Business leaders are being told that AI agents will transform marketing, sales, and operations.
The promise is compelling.
An agent can research a prospect, review CRM history, prepare an email, update a record, create a task, analyze campaign results, and coordinate work across multiple applications. Some agents can now operate software through the same buttons, menus, and fields a person uses. Others can divide a complex assignment among specialized agents and bring the results back together.
These capabilities are real. They are improving quickly.
But they also create a tempting—and costly—assumption:
Once we deploy the right AI agent, it will fix the gaps in our marketing and sales process.
It will not.
An agent can accelerate a well-designed revenue system. It can also accelerate bad data, unclear positioning, inconsistent follow-up, disconnected technology, and weak decision-making.
That distinction matters, especially for small and mid-market companies that cannot afford to spend months experimenting with technology that creates more activity without creating more revenue.
The most important question is no longer whether AI agents are capable.
They are.
The more important question is whether your business is organized well enough for them to be useful.
AI agents are now capable enough to matter
Until recently, most business conversations about generative AI centered on content.
Write an article. Summarize a meeting. Draft an email. Generate a social post. Rewrite a proposal.
Those tools helped people produce work faster, but the person still managed the process.
AI agents change that relationship.
An assistant responds to a prompt. An agent can interpret an objective, decide which steps are needed, select tools, retrieve information, take actions, inspect the results, and continue until it reaches a stopping point.
OpenAI describes modern agents as applications that can plan, call tools, collaborate across specialists, and maintain enough state to complete multistep work. Its current agent infrastructure can also place agents in controlled computer environments where they inspect files, run commands, edit code, and work through longer assignments.1
Anthropic makes a useful distinction between workflows and agents. In a workflow, the steps are largely predetermined. In an agentic system, the model has more freedom to decide how it will complete the task. Anthropic’s experience building production systems suggests that the most successful implementations often begin with simple, composable patterns rather than elaborate autonomous systems.2
That point deserves more attention.
Businesses do not need the most autonomous agent they can build. They need the simplest agent that can reliably improve a meaningful process.
What agents can do today
The current generation of agents can perform real operational work, particularly when the task has a clear objective, reliable information, appropriate tools, and a result that can be checked.
Research companies, markets, and accounts
Agents can search public information, internal documents, databases, files, and connected business applications.
A well-designed research agent can:
- Investigate a prospective customer
- Summarize its business model
- Identify relevant industry developments
- Compare the company with an ideal-customer profile
- Review previous interactions
- Find potential operational needs
- Prepare a briefing for sales
For broad research assignments, several agents can work in parallel. One might examine the company, another its industry, another recent financial or operational developments, and another the competitive environment.
Anthropic has documented a research architecture in which a lead agent delegates separate lines of investigation to specialized subagents and then synthesizes their findings.3
This does not mean every prospect should receive an expensive, multi-agent investigation. It means businesses can now match the depth of research to the value and complexity of the opportunity.
Qualify and route leads
Agents can review form submissions, company attributes, engagement history, CRM information, website behavior, and documented qualification criteria.
They can then:
- Categorize the inquiry
- Identify the likely service need
- Assign a priority
- Flag missing information
- Recommend the next step
- Route the record to the appropriate person
- Prepare a summary for review
This can remove a meaningful amount of administrative work from sales teams.
But the agent must be given an actual definition of a qualified lead.
“Find the best opportunities” is not a business rule.
A useful qualification process needs criteria such as company size, industry, geography, operational need, urgency, buying authority, strategic fit, previous engagement, or expected account value.
An agent can apply those rules consistently. It cannot rescue an organization that has never defined them.
Prepare personalized outreach
Agents can combine account research with approved messaging, service information, CRM history, case studies, and campaign goals.
Instead of inserting a company name into a generic template, an agent can prepare outreach around a relevant business condition.
For example, it might recognize that:
- A manufacturer is expanding into a new region
- A finance team is hiring heavily for receivables roles
- A company recently changed leadership
- A prospect is using a platform your business supports
- An existing customer has not adopted a service that solves a known need
That context can make outreach more useful and less intrusive.
The agent should still work within clear boundaries. It should know which claims are approved, which services apply, which examples are relevant, and which subjects require human review.
Personalization without control is simply a faster way to send the wrong message.
Update CRM records and execute workflows
When agents are connected to business applications, they can do more than recommend actions.
Depending on their permissions, they can:
- Create and update records
- Add research and meeting summaries
- Change classifications
- Assign owners
- Create follow-up tasks
- Identify duplicate or incomplete data
- Retrieve account history
- Start an approved workflow
- Escalate an exception
- Prepare a report
- Pause for human approval
This is where agents begin to provide value beyond personal productivity.
