The market has settled on how much Intelligence Capital matters. Writing in California Management Review, Accenture’s Teresa Tung and Philippe Roussiere call tacit knowledge “your next competitive moat,” and they are right.

The harder problem comes before protecting it: most of your IC was never captured in the first place, and what is not captured cannot be automated.

Inside an operation, IC is not the strategy on the slide. It is the daily micro-decisions underneath it: how you treat a customer technically outside policy, how the difficult calls get made when the written rules run out, and what gets prioritized when everything is urgent.

Some of that is already written down, in your policies and terms, standard operating procedures, contracts, eligibility rules, underwriting guidelines and pricing conditions. 

But much of what gets written follows regulation, market practice and the same templates your competitors work from, which makes it the less distinguishing half.

What is genuinely yours is what the operation learned and never recorded: the exception your best assessor allows, how the difficult decisions actually get made, the reason a rule gets overridden in one region and not another.

That undocumented half is the ceiling on your autonomy. 

An agent can only act on judgment that exists somewhere outside a person’s head, so every decision that still depends on what somebody knows and never wrote down is a decision that cannot run without them. This is a hard constraint on scope, not a maturity gap you grow out of.

Known IC you extract, unknown IC you only catch in the act

Known Intelligence Capital is what the organization has already committed to paper. Getting it into a system is tractable, because the source material exists.

Unknown Intelligence Capital is everything that never made it into a document: the reasoning behind an override, the local practice that grew up because the written rule gives the wrong answer in one situation. 

It becomes visible only when it is applied to a real case, and most clearly when somebody corrects a decision.

Known IC is an extraction problem: the source artifacts exist and can be parsed, so you can get to it whenever you choose. 

Unknown IC is an instrumentation problem: there is no artifact to parse, so you have to catch it where it is applied, by working through real cases with your experts while you build and by recording every correction once you run. If you are doing neither, you are not collecting it at all.

Catching Intelligence Capital has been tried before, and it failed

At AAAI-26, Ted E. Senator published a post-mortem of the expert systems boom and bust, the 1980s attempt to write expert judgment into machine-executable rules, which produced real successes before the field collapsed into an AI winter. Three of his reasons still explain why most attempts fail today.

• Getting the knowledge out was the bottleneck. Acquiring and representing expert knowledge was “the major cost for developing an expert system,” and the knowledge-engineering bottleneck “was believed to be the major impediment.” That is the undocumented half, and it is where most attempts still fail.
• The systems broke at the edges. Handling “something at the boundary or outside the scope of their knowledge,” results were “unreliable at best.” The failure was silent, which is what made it dangerous.
• They were expensive to keep, not just to build. The systems proved “expensive to develop and maintain, and not easily adaptable.”

There is a fourth risk that did not exist then. The model your system runs on will be switched off by its vendor on a date you do not choose, and whatever you built around that one model’s behavior then has to be rebuilt.

None of this is only history. Gartner forecasts more than 40% of agentic AI projects canceled by the end of 2027, on escalating costs, unclear business value and inadequate risk controls, and Forrester names the binding constraint: “Every autonomous action has to be logged and defensible to an auditor, and right now that cost is too high.”

Capture Intelligence Capital in five steps, without guesswork

Most programs stall at the start, because the choice is framed as all or nothing. Gartner’s Shiva Varma names it: “Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure.”

There is a third option, and it is also how you capture.

Run the system in shadow, so that your autonomous system works every decision alongside the person who makes it today but holds no authority to act, logging each outcome with its reasoning and its citation back to the source text, and recording every correction your people make and what they changed it to.

That does two things at once. Your people carry on as normal, so nothing rides on the system and nobody has to trust it on day one.

And every correction they make records undocumented Intelligence Capital at the moment it is used, including the judgment nobody thought to mention while the system was being built. Agreement becomes something you measure rather than assert, and you widen the system’s authority against that measurement.

A shadow run puts every decision on record, but capturing is only the first of five stages, and the judgment can still be lost at any of the other four:

1. Capture. Record not merely that somebody overrode a decision, but what they were looking at and the reason they gave. The written half is drafted from the policy and a sample of decided files. The unwritten half starts with your experts working through real cases while the system is built, rather than describing their judgment in the abstract, and every recorded correction adds to it. Then sort each correction by where it belongs: a flaw in the source logic means the rule has to change, while a judgment the source never covered belongs in a documented exception on top of the rule, not in a rewrite of it.
2. Represent. Put the decision rule where the person accountable for that policy can read it, disagree with it and change it. If the only readable version is code or a prompt, you have moved the bottleneck rather than removed it. When the rule comes from a signed contract, you can change it by issuing a new version for future customers. Each case has to be decided against the terms that applied to it, so every decision should record which version it ran on.
3. Decide. Route on evidence rather than on the model’s opinion of itself. A confidence score describes the model’s own output distribution; it says nothing about whether the evidence the rule depends on is present. Formalize the rule as explicit conditions, then evaluate whether each one can be satisfied from the document in front of it. An unsatisfied condition is a determinate state rather than a low score, and it goes to a person with a stated reason. 
4. Evidence. Emit provenance as a by-product of deciding rather than as a project afterwards: the rule version that applied, the inputs it read, and the passage of source text each input came from. An audit is then answerable from the record itself rather than from a reconstruction.
5. Endure. Keep the judgment layer decoupled from the inference layer, so that replacing a model is component substitution rather than migration, and your rules and decision history survive it untouched.

What changes when your Intelligence Capital becomes deterministic

Encoding what you captured is what turns a record of past decisions into something that decides, and it is a reviewed step: the people accountable for each rule decide what a correction changes. 

Otera formalizes your terms and policies, together with the judgment your domain experts add, into an Agentic Decision Graph: a deterministic representation of the logic your decisions follow. The model does the reading and the formalizing. The graph does the deciding.

That shifts the business from a probabilistic paradigm to a deterministic one, and the distinction is mechanical rather than rhetorical. 

Generative models are non-deterministic by construction: because generation samples from a distribution, the same input can produce different output across runs. Confining the model to formalization, and settling decisions by evaluating the graph, moves that variability into a one-time artifact a person reviews and out of the decision itself. The same input then produces the same decision, for the same stated reason, with the same citation behind it, which is what consistency means in a regulated environment.

If you want a first move, pick one decision your team makes many times a day and start recording not the outcome but the reason behind every correction to it. That record is the asset that everything else is built on.

Intelligence Capital you never captured cannot be improved, cannot be automated, and cannot even be protected. Intelligence Capital encoded in a Decision Graph, with every decision on an audit trail the business keeps, is all three.

‍