How to build an AI agent: from goal to live deployment

Building an AI delegate is a multi-step process. Marketers need to define the agent’s task, nexus approved data and tools, and set limits on its actions. Once they expert the basics, anyone — marketers, revenue reps, or client teams — can build agents alongside ease.

Today, AI delegate acceptance is growing. A 2026 study of 5,000 U.S. adults from Menlo Ventures established that one in four AI users has used an agent, alongside 24% of respondents citing regular use.

To create the most of the tech, teams need to discover the correct use cases and commencement experimenting. This article covers how to build an AI delegate for those fair getting started. Discover how to deploy, test, improve, and set guardrails for your archetypal AI agent.

Table of Contents

• What is an AI agent?
• Is an AI delegate the correct fit for your workflow?
• How to Create an AI Agent
• How to Build an AI Agent Without Code
• Test, Evaluate, and Improve
• Add guardrails and individual handoff.
• How to Deploy an AI Agent to a Real Workflow
• AI Agent Use Cases for Marketing, Sales, and Service
• When to Consider Models, Protocols, or Multiple Agents
• Frequently Asked Questions About How to Build an AI Agent
• Start alongside one delegate your squad can depend on.

What is an AI agent?

An AI delegate is application that uses an AI example to activity toward a goal autonomously, or without individual input. Agents can choose among approved tools, assessment the results, and decide what to do next until they complete the project or attain a set limit.

Agents need pre-defined goals or functions. For a promotion team, that could be to assessment a campaign’s achievement and propose optimizations. The delegate would need to analyze conduit metrics, difference results alongside the run objective, acknowledge apt causes, and explain the recommended change.

Looking to build your archetypal agent? HubSpot users now have admission to a no-code AI delegate builder. Marketers can depict the automation they desire to create, and Agent Builder volition do the rest.

AI Agent vs. Chatbot vs. Rules-Based Automation: Explaining the Difference

Agents, chatbots, and workflows have distinct problem-solving capabilities and assist distinct functions:

• Chatbots react to queries or conversations, answering questions that the person inputs.
• Agents are capable to create decisions on their own and problem-solve independently.
• Rule-based automations execute a predefined project according to a set trigger. For example, if a buyer changes agreement stages, an email may be sent alongside a discount code.

Agents, chatbots, and automations can additionally activity together. A chatbot, for example, can act as the conversation opening for an agent, as fine as a predefined workflow can additionally contain an agentic step

The Four Building Blocks of an AI Agent

An productive tradition AI delegate needs a repeatable job, applicable data, a apparent starting point, and a measurable result, as defined by Elle Oshiro, a prof at HubSpot Academy. These four construction blocks assistance marketers rotate an idea into a workflow they can put to use.

1. Repeatable Work

AI agents excel at regular tasks that need to be performed consistently. For example, a promotion squad power assessment misplaced deals all duration to acknowledge buyer concerns. Even although all agreement contains distinct evidence, the agent’s job stays the same: analyze the concern, inspect existing assets, and propose a next step.

That said, choose a project that the squad can depict plainly and expects to perform complete and over.

2. Data Already in Your CRM

AI delegate betterment relies on high-quality training data. CRM data gives the AI environment it can use to create decisions. Relevant data power contain agreement properties, business details, revenue notes, and recorded client conversations.

Before construction an agent, inspect whether the necessary data exists, is current, and is accessible to the agent.

3. A Trigger That Fires Under the Right Conditions

A trigger determines whenever the delegate starts its work. It could be a manual request, a schedule, or a alter to a CRM record. With HubSpot, marketers and revenue ops can use any of these options through Agent Hub workflow triggers.

4. A Clearly Defined, Measurable Outcome

Define what the delegate must deliver, what a accurate outcome looks like, and how the squad volition measure success. For a run review, the output could be a study that flags underperforming channels and recommends one alter to test. Each advice should citation the applicable metrics, explain the rationale, and acknowledge data gaps.

Is an AI delegate the correct fit for your workflow?

AI agents are value a nearer appearance whenever a project is repeatable, tedious, and requires interpreting information and deciding what to do next. If all you need is a summary of content or data inspection you paste into a chat, a sole AI immediate may do the job. Even Anthropic recommends starting alongside the simplest method that meets the need, since agents can add disbursal and delay.

Three questions to ask before you build:

• Where does the project necessitate judgment?
• Can the delegate admission adequate data to do the job?
• Will it preserve activity following individual review?

