跳到正文
MarTech.org· Susan Ferrari·· 4 天前AI 评分40

最聪明的 AI 也需要结合企业自有业务知识

Even the smartest AI needs your business knowledge

AI 导读

企业落地 AI 的关键不在于等待更先进的模型,而在于让 AI 智能体调取内部沉淀的客户反馈与业务知识解决实际问题。黄仁勋在对话中指出,依托 Box 上下文引擎可使智能体的 token 消耗减少高达 90%。同时,企业应为智能体划定清晰职责、工作流程与零信任访问权限,由人工保留关键审核与决策权。

正文 · 原文

Robots reading books.

Businesses have spent years collecting customer research, feedback, and other valuable information. Yet much of that knowledge sits in files nobody opens. Without access to that information, AI has little understanding of what makes your customers and business different.

For me, one of AI’s biggest opportunities is helping businesses put what they already know to work.

That matters more than deciding which AI model is smartest. ChatGPT, Claude, and Gemini will keep improving, but choosing a model isn’t the same as figuring out how to use it. A marketer still needs a question to investigate, information worth examining, and a reason to trust the result.

That was my biggest takeaway from a BoxWorks keynote conversation between former NVIDIA CEO Jensen Huang and Box CEO Aaron Levie. Huang argued that today’s models are already capable of useful work. Businesses don’t need to wait for the next breakthrough to find practical applications.

He compared this stage of AI to the early days of personal computing and the iPhone. People found uses for those technologies long before they reached today’s capabilities.

It reminds me of having a Ferrari in the driveway and debating next year’s engine instead of taking it for a drive.

Start with the customer knowledge you already have

Consider customer retention. Asking AI for general advice is one thing. Asking it to examine your customer feedback, support issues, and cancellation reasons from the past year is more useful. You can look for recurring problems and decide which deserve attention, with the evidence in front of you.

Much of that information is unstructured data — information that doesn’t fit neatly into spreadsheet rows or database fields. Finding relevant answers often means searching folders, digging through emails, and tracking down research.

I’ve spent more than 25 years in customer insights, analytics, and research. It’s frustrating to watch companies pay for research, then struggle to use it when a decision comes up. That’s the problem I want AI to help with.

10X your SEO with Semrush for Enterprise.

The world’s most powerful SEO platform, purpose-built for Enterprise.

Request demo

Find the relevant information

That’s why the discussion of context engines caught my attention. If you have a question for a librarian, you want the relevant books, not the entire library. An AI agent needs that kind of help too.

A context engine helps an agent find and use the relevant information without processing every document.

Huang discussed how Box’s context engine could reduce an agent’s token usage by up to 90%. I wouldn’t treat that keynote claim as a promise for every business. The result depends on the task, configuration, and information involved.

If an agent processes less irrelevant material, a task could take less time and cost less to run.

Huang also described NVIDIA’s computing advances as a way to lower processing costs and give AI more capacity for difficult problems. For businesses, the question is whether those changes make a particular use affordable.

Give agents a defined job

Huang described the model as the brain and the supporting system as the harness. The agent also needs company knowledge and access to tools. Without those, it can’t do much with what it knows.

An agent handling an order question could check the customer’s history and delivery status, then help resolve an issue. If an employee needs to take over, the agent could prepare the background so the customer doesn’t have to start again.

But I don’t want an agent inventing a process whenever it gets a request.

Take customer onboarding. Information must be collected, accounts created, appointments scheduled, and records updated. The business should define those steps and decide which actions require approval.

A kitchen and bath design company, for example, could use an agent to gather preferences before a consultation, organize project details, and prepare routine follow-up for review.

The designer still does the design work. The owner and employees still make decisions and approve what goes to customers. I’d want the agent to reduce the scheduling, searching, and follow-up that eat into their day.

Keep useful experience from disappearing

Huang also discussed institutional memory. When a team resolves a customer problem, a documented solution could help an agent handle a similar issue later.

That matters when an experienced colleague leaves. Too often, everyone knows that person had the answer, but nobody knows where to find it.

Someone still has to check that the information is accurate and current. An old policy doesn’t become useful just because an agent can retrieve it quickly.

Decide what the agent is allowed to do

I wouldn’t connect an agent to company files without first deciding what it should be allowed to see and do.

Huang and Levie discussed Box’s protections for business information and NVIDIA’s OpenShell approach to more secure agent environments. Huang emphasized zero trust.

I think of access in terms of the job. Someone hired in accounting doesn’t need every employee file or confidential contract. An agent shouldn’t have unrestricted access either.

Businesses need permissions, monitoring, and a way to stop an agent’s activity. They also need to decide when a person must review or approve an action.

Know what’s worth customizing

Huang argued for using general models while developing specialized capabilities around a company’s expertise.

My restaurant comparison is simple: Buy the oven. Protect the family recipe. A business doesn’t have to build every tool it uses, but it should know which knowledge makes it different.

He discussed open models as an option for customization. That doesn’t mean every small business needs engineers or a model of its own.

I’d start by asking what our team knows about customers that a general model won’t know, and where that knowledge would help us do a better job.

What would you do with the time you saved?

I liked Huang’s argument that productivity shouldn’t be measured only in time or money saved. It can also give a team room to try work it couldn’t previously afford to do.

If a marketing team usually has time to investigate three campaign ideas, AI assistance might let it examine more. I’d care about whether that extra work led to a better decision, not how many ideas the tool generated.

Huang used radiology as an example of how time saved could create room to serve more patients. That outcome isn’t automatic. Testing and monitoring still matter.

For a marketing team, I’d ask where the saved time went. Did we talk to more customers? Test an idea we would have skipped? Fix something customers kept complaining about?

Learning to use AI is part of the job

Huang was optimistic about people who learn to use AI, while acknowledging that jobs will change and some people will lose them.

I don’t want to brush past that. People need training and support, not just instructions to adapt. Helping nontechnical colleagues get comfortable with AI is a part of this work I care about.

You don’t have to become a programmer to begin. Choose a task you understand well enough to judge the result. Try the tool, check its work, and see whether it actually helps.

Where I’d start as a marketer

  • Pick a customer question you haven’t answered well. I’d start with a recurring frustration in feedback or a reason customers cancel. Give the work a purpose before choosing a tool.
  • Find the information that could answer it. Gather the relevant research and support records. Check dates and permissions. I’d rather work with a small set of useful, current sources than a folder full of everything.
  • Try one task and review the result. Ask for a summary of recurring issues or a follow-up draft. Compare it with the source material and decide what needs correcting before anyone acts on it.
  • Decide what to do with the time saved. I’d use it to speak with customers or test an idea we haven’t had time to explore. Producing more content wouldn’t be my default.

The work I want AI to help with

I’ve spent years helping businesses understand their customers. The hardest part often isn’t collecting information. It’s getting people to use it when they make a decision.

That’s why this conversation stayed with me. I want to see whether AI can help a team find the research, understand the complaint, and follow through on what it learns.

I don’t need another impressive demo. I want to see what changes for the customer.

The post Even the smartest AI needs your business knowledge appeared first on MarTech.

来源:MarTech.org · martech.org

EVAOO 主站实战落地 · AI Agent 与研发提效

看完情报?回到 EVAOO 主站直接上手工具与完整 SOP

返回 EVAOO 主站