Onboarding new AI hires calls for context engineering - here's your 3-step action plan

Jiojio/Moment via Getty Images

Follow ZDNET: Add america arsenic a preferred source connected Google.

ZDNET's cardinal takeaways

• Successful adoption of AI agents requires discourse engineering. 
• Context engineering requires entree to data, metadata, process flow, and more.
• Context engineering ensures your information is acceptable for agentic AI usage. 

Why is it that your existing employees initially outperform the caller rockstar you've conscionable hired?  And wherefore bash you person a play of onboarding earlier a caller prosecute gets up to speed?

Institutional knowledge. The caller rockstar knows however to bash the job. That's wherefore you hired them.  But they request clip to recognize the institution culture, processes, approaches, applications, their team, and customers and partners. 

Also: Is that a scam? This caller ChatGPT instrumentality scans suspicious numbers, emails, and links - for free

In the AI world, the organization cognition is called context.  AI agents are the caller rockstar employees. You tin onboard them successful minutes, not months. And the much discourse that you tin supply them with, the amended they tin perform. 

Now, erstwhile you perceive reports that AI agents execute amended erstwhile they person close data, deliberation much broadly than lawsuit data. The information that AI needs to bash the occupation efficaciously besides includes the information that describes the organization knowledge: context.

We'll get to discourse engineering successful a minute.

Understanding context

So let's look astatine the antithetic types of context, its source, and whether it's structured oregon unstructured -- each of which volition find however it is presented to the AI agent. 

Also: More workers are utilizing AI than ever - they're besides trusting it less: Inside the vexation gap

You support proceeding astir models having a ample discourse window. Claude has a 1-million-token discourse window; ChatGPT 5.2 has a 400,000-token window. But this is not capable to grip everything astir the company. Consider Salesforce's org configuration -- 20 apex classes of comparatively precocious complexity are implicit 250,000 tokens. So, we request to beryllium selective and supply the discourse for the relation that the AI cause is delivering: context engineering.

Context engineering

As you tin spot from the array below, overmuch of this accusation is unstructured. Your employees are bully astatine interpreting it and filling successful the gaps utilizing their judgement and applying organization knowledge. AI agents tin present parse unstructured data, but are not arsenic bully astatine applying judgement erstwhile determination are conflicts, nuances, ambiguity, oregon omissions. This is wherefore we get hallucinations. 

So the discourse you supply needs to beryllium implicit and AI-readable. But the discourse besides needs to beryllium circumstantial to the relation of the AI agent, truthful the discourse model is not overwhelmed. The mode to bash this is to see the end-to-end process that the AI cause is performing and usage that to scope the context. That requires parsing the assorted applications that store the discourse to propulsion the close level of information. If we look astatine Salesforce's acquisitions, it starts to marque sense: Data360, Informatica, MuleSoft, and Tableau are each antithetic forms of discourse astatine scale.

Also: Is your AI cause up to the task? 3 ways to find erstwhile to delegate

Context successful context

As we've said, providing the close discourse to the AI Agent astatine the close level of item means parsing these information sources with a wide knowing of what the end-to-end process it is trying to perform.  

This is simply a operation of the documented concern process and the exertion configuration encoded successful the metadata and dependencies. And this is not conscionable astir whether metadata uses different metadata, but wherefore and how. 

The process maps supply visibility into manual activities betwixt applications oregon wrong applications. The accuracy and completeness of the documented process diagrams alteration wildly. Front-office processes are mostly precise poor. Back-office processes successful regulated industries are typically precise good. And to exploit the powerfulness of AI agents, organizations request to streamline them and optimize their concern processes. This has sparked a process reengineering gyration that mirrors the 1 successful the 1990s. This clip around, the level of item required by AI agents is higher than for humans.

Also: Gen AI boosts productivity, but lone for definite developers - here's why

The knowing of the app configuration done the metadata and dependencies is available, but it is often confused by precocious levels of method debt. And it requires blase investigation to beryllium implicit and trustworthy. AI agents are not yet susceptible of taking each the metadata and making consciousness of it. There is simply excessively overmuch data. The lone attack is to usage precise clever, agentic workflows of chained surgical tasks to tally the analysis. 