When Rules-Based Automation is Better

Choose rules-based automation whenever the next steps are clear. Automations activity fine whenever outcomes prosecute one or multiple defined branches. Each stage of the procedure should prosecute a decision tree that the squad can define ongoing and automate.

Common examples for rule-based automation include:

• Assigning a guide to a local squad according to its nation field
• Creating a follow-up project following a demo request
• Notifying an document owner whenever a renewal date approaches
• Excluding unsubscribed contacts from a promotional email audience

Even a complex lead-routing procedure can activity fine alongside rules-based automation. A workflow power inspect territory, business size, and merchandise interest, afterward allocate the guide to an eligible rep. If nobody accepts a guide inside a set period, it can vigilant a director or reassign the contact.

How to Build an AI Agent Without Code

For teams learning how to set up an AI agent, commencement alongside a apparent goal. Define what an delegate has to do and its output. Then, compose apparent instructions for the delegate and test it before complete deployment.

Looking to create a tradition agent? HubSpot now offers an AI delegate builder. Just depict your goal in plain language, and Agent Builder volition do the rest. HubSpot users additionally have admission to pre-built agents that can assistance alongside tasks akin data management or prospecting.

Define a narrow goal and achievement criteria.

First, compose downward the manual project you desire the delegate to obtain over. For example, assessment misplaced deals and propose satisfied that addresses buyers’ unanswered questions.

Before you open the Agent Builder, put together a abbreviated concise for the task. That clarification should shield the agent’s goal, scope, and what you anticipation it delivers. See an example below.

• Goal: Help marketers acknowledge satisfied needs from documented buyer concerns.
• Scope: One selected closed-lost agreement and an approved satisfied catalog.
• Deliverable: A abbreviated concise alongside the concern, supporting evidence, and a advice to reuse, update, or create content.
• Reviewer: A marketer who checks the advice before anyone acts on it.

Note: Give your archetypal delegate admission to fair one deal, so reviewers have a manageable set of records whenever norm assessing the agent’s work.

Next, define your success criteria, or the checks you’ll use to fairness whether the delegate completed the project correctly. For this example, here’s what achievement looks like:

• Each buyer involvement has a origin citation you can verify.
• Any recommended existing asset appears in the approved catalog and addresses the concern.
• A projected satisfied concise states the buyer’s inquiry and what the satisfied should explain.
• The delegate flags missing evidence or problems that satisfied cannot solve.
• Include cases anywhere no satisfied advice is appropriate

Write apparent instructions.

Learning how to train an AI agent? Start by turning your notes into apparent instructions the delegate can follow. Explain how the delegate should analyze the evidence, choose a recommendation, and grip missing information.

For marketers using HubSpot, open the selected delegate in Agent Builder and assessment its existing instructions. Update the Instructions field to indicate the project you defined. Cover the role, goal, approach, and expected output.

Here’s a starting item for our example. Notice how these instructions explain what to do in distinct situations:

Role and goal: You assistance marketers acknowledge satisfied needs from one selected closed-lost deal.

Approach: Read the accessible agreement particulars and applicable connected records. Identify a documented buyer concern, afterward difference it alongside the approved satisfied catalog. Distinguish straightforward buyer feedback from a salesperson’s interpretation.

Recommend an existing asset whenever it answers the buyer’s question. If an asset lone partially answers it, define what needs an update. Propose new satisfied lone whenever the supplied catalog lacks a suitable asset.

Missing evidence: If the records don’t assistance a conclusion, explain what data is missing. If satisfied cannot location the concern, province why. Do not assume that missing satisfied caused the agreement loss. A petition for an unsupported integration have to be flagged as a merchandise limitation.”

Output: Return a abbreviated concise alongside four sections: Buyer concern, Supporting sources, Content recommendation, and Information gaps.

Add tools and permissions.

Depending on the job, an delegate power need to hunt the web, recover CRM records, or update client information. It may additionally need connectors to attain another apps, specified as Zoom, Appcues, or Google Docs. For our lost-deal review, the setup is simple: the delegate lone needs to recover applicable CRM data and difference it alongside approved promotion content.

In Agent Builder, use Knowledge for reusable citation matter and Inputs for data that changes alongside all run, specified as the client or agreement to review. For example, a assistance agent power citation your assistance policies during reviewing a selected customer’s issue.

Agent Builder lets users create a cognition vault supporting images, website pages, blog articles, and uploaded documents, including spreadsheets, PowerPoints, and PDFs.

With HubSpot, teams can additionally set AI environment for agents. With Context filled out, AI volition cognize practically everything concerning you and your business.