They become part of the operating process.
Imagine an inbound inquiry arriving through a company website. A connected agent could:
- Review the submission.
- Research the company.
- Check the CRM for existing contacts and opportunities.
- Compare the inquiry with qualification criteria.
- Identify the most relevant service.
- Draft a response.
- Create a follow-up task.
- Recommend an owner.
- Ask a salesperson to approve the communication.
- Record the final action in the CRM.
That is not a chatbot. It is coordinated revenue work.
Operate software through the user interface
Some agents can now interact with software visually. They can inspect screenshots, click buttons, enter information, scroll, and navigate menus.
This matters because many businesses still depend on systems that do not have complete APIs or modern integration options.
Computer-use capabilities can help bridge some of those gaps. However, they are generally less dependable than a direct API connection.
Interfaces change. Pop-ups appear. Sessions expire. Fields move. A visual action that worked yesterday may fail after a software update.
Computer use is valuable, but it should be treated as one available tool—not as evidence that an agent can safely operate every application without oversight.
Coordinate specialized agents
Agent interoperability is also improving.
Google’s Agent2Agent protocol is designed to allow agents to discover one another, exchange information, delegate work, and coordinate actions across platforms. The protocol has since moved into a broader open-source ecosystem supported by multiple major technology companies.4
This points toward a future in which a company may not have one universal agent.
It may have:
- A research agent
- A CRM agent
- A lead-triage agent
- A content agent
- A reporting agent
- A customer-support agent
- A compliance or approval agent
Each agent may have different tools, instructions, permissions, and performance expectations.
The strategic opportunity is significant.
So is the potential for chaos if those agents are working from different definitions, different records, and different versions of the truth.
The problem is no longer access to AI
Most businesses can now access capable AI.
The harder problem is preparing the business around it.
An agent needs more than a well-written prompt. It needs usable context.
That context may include:
- The company’s positioning
- Ideal-customer definitions
- Approved brand language
- Products and services
- Qualification criteria
- CRM records
- Customer history
- Sales stages
- Website behavior
- Pricing rules
- Standard operating procedures
- Approval requirements
- Examples of acceptable work
- Clear permission boundaries
Many organizations possess this information, but not in a form an agent can reliably use.
Some of it lives in CRM fields. Some is buried in old proposals. Some is stored in individual inboxes. Some exists only in the experience of a salesperson or company founder. Some is contradicted by the website. Some has not been updated in years.
That is not primarily an AI problem.
It is an operating-system problem.
An agent cannot use knowledge your business has never organized
Business leaders often assume that adding more information will make an agent more accurate.
That is not always true.
Context is limited, and poor context can be as damaging as missing context. Conflicting instructions, outdated documents, overlapping tools, unnecessary data, and vague descriptions make it harder for an agent to determine what matters.
Anthropic’s guidance on context engineering emphasizes carefully selecting the information available to an agent rather than loading everything into its working context. It also recommends designing tools with clear purposes and boundaries so agents can choose the right action.5
This has a direct business implication:
Your company does not need to give an agent everything it knows.
It needs to give the agent the right information for the work it has been assigned.
A lead-qualification agent does not need access to every company document. It needs accurate qualification rules, relevant CRM data, approved customer definitions, and a clear escalation path.
A content agent does not need permission to modify financial records. It needs positioning guidance, subject-matter sources, editorial standards, and a defined publishing process.
A reporting agent does not need authority to send customer communications. It needs trustworthy data, agreed metrics, and rules for identifying anomalies.
The clearer the role, the more useful the agent can become.
Disconnected systems create disconnected agents
Consider a company that wants an agent to improve inbound lead response.
The agent may be able to read the form, research the company, check the CRM, prepare a response, and create a sales task.
But what happens when:
- The form does not collect the information sales actually needs?
- Campaign attribution is missing?
- The CRM contains several versions of the same company?
- Lead stages mean different things to different employees?
- Salespeople do not consistently record their activity?
- Website messaging no longer reflects the company’s actual services?
- No one has defined what makes an opportunity qualified?
- The agent cannot determine whether someone already contacted the prospect?
- Different departments use conflicting customer categories?
The agent does not repair those problems.
It inherits them.
Worse, it may make the process appear more sophisticated while producing unreliable decisions underneath.
A polished summary is not proof that the underlying information was correct.
A personalized email is not proof that the prospect was a good fit.
An automatically updated CRM field is not proof that the field should have been changed.
Fluency can make weak processes harder to see.
That is why agent implementation should not begin with the question:
Which AI platform should we buy?
It should begin with:
Which business process are we trying to improve, and is that process clear enough for an agent to participate in it safely?
Your revenue system is the agent’s operating environment
For an agent to support revenue growth, it needs a connected system around it.
