Why has the AI that cost tens of millions of yuan become a white elephant that no one dares to use?

In the internal systems of many large enterprises, there exists a kind of costly silence: the AI platforms deployed at huge costs see their access volume plummet sharply after the initial novelty period, gradually settling into an awkward active user figure on the reports.

Employees are afraid to use them, as they are not sure whether the judgments output by the system will become evidence held against them in future accountability investigations; they are reluctant to use them, as they cannot see the logical chain behind the output results; they do not want to use them, because the existing incentive, rights and responsibility system does not give clear feedback no matter whether they use the system well or poorly.

IBM thus puts forward the assertion that without trust, the AI strategies of most enterprises are doomed to fail.

This assertion is not only about the level of technical compliance. It actually points out the overall shortage of "institutional public goods" encountered in enterprise AI deployment, while the vast majority of decision-makers still use the method of purchasing standardized software to deal with a transformation that requires rebuilding internal contracts.

In my view, the watershed of current enterprise AI competition has shifted from the competition of computing power reserve to the management capability of "organizational friction cost". The following article will gradually analyze why trust has become the most fatal yet most underestimated strategic variable from four dimensions: involution of transaction cost, internal lemon market, principal-agent gap, and the positive spiral of governance and innovation.

How the Lack of Trust Reduces AI to an Expensive Ornament

Why are enterprise AI applications stranded on a large scale? The classic question raised by Ronald Coase in 1937 still remains extremely incisive: he asked why enterprises exist? Because internal instructions and authority relationships can save the repeated search, negotiation and contracting costs in market transactions.

However, when AI systems enter the organization as "electronic agents", an unexpected reversal occurs. Instead of automatically reducing internal coordination costs, they generate a large number of new transaction costs between people and the systems. These costs are highly hidden and will not be directly recorded in any financial statement.

The first form of cost is psychological auditing.

When a procurement manager gets the supplier list recommended by AI, his first reaction is not to evaluate the pros and cons of the plan, but to start figuring out which historical data the model used during training, and whether some past grey operations have been learned into the parameters; if a certain recommendation is adopted and the supplier has an accident later, during the audit traceability, will the person signing for responsibility be me or the algorithm development team.

Similar scenarios also happen in the human resources department. After the recruitment algorithm outputs a sorted list of candidates, the HR director has to spend two extra hours manually rechecking from the pile of resumes whether the top candidates really conform to common sense before clicking "approve" in the system.

These mental consumptions of repeated verification, self-evidence and reserved explanation calibers are in themselves a kind of defensive operation cost.

Every time such operation happens, the efficiency gain originally promised by AI will be eroded by a part.

The second type of cost is more hidden but has a deeper impact, which is the buck-passing game.

Due to the lack of a transparent interpretation mechanism and pre-defined responsibility boundaries, once AI output is involved in cross-departmental collaboration, it will easily slide into collective silence. The marketing department thinks the pricing strategy given by the model is too radical and wants to raise the price, but worries that if the sales volume declines, it will be held accountable for "ignoring data science suggestions"; the data team states that the model is only for decision assistance, and the final decision-making right lies in the business side.

Neither side makes the final decision, so the pricing scheme output by AI flows back and forth in the approval process until its timeliness is completely lost.

What the organization pays for this is not the error cost of a single person, but the time lag and rigidity of the entire decision-making system.

When we piece these phenomena together, we will find that many enterprises' AI pilots have a very high rate of return, but there is a cliff-like decline when they are scaled up for promotion.

The reason behind this is not complicated, because pilot projects are usually undertaken by highly trusted innovation teams and run in a closed environment with abundant resources and risk exemption, so those defensive costs are artificially isolated.

Once promoted to the whole company, ordinary employees have to use their own job security to support a system that they cannot understand and whose rights and responsibility boundaries have not been clearly informed, then defensive instinct will override everything else.

At this time, if the organization fails to provide a set of prepositive "institutional trust public goods", such as a transparent interpretation mechanism, clear responsibility attribution, and unobstructed appeal channels, collective inaction and symbolic use will become the dominant strategy for every rational individual.

The AI system obtained with a large amount of investment will be reduced to a precision engine that no one dares to actually start.

That is to say, the expansion of internal friction costs in enterprises has squeezed the technical possibility into organizational infeasibility.

The continuous rise of friction costs not only leads to a pile of idle tools.