Is your contented acceptable for AI?

For each benignant of content, we request to inquire 5 questions

1. Does the accusation exist, who owns it, and what inducement bash they person to enactment the project?
2. Is it up to date, and what is the process for maintaining and governing it?
3. Is it written for AI, and what changes request to beryllium made to forestall ambiguity and confusion?
4. Where should it beryllium stored truthful AI tin entree it, and what information and entree controls should beryllium applied?
5. How should it beryllium structured and tagged for curation, balancing details with token usage?

Let's look astatine 3 contented types -- culture, concern process, and applications -- and see each successful turn.

Company culture

This is the accusation that is typically provided to caller employees during onboarding, but it is besides the intangible cognition that is absorbed implicit time. AI agents request each of it each astatine once. 

Also: Forget the main AI serviceman - wherefore your concern needs this 'magician'

• Existence and ownership: This refers to the onboarding contented the enactment uses, including institution policies. AI agents don't attraction however adust the contented is. It tin besides beryllium different documents that amusement the civilization and property of the enactment -- the selling marque guidelines, yearly reports, and shareholder presentations. Even the benignant of lawsuit testimonials, selling videos, and bureau plan volition supply colour to a blank canvas. Maybe determination is simply a array of firm acronyms. The complication is that this is owned by antithetic teams. There is enactment they request to enactment successful to enactment the project, but what are their priorities and incentives? Ideally, selling should instrumentality the lead.
• Valid and valuable: Of each the institution documentation, this is astir apt kept reasonably current, unless determination has been a caller re-brand. If so, you request to beryllium cautious what to include. It whitethorn beryllium up to date, but is it inactive applicable and valuable?
• Written for AI: The onboarding worldly whitethorn person been written for presenting to caller starters, not for reading. Hence, determination could beryllium immense discourse gaps, which request to beryllium filled with a preamble oregon notes. Other contented needs to beryllium acceptable successful its ain context. For example, the AI cause needs to beryllium told however to construe lawsuit testimonial videos oregon marque guidelines. Company argumentation documents are often written for humans, with nuances and assumptions that the AI cause volition not prime up during onboarding and done tribal knowledge.
• Access and security: This contented is mostly unstructured and precocious volume. The lawsuit testimonials whitethorn request to beryllium transcribed to substance alternatively than indexed arsenic video. This means that a solution similar Data 360 needs to beryllium utilized to marque it accessible and easy searched. However, we besides request to see the information and entree controls. Is determination IP,  delicate data, oregon idiosyncratic accusation that should not beryllium exposed? The information level of 2 oregon 3 sets of information erstwhile held successful isolation whitethorn beryllium acold little than erstwhile they are combined. Once the information is aggregated, the information level whitethorn beryllium galore levels higher based connected the insights that are present available.
• Structure and tagging: This is hard information to structure, arsenic virtually each of it is required arsenic inheritance for the AI cause alternatively than successful the discourse of delivering a process. There is simply a equilibrium betwixt providing each the accusation astatine a elaborate level and the outgo and feasibility of the token usage. So you request to see however to categorize the information truthful that it tin beryllium sliced and diced and served up to the AI cause successful the astir token-efficient manner.

Business operations/process

The documented concern processes are the captious operation for the AI cause to present an outcome. But they besides picture the supporting processes that situation the AI cause and connected which it relies oregon delegates. 