That system does not need to be perfect. Most businesses do not have the time or budget for a complete technology transformation before testing AI.
But the part of the system assigned to the agent must be understandable, measurable, and controlled.
Positioning gives the agent direction
An agent needs to understand:
- Who the company helps
- Which problems it solves
- How its offerings differ
- Which claims are defensible
- Which customers are a poor fit
- What language the brand uses
- What the company should never promise
Without this foundation, AI-generated personalization often becomes generic language wrapped around superficial research.
The agent may mention the prospect’s industry or recent announcement, but it still will not understand why your company is relevant.
The website defines the conversion path
Your website teaches both people and machines how the business is organized.
It should help a visitor understand:
- Whether the company serves businesses like theirs
- Which problem each service addresses
- What the next logical action is
- What evidence supports the company’s claims
- What information should be provided when requesting help
If the offer is confusing, an agent will not make it clear.
If the form creates unnecessary friction, an agent cannot recover the leads that never submitted it.
If service pages attract the wrong audience, automating follow-up may only help the company process poor-fit leads more efficiently.
CRM structure gives the agent memory
The CRM should provide an understandable record of the relationship.
That includes:
- Contact and account information
- Source and campaign data
- Lifecycle stage
- Ownership
- Activity history
- Qualification status
- Opportunity details
- Next steps
- Relevant customer context
An agent also needs to know which fields are authoritative and which are optional, outdated, calculated, or unreliable.
Poor CRM data does not become better because an AI system can read it faster.
Revenue operations give the agent rules
Agents work best when the business process has defined decision points.
For example:
- What happens when a lead arrives?
- Who owns it?
- What makes it qualified?
- When should it be escalated?
- Which follow-up sequence applies?
- When should a person intervene?
- What information must be recorded?
- What happens when the agent is uncertain?
These do not have to become rigid scripts.
They do need to be clear enough that employees—and agents—can make consistent decisions.
Analytics show whether the agent helped
Agent activity should not be mistaken for agent value.
An agent may conduct hundreds of searches, update thousands of fields, or produce dozens of reports. None of that proves it improved the business.
A useful deployment should be tied to an operational or revenue outcome, such as:
- Faster lead response
- Less manual research
- Higher CRM completeness
- Better routing accuracy
- More consistent follow-up
- Increased qualified-meeting conversion
- Reduced support resolution time
- Improved sales preparation
- Fewer missed opportunities
- Lower cost per completed process
Start with a baseline. Then measure the change.
Otherwise, the company may celebrate how much the agent did without knowing whether any of it mattered.
Human approval is a design choice—not a failure of automation
The conversation around agents often treats full autonomy as the goal.
That is the wrong goal for many business processes.
The appropriate level of autonomy depends on the cost of being wrong.
An agent may be allowed to act independently when it:
- Classifies an inquiry
- Summarizes a conversation
- Adds research to a record
- Identifies missing information
- Prepares a task list
- Drafts a response
- Flags an unusual metric
Human approval may still be appropriate before it:
- Sends an external message
- Changes an opportunity stage
- Publishes content
- Changes pricing
- Deletes information
- Issues a refund
- Commits company resources
- Makes a legal or contractual statement
- Shares confidential data
This is not unnecessary friction.
It is responsible system design.
The best human-in-the-loop processes do not force employees to redo the agent’s work. They ask the person to review the small number of decisions where judgment, authority, or risk matters most.
The agent handles the preparation. The person owns the decision.
Security becomes more important when AI can act
A system that generates text creates one level of risk.
A system that can retrieve data, use credentials, communicate with customers, modify records, and activate other tools creates a different level of risk.
NIST has launched an AI Agent Standards Initiative focused on agent identity, authorization, interoperability, and secure adoption. NIST research has also demonstrated how malicious instructions hidden in websites, emails, repositories, and other external content can redirect an agent toward unintended actions.6
OWASP’s guidance for agentic applications identifies risks such as agent goal hijacking, tool misuse, excessive privileges, memory poisoning, insecure agent communication, cascading failures, and exploitation of human trust.7
These concerns should not prevent businesses from using agents.
They should shape how agents are deployed.
Practical controls include:
- Giving agents only the access required for their role
- Separating read permissions from write permissions
- Requiring approval for consequential actions
- Restricting which tools an agent may use
- Logging actions and decisions
- Testing the agent against realistic failure scenarios
- Reviewing external content before it influences sensitive actions
- Providing a clear way to stop or revoke the agent’s access
- Monitoring performance after deployment
The agent’s potential impact should determine the strength of the control.
Agents are improving quickly, but reliability still varies
Agent capability is advancing at an unusual pace.