It breeds more hidden consequences: when ordinary employees cannot judge the real quality of AI output at all, a systematic adverse selection is quietly staged inside the enterprise.

Why High-value AI Applications Are Collectively Neglected

In 1970, George Akerlof used the "lemon market" model to reveal how information asymmetry allows defective products to squeeze out high-quality products.

Applying the same logic to enterprise AI application scenarios, the situation is roughly the same: the technical department masters all the details of model performance, while the end users on the business side can only observe the input and output like a black box.

This severe information asymmetry makes users naturally tend to treat all AI suggestions as "lemons" with hidden risks, because the cost of distinguishing good from bad is too high.

Then this mentality gives rise to a highly destructive usage pattern.

Employees will strictly confine AI to the most marginal and insignificant tasks, such as letting generative AI draft holiday greeting emails, sort out first drafts of interviews, and summarize the key points of non-sensitive documents.

Even if errors occur in these scenarios, they will not trigger compliance risks, customer claims or regulatory accountability.

However, high-value scenarios that can really leverage the profit pool, such as credit risk assessment, dynamic pricing strategy, and supply chain vulnerability scanning, are collectively neglected. The seriousness of the problem is that this avoidance is a rational self-protection behavior of users, but from the overall perspective of the organization, it constitutes a typical adverse selection: high-quality AI systems cannot get the opportunity to prove themselves in high-value scenarios, because everyone uses them according to the worst-case assumption; while some mediocre systems that even have hidden defects can run smoothly in the comfort zone of low-sensitivity scenarios, because they have never been pushed to the real front line of inspection.

The entire AI investment portfolio is locked in an interval with extremely low marginal benefit.

This also explains why many enterprises claim that "AI has been fully implemented", but no structural improvement on the profit statement can be seen for a long time. A large amount of resources are poured into minor tweaks that cannot produce significant value increment, while the efficiency logic of core operation processes remains almost unchanged.

The real power of governance and transparency mechanism lies precisely in being able to break this vicious circle.

When users can continuously observe the decision-making basis and risk boundary of the model through channels such as interpretability reports, audit logs and fairness tests, they actually obtain a signal system that can reliably judge the quality of AI output.

High-quality systems can be released from the state of "default suspicion" and gradually enter the core decision-making chain.

Governance does not play a restrictive role here, it performs the function of "quality signal release", which disintegrates the lemon market from the inside.

However, even if the problem of quality signal is solved, the inner activities of employees are far from over.

There is a more sharp question hanging over every human-machine interaction: if something really goes wrong in the end, who will bear the consequences?

This attribution vacuum is the most stubborn gap.

In the AI Era, Unclaimed Responsibilities

From the perspective of corporate finance and incentive theory, introducing an AI system is equivalent to putting a special agent whose behavior cannot be fully contracted inside the organization.

In the traditional principal-agent framework, the principal can align the interests of the agent with his own through salary design, supervision and assessment.

But AI systems have no intentions and do not pursue private interests, yet they do generate outputs with economic consequences.

As a result, a structural vacuum zone appears in responsibility attribution.

For example, a credit examiner rejects the application of a small business owner according to the AI's loan rejection suggestion, and later the regulatory department determines that the decision constitutes discriminatory loan rejection.

When investigating the responsibility, the examiner will say, "I just operated according to the system prompts, and the internal process of the company requires that AI opinions must be recorded." The data science team defends that "the preset fairness indicators of the model have reached the standard, and the deviation in individual cases is a problem of statistical distribution." The compliance department says, "We reviewed the overall compliance status of the model when it went online, and cannot monitor each specific output in real time."

All three parties are logically valid in their own positions, but the organization cannot find an entity that can clearly take responsibility.

The fact here is that no one has effectively priced the tail risk of AI decisions.

The pricing mentioned here is not only the calculation of risk probability, but also the fundamental property right issue of "once the loss occurs, who will bear the financial and reputational consequences, in what order and in what way".

IBM emphasizes that governance must be embedded in the system architecture from the very beginning rather than being attached afterwards. Its profound insight lies in that this is essentially designing an ex-ante responsibility allocation mechanism.

By clarifying in advance the trigger conditions of veto power, escalation processing paths, and responsibility signing rules at different decision nodes, the organization can build a very valuable "accountability option", that is, when the AI output causes adverse consequences, the enterprise can quickly start the preset investigation, correction and compensation process, reduce the chaos to a controllable range, and turn mutual accusations into orderly attribution.

