How should enterprises make their AI investment plans for 2027? Are you still making pie-in-the-sky promises to your boss?

The end of the year is just around the corner again.

At this time in previous years, the biggest headache for CIOs was figuring out how to calculate the IT budget for the next year. This year is different: they have to add the planning of AI investment for 2027 to the agenda.

Lao Yang recently communicated with several CIO friends from different enterprises, and found that compared with previous years, the atmosphere this year is rather weird. The management has been bombarded by various press conferences and self-media content throughout the year, and their expectations for AI have been pushed to the maximum. Business departments have also tasted the benefits of hands-on use of tools like WorkBuddy, and their appetite is growing. If the information department caught in the middle still writes in its 2027 plan that "we will introduce large models, build a middle platform, and embrace AI-native", Lao Yang can bet that this plan will be 90% likely to be directly rejected by the management.

In Lao Yang's view, enterprise AI implementation in 2027 has entered the deep water zone. Management pays more attention to the input-output ratio, which means the management's purse strings are tightened. Anyone who tries to fool others with concepts will probably be eliminated first. Therefore, the planning must shift from "technology-driven" to "business-driven".

I. How to make the 2027 plan: do not seek oversize, but pursue precision

Under the current market environment, Lao Yang believes that the enterprise digital budget for 2027 needs to be carefully managed. The budget is divided into two major parts, as detailed below:

The first part: securing the foundation

The computing power infrastructure in 2027 must be optimized where necessary. Data governance must be carried out no matter how difficult it is, and the security and compliance line must be firmly established. If the current trend continues, with the in-depth application of AI agents like workbuddy, the out-of-control shadow IT of business departments will happen sooner or later, and the information department will still be the one to take the final responsibility. Data governance is not a task that can be completed in one or two years. If the data quality does not meet the standards, any AI tool you use will be useless. Security and compliance under AI is not a slogan, it must be written into the hard clauses of the plan, clearly stating who is responsible for construction, who is responsible for review, and who is accountable for accidents, in black and white, and it must be enforceable.

The second part: generating tangible business value.

That is, to create visible value. The biggest fear for enterprises when doing AI projects is to pursue comprehensiveness and want everything, but end up with nothing. Therefore, Lao Yang suggests that in 2027, AI landing scenarios should only focus on 1 to 3 core scenarios that can bring clear business increment or significantly reduce costs, such as intelligent quality inspection and production scheduling in the manufacturing industry, and intelligent credit review in the financial industry.

In particular, the cost of computing power investment in 2027 is also the key focus of the budget. Do not be keen on buying expensive GPUs to build large models. Do not make AI assets too heavy. Change the strategy to make computing power like tap water, pay for what you use, and turn heavy assets into light operations. Of course, for some special industries, such as industries with highly confidential data and high compliance requirements, it is still necessary to build their own systems.

The most notable point is: it is best to do as few so-called AI-native reconstructions as possible. It is an objective fact that old systems cannot be connected and old business logic cannot be migrated. Therefore, CIOs must clearly state in the plan that they should use means such as MCP protocol and AI adaptation layer to attach large models externally to activate the existing old ERP and CRM systems. Do not rush to reconstruct at every turn. The larger the reconstruction scope, the faster the project will fail.

II. How to evaluate AI performance: do not count Tokens, calculate the output

How to assess the implementation of enterprise AI in 2027? This is the most headache problem for most enterprises at present, and also the most prone to disputes.

In Lao Yang's view, if an enterprise assesses AI performance based on Token call volume, it is an extremely stupid act of lazy management. AI is not utilities like water, electricity and gas, and it is not the case that the more you use, the better. If the assessment is really based on Token usage, in order to meet the standards, business departments will treat large models as playgrounds, let AI write nonsense every day, and frantically generate useless data in the test environment. In the end, the computing power bill will skyrocket, and all the output will be worthless digital garbage.

The problem is that if we only assess the output effect, it is extremely difficult to quantify in the early stage. Because the value of AI in many scenarios is to help people do their work more smoothly, not to directly replace a position. If the purpose of an enterprise leader to promote AI is to lay off employees, this project will never succeed.

So how to assess the AI implementation and application in 2027?

Lao Yang's suggestions are as follows:

At the bottom layer, make computing power cost transparent. The IT department builds the "water plant", and the business department buys the "water". The Token consumption is directly included in the operation cost of the business department. Do not assess the usage volume, use financial leverage to force them to be cost-conscious.