METR measures how long AI agents can work on software tasks before their reliability declines. Its research shows a strong upward trend in the length and complexity of tasks frontier systems can complete.8
That progress is meaningful, but it should not be interpreted too broadly.
Software tasks often have clearer instructions and more objective success criteria than real marketing and sales work.
A program either runs or it does not. A test passes or it fails.
A sales opportunity is less tidy.
The prospect may have incomplete information. The stated need may not be the actual need. Several people may influence the decision. The best next action may depend on history that was never recorded. The technically correct answer may still be inappropriate for the relationship.
Agents tend to perform best when the work is:
- Well-scoped
- Supported by reliable data
- Broken into understandable steps
- Connected to appropriate tools
- Evaluated against a visible outcome
- Escalated when uncertainty is high
This is why agents should be introduced process by process—not as a vague companywide mandate to “use AI.”
Start with the process, not the platform
For most small and mid-market businesses, the right first agent is not the most impressive one.
It is the one that solves a specific, recurring, measurable problem.
Strong starting points may include:
- Researching inbound companies before sales follow-up
- Summarizing calls and creating CRM tasks
- Detecting incomplete lead records
- Categorizing support inquiries
- Preparing campaign performance summaries
- Identifying overdue follow-up
- Drafting personalized—but reviewable—outreach
- Comparing new opportunities with qualification criteria
- Routing inquiries to the right employee
- Monitoring a defined set of business signals
These processes are narrow enough to control and valuable enough to measure.
They also create an opportunity to learn what the organization needs before giving agents broader responsibility.
An AI agent readiness checklist
Before introducing an agent into a marketing, sales, or revenue process, answer six groups of questions.
Strategy
- What business problem are we solving?
- What result should improve?
- Is the target customer clearly defined?
- Is our positioning consistent?
- Do we know what the agent should not do?
Process
- Is the existing workflow documented?
- Are ownership and handoffs clear?
- Are the important decision points defined?
- Is there an escalation path?
- Can we describe what a successful outcome looks like?
Data
- Is the required information available?
- Is it accurate enough to support decisions?
- Do we know which source is authoritative?
- Can the agent distinguish current information from outdated material?
- Are critical CRM fields consistently populated?
Technology
- Can the agent securely access the systems it needs?
- Is an API available for actions that require reliability?
- Are tools clearly named and separated by purpose?
- Can the system preserve state across steps?
- Can the agent stop and request help?
Governance
- What can the agent read?
- What can it create or change?
- Which actions require approval?
- How will its activity be logged?
- How can access be revoked?
- What happens when the agent encounters suspicious instructions?
Measurement
- What is the current baseline?
- Which metric should change?
- What error rate is acceptable?
- How will outputs be reviewed?
- How will we detect declining performance?
- Does the value justify the technology and oversight required?
These questions are not intended to slow AI adoption.
They prevent a company from automating a process it does not yet understand.
The real competitive advantage is not the agent
AI agents are becoming broadly available.
That means access alone will not create a lasting advantage.
The advantage will come from what surrounds the agent:
- Clear positioning
- Organized knowledge
- Reliable data
- Connected systems
- Defined workflows
- Appropriate permissions
- Human judgment
- Useful measurement
An agent placed inside that environment can reduce repetitive work, improve response times, strengthen follow-up, and give employees more time for decisions that require experience.
An agent placed inside a disconnected environment will produce more work around the same underlying problems.
That is the perspective business leaders need as agent adoption accelerates.
Do not ask AI to fix a revenue system that leadership has not yet connected.
Connect the system. Define the work. Establish the controls. Then give the agent a meaningful role within it.
That is how AI moves from an interesting tool to a dependable part of business growth.
Is your revenue system ready for AI agents?
Virbion helps small and mid-market businesses connect marketing strategy, AI visibility, website conversion, CRM, revenue operations, analytics, sales execution, and practical AI agent workflows.
The objective is not to add AI everywhere.
It is to identify where an agent can create measurable value—and make sure the surrounding system is ready to support it.
Request a Revenue Systems Review
Sources and further reading
- OpenAI, The Next Evolution of the Agents SDK, and Agents SDK documentation.
- Anthropic, Building Effective Agents.
- Anthropic, How We Built Our Multi-Agent Research System.
- Google Developers Blog, Announcing the Agent2Agent Protocol, and How A2A Is Building a World of Collaborative Agents.
- Anthropic, Effective Context Engineering for AI Agents, and Writing Effective Tools for Agents.
- National Institute of Standards and Technology, research and guidance on securing AI agent systems.
- OWASP GenAI Security Project, OWASP Top 10 for Agentic Applications for 2026.
- METR, Task-Completion Time Horizons of Frontier AI Models.