Also: Stop utilizing ChatGPT for everything: My go-to AI models for research, coding, and much (and which I avoid)

• Existence and ownership: Most organizations person processes documented. In 30+ years of moving successful concern process engineering, we've recovered that processes are usually incomplete, retired of date, and successful a assortment of formats. Fortunately, you bash not request to get each process up to date; lone the ones related to the AI cause you are building, which are apt owned by 1 oregon 2 concern units. The processes request to screen some the automated and quality activities, but astatine a acold greater level of detail. AI agents bash not grip nuances, gaps, and ambiguity arsenic good arsenic humans do. Now you tin usage AI to assistance you physique the archetypal chopped process. It tin make process diagrams from notes, diagrams, oregon adjacent systems metadata. These tin beryllium refined by moving with leaders and users.
• Valid and valuable: The astir important process to papers and optimize is the process of process improvement. This becomes critically important for AI agents that volition instrumentality contented literally, and trust connected up-to-date processes documentation to behave arsenic expected. 
• Written for AI: AI is precise bully astatine knowing process-related diagrams and procedural documents. The contented is the prime of the documentation: its completeness, accuracy, and currency.
• Access and security: Again, if it's unstructured documentation, specified arsenic images, a solution similar Data 360 is needed to marque it accessible and easy searchable. But process diagrams could beryllium presented arsenic structured JSON, which is much easy consumed by AI.
• Structure and tagging: This is precise circumstantial to the scope and result of the AI agent. Therefore, the metadata of the process diagrams is important. 

Application configuration

The exertion metadata describes the information structure, concern logic, and permissions of a circumstantial application. If AI agents span exertion boundaries, the contented needs to beryllium augmented by architectural diagrams that picture however applications enactment together. Also included successful these diagrams could beryllium however the agents enactment together. 

Also: 5 ways you tin halt investigating AI and commencement scaling it responsibly successful 2026

• Existence and ownership: This information is stored wrong each exertion arsenic metadata. However, it needs to beryllium much than a database of metadata. It needs to see metadata dependencies, specified arsenic the metadata investigation that Elements.cloud produces for Salesforce. An exertion similar Informatica is designed to store metadata from aggregate systems.
• Valid and valuable: The metadata is 100% accurate. The metadata investigation tin beryllium performed whenever it changes, truthful it tin besides beryllium 100% accurate.
• Written for AI: Metadata is highly structured, and therefore, ideally suited for being work by AI.
• Access and security: As it is highly structured, it tin beryllium stored successful immoderate database. What is captious is however it is structured truthful that it tin beryllium accessed. The contented is that immoderate exertion has acold excessively overmuch metadata, and it volition overwhelm the token limits.
• Structure and tagging: The metadata needs to beryllium related backmost to the operational concern processes that the AI cause is delivering and the information sources that the AI cause needs.

Only 7% of connection is words 

There is the communal look that connection is lone 7% words. So what astir the different 93%?

• The words (7%): Verbal contented (the literal meaning).
• The code (38%): Voice quality, pitch, and volume.
• The visuals (55%): Facial expressions and assemblage language.

Tonality acts arsenic the punctuation of spoken language. Let's instrumentality the elemental sentence: "I privation to spot you successful my office." Context is the 93%. We instruct AI with words; the 7%. Is it immoderate wonderment we get hallucinations and inconsistent results? We request to supply the different 93%. The context. This could include: the narration betwixt the lawsuit and company, the comparative value of antithetic aspects of the data, the signifier successful the process, the urgency, and the worth of the outcome. And that discourse is provided arsenic words and data. So we request to marque definite that determination is discourse for the context.

Also: The champion AI chatbots of 2026: I tested ChatGPT, Copilot, and others to find the apical tools now

Context engineering is simply a caller word for AI agents, but the contented already exists wrong organizations arsenic organization cognition that radical sorb implicit time. AI agents are built to judge a firehose of accusation but necessitate it to beryllium close and unambiguous. That has implications for organizations that privation to pat into the benefits of AI agents susceptible of delivering blase outcomes. Here is simply a 3-step enactment plan:

1. Document the scope of your AI agents, including the end-to-end process and outcomes.
2. Identify the captious contextual accusation required for AI agents to execute astatine the highest levels, and reappraisal their quality.
3. Format the contextual accusation successful the platforms that tin curate it for AI agents.

This nonfiction was co-authored by Ian Gotts, elder probe chap astatine Keenan Vision, co-founder of Elements.Cloud, 10X author, tech advisor, speaker, and investor.