At the middle layer, process indicators. Focus on AI penetration rate and business adoption rate: what percentage of customer service problems are pre-processed by AI? How much code generated by R&D with AI is adopted? How much time does the finance department save on account reconciliation with AI?

At the top layer, business ROI, which is only assessed for core projects. For example, after the AI production scheduling goes online, how many percentage points has the inventory turnover rate increased, and how much has the energy consumption decreased. Do not assess ROI from the very beginning, which will only force the using departments to create superficial gimmicks. It should be noted that the real core business processes, such as intelligent supply chain scheduling and R&D simulation, have long cycles and high risks, and the benefits cannot be clearly calculated in the first two years. Therefore, Lao Yang believes that giving them an observation period is better than forcing them to generate false data.

To sum up in one sentence: technical indicators step back to the background, and business indicators move to the foreground.

III. Two "invisible projects" cannot be saved on costs

In addition to technology and scenarios, Lao Yang believes that two shortcomings must be made up in 2027.

First: organization and AI literacy

It should be noted that the biggest resistance to enterprise AI implementation comes from people. Therefore, CIOs must clearly state in the plan: how to carry out hierarchical training, and how to adjust performance assessment, so that employees are willing to input their practical experience into AI. Training is most likely to become a mere formality. Holding a general meeting for all staff to talk about the AI prospects for two hours is far less effective than letting business backbones participate in a real project for two weeks. At the same time, in 2027, enterprises need to introduce from outside or internally cultivate business-savvy "translators", that is, Frontier Deployment Engineers, FDE. These talents are more precious than scarce computing power now, and cannot be recruited from the market, so they need to be cultivated slowly from inside the enterprise, which CIOs need to specify in the plan.

Second: governance and security compliance

From the current trend, it has become an inevitable trend for business departments to develop shadow IT. Therefore, CIOs must clarify in the 2027 plan: how to build an enterprise-level AI sandbox, how to desensitize sensitive data and add watermarks, and how to build a full compliance defense line covering pre-access review to post-event audit.

IV. How can the information department keep its position

In the general environment of cost reduction and efficiency improvement, if the plan of an enterprise's information department is still limited to "buying software, implementing projects, and repairing computers", it is very likely that IT will be among the first batch of departments to be optimized in 2027. Because these tasks can also be done by business departments with AI at a lower cost.

Therefore, if you do not want to be optimized, the information department must complete three role transformations.

1. Transform from a builder to an architecture and governance manager

Business departments want to build their own Agents? That's fine, IT will not stop it, but it must run on the security base provided by IT. IT holds the company's core data sources, API interfaces and identity permissions. As long as you control the data export and the compliance bottom line, you will be the unavoidable gatekeeper. But a gatekeeper is not a blocker. After IT sets up the base and rules, business departments can operate freely within the scope, but any operation that crosses the boundary must be blocked. This is delegating power with a bottom line, not a one-size-fits-all rejection. What the management wants is controllable innovation, not a stagnant pool where no one dares to make any changes.

2. Transform from self-satisfied technology pursuit to business partnership

Stop reporting to the management how many billions of parameters the large model you used has, and report how much money you have saved for the business and how much efficiency you have improved. When you can use AI to help the management make profits, no one will dare to lay you off.

3. Precipitate core capabilities into the enterprise brain

Stop being assembly line workers who only receive requirements and write CRUD codes. Encapsulate the company's core business logic into standardized interfaces that can be called by AI. When business departments need to build new applications, they find that all the underlying components they call are the things you have encapsulated, and you will become the hub of enterprise intelligence.

It is worth noting that role transformation is not a slogan. If the 2027 plan does not clearly state who leads each step, what the pace is, and how to accept the results, you will still stay where you are at the end of the year.

Final conclusion:

The 2027 AI plan is a major exam. It tests not only the CIO's judgment on technology trends, but also the insight into the essence of enterprise management.

Abandon those flashy concepts and return to business common sense. The IT team that can make money and save money with AI, and also hold the security bottom line in the chaos, will not only avoid being optimized in this wave, but also become the core engine driving the enterprise forward.

This article is from the WeChat official account "Xiangjiang Digital Review" (ID: benpaoshuzi), written by Lao Yang, and authorized for release by 36Kr.