational continuity case. The knowledge leg gives the company something durable to stand on, regardless of who stays and who leaves.

What changes when the knowledge leg is in place

Once the knowledge leg exists, the shape of the AI program changes.

Pilots stop stalling on real cases. The agent has a substrate for the hard ones, not just the easy ones. It knows why a decision gets made, not just what the decision is. It stops guessing when it hits the boundary, because the captured expertise tells it where the boundary is.

Governance becomes possible. An action layer built on captured expertise can be audited. You can point to the specific piece of operating knowledge that produced a decision, show which system was read, show why the agent stopped and escalated. This is what governed agent operations actually means. Not a policy document, a system where every action is traceable back to the substrate underneath it.

Onboarding accelerates. The same operating knowledge that trains the agent can train the next generation of human operators. A new hire is no longer learning from scattered documents and hallway conversations. They are learning from a structured, validated capture of how the real work happens.

Continuous learning becomes real. As the business changes, the operating knowledge changes with it, and the agent's behavior updates in step, under human approval. Nothing drifts silently. Nothing goes stale without someone signing off on the update.

The action layer gets simpler. When the substrate carries the judgment, the action layer does not have to. It becomes a set of narrow, governed specialists that read from the profile, do their work inside real systems, and ask for approval before anything that matters. That is a much smaller engineering problem than what most agentic AI programs are trying to solve today.

The programs that get this right will not look like the 95% MIT NANDA is measuring. They will look like the outliers, the ones where the business impact is measurable, the governance is real, and the system survives contact with the operation.

The category choice ahead

There is a strategic question underneath all of this, and CAIOs and CTOs are going to have to answer it in the next 12 to 24 months.

You can keep building the action leg on the assumption that the knowledge layer will eventually be solved by better models, longer context windows, and more RAG. This is the current default. The Gartner projection suggests where it leads.

Or you can build the knowledge leg first: capture the operating expertise that lives in your best people, structure it into machine-readable profiles, and layer a governed action system on top of it. This is what the working enterprise deployments look like once you strip the marketing off them.

The difference between the two paths is not a technology choice. It is a substrate choice. And substrate choices compound. Every agent you build, every workflow you automate, every governance framework you write is going to inherit the properties of whatever knowledge layer sits underneath it.

The market will keep talking about the action leg because it is the visible half. The teams that quietly win will be the ones who did the unglamorous work first, the work of capturing what their best people actually know, in a form an AI can stand on.

That is the knowledge leg. Build it first, and the action leg starts working. Skip it, and you become one of the 95%.

Mine the expertise. Then operate differently.