With this option, the AI project is equivalent to buying a cheap institutional insurance for its own tail risk.

For projects that lack this mechanism, when calculating the net present value, a higher risk-adjusted discount rate must be applied to their future cash flows, because they are not only burdened with market fluctuations, but also an unquantifiable organizational responsibility crisis that may detonate at any time.

Therefore, enterprises with poor governance have already lost their AI projects at the valuation stage. The capital market has already marked a higher implicit discount for them through aversion to uncertainty.

Is this quite counterintuitive? Organizations that can sort out responsibility attribution clearly are not putting heavier shackles on themselves, but paving a longer runway for innovation.

When people no longer need to consume energy on the question of "who will bear the consequences after making a mistake", governance completes its most critical leap: transforming from defensive control to the base of offensive capability.

Positive Spiral of Governance Mechanism and AI Innovation

For a long time, the management has generally been trapped in a cognitive inertia that supervision is equivalent to restraint, and control is equivalent to deceleration.

But from the empirical research of large-scale organizational behavior, the situation is just the opposite.

Clear boundaries are often the most important infrastructure for psychological security.

When front-line business personnel know exactly in which scenarios the suggestions given by AI are for reference only, under what conditions they can exercise veto power, where to feed back abnormal situations, and they will not be held personally responsible beyond compliance, they will truly let go of their defenses, and carry out deep confrontation and integration between their years of accumulated subtle intuition and tacit knowledge and algorithm output.

This "confrontational integration" is exactly the creative value that enterprises most desire to obtain from AI investment.

Vague zones will only bring about collective behavioral retreat. Once employees need to secretly guess "will there be trouble if I use it", their brain bandwidth will be consumed in risk speculation and self-protection strategies, and the cognitive resources used to solve substantial business problems will shrink sharply.

In contrast, those enterprises that have taken the lead in establishing clear AI usage boundaries, appeal channels and feedback closed loops inside are reaping an implicit acceleration effect: front-line managers dare to try more radical parameter combinations in supply chain simulation, market analysts are willing to let AI detect those correlation signals ignored by traditional analysis frameworks, and risk management personnel dare to collide model early warnings with subjective judgments, and then calibrate the model in reverse.

This high-frequency, low-risk test cycle is the fundamental driving force for innovation to fall from paper to financial statements.

In the next five to ten years, what determines the AI competitiveness of enterprises may no longer be the parameter quantity or multi-modal capability of the model, but a measurable and operable indicator, such as the "trust elasticity coefficient" of the organization.

It measures to what intensity of business uncertainty, employees, customers, regulators and the capital market still maintain well-founded confidence in the operation of AI systems.

The improvement of this coefficient cannot rely on rounds of internal publicity or technical upgrades. It needs to be deliberately accumulated and audited as a brand new factor of production.

I even judge that in the near future, a new C-suite role will emerge, perhaps the Chief AI Trust Officer, who is specially responsible for monitoring the compliance rate of interpretability indicators, the trend of adoption depth curve, the responsibility closure rate of key decision nodes, and incorporating these data into the board review cycle with the same frequency as financial reports.

This means that trust management will step out of the narrow lane of moral advocacy and compliance inspection, and enter the core area of enterprise value creation.

Now, those enterprises that are still rolling out AI systems based on speed intuition may be able to deliver a beautiful deployment coverage report in the short term. But when the track shifts from "whether there is AI" to "whether AI is used in depth", their organizational resilience will face severe tests.

Those enterprises that have already started to build institutional public goods, straighten out responsibility attribution, and tear open the lemon market with transparent signals are now quietly widening a hard-to-catch advantage gap.

The generation gap of technology can be easily bridged with capital, but the gap of trust can only be filled inch by inch with systems.

Therefore, the success or failure of enterprise AI strategy is not written on the competition list of model accuracy, nor engraved on the scale of computing power infrastructure. It is written in the moment when every ordinary employee faces the decision suggestion pushed by AI next time, and is willing to solemnly click "adopt".

Making that moment expectable and deliverable is exactly where all institutional construction about trust will eventually lead.

This article is from the WeChat official account "Dongzhen Strategy", the author is Dongzhen Strategy, and 36Kr publishes it with authorization.