Teams must additionally decide who can use an delegate and who has the capability to edit it. Agent Builder allows users to designate who on the squad has permission.

Test alongside delegate scenarios.

Try the delegate on records you comprehend fine adequate to fairness its answer. Check that the delegate reviewed the intended agreement or metrics, used the approved references, and supported its recommendation. Be certain to vet the cited matter yourself.

Pro tip: Test runs in Agent Builder don’t disbursal additional credits. Repeat evaluation until the output is accordant and meets your requirements.

Compare the period needed to inspect and accurate the agent’s outcome alongside the period needed to do the project manually. A helpful delegate should create dependable activity and preserve the squad several effort.

Iterate on range and instructions.

A failing test isn’t necessarily a failure! It’s an chance to acknowledge what needs fixing. This is an chance to revise instructions and retool prompting. Change one item at a time, afterward rerun the unsuccessful example and a earlier prosperous one. This helps teams see whether the alter fixed the issue without creating another.

Test, Evaluate, and Improve

After an delegate is built, ongoing evaluation helps teams decide whether it plant correctly and delivers value. Teams should assessment agents following their instructions are edited and whenever tools, permissions, models, or connected data sources change. Schedule monthly reviews and run a test following a reported error.

Build a applicable test set.

A test set is a gathering of examples of the activity an delegate typically handles, specified as analyzing closed-lost deals. In this case, all closed-lost agreement counts as one example in the test set.

Keep the test set in a spreadsheet so you can difference what you expected alongside what happened. Analyze all example before operating an agent. Fill in your expectations in the spreadsheet. Next, run all agreement through the delegate and mark its output.

Pro tip: Run selected cases multiple times. Model outputs can change between attempts.

Review failures and border cases.

Keep a nonaccomplishment log in the test sheet. For all problem, add a pillar for “reviewer notes” and explain what went incorrect and what needs checking. Be extremely particular concerning mistakes. Next, collection akin failures, examine the cause, and use the suitable fix. Rerun the affected cases following changes are made. Record the alter and the new outcome alongside the first so you can difference them.

Common agentic failures and fixes are as follows:

• Outdated answers. Check which origin an delegate used. Replace obsolete documents or eliminate them from its approved references.
• Missing information. Check whether the required origin is connected and accessible. If the data doesn’t exist, instruct an delegate to petition it or emblem the gap.
• Wrong document or recipient. Check the input passed to an delegate and how it identifies its target. Use a distinctive document ID.
• Unsupported conclusions. Require supporting evidence and explain whenever an delegate should depart a inquiry unanswered.
• Unwanted actions. Remove unnecessary tools or restrict their permissions. Require individual endorsement for consequential changes.

Track achievement metrics.

Use the test spreadsheet to measure whether an delegate completes helpful activity and reduces the attempt your squad spends on it. Add columns for Review time, Corrections needed, Run cost, and Agent version.

After all test, document how lengthy you spent checking and correcting the result. Enter the run disbursal if available. Then summarize these measures in a distinct tab, calculating project achievement rate, correction rate, individual period spent, and disbursal per prosperous task.

Compare individual period alongside the period needed to complete akin tasks manually. For example, if preparing a client handoff takes 20 minutes manually and eight minutes alongside an agent, that’s 12 minutes saved — provided the outcome meets the identical norm requirements.

Add guardrails and individual handoff.

AI guardrails define what the delegate may handle, what data it shouldn’t touch, and whenever a individual should obtain over. Set these rules before initiate to decrease the hazard of unauthorized actions, data exposure, and malicious instructions hidden in prompts or origin material.

Here’s how teams should method environment guardrails:

1. Define the job’s boundaries in the agent’s instructions, including how to grip unrelated requests and instructions established in origin material.
2. Limit the agent’s tools and permissions to what the job requires.
3. Enforce admission restrictions through the connected apps’ controls.
4. Share lone necessary personal data and restrict who can perspective the output.
5. Specify whenever the delegate must halt and who should obtain over.
6. Require an authorized person’s endorsement before delicate or high-impact actions, enforced through the workflow or tool.
7. Test the complete handoff.

Agent Builder allows teams to set guardrails throughout four dimensions: instructions, actions, cognition vault, and permissions. Build harmless agents for promotion and revenue purposes through Agent Hub.

Choose the archetypal deployment surface.

A deployment exterior is anywhere group use an delegate or obtain its work. An inner inbox, a CRM workflow, and a customer-facing conversation autumn under distinct deployment surfaces. Choose one exterior that your delegate supports.

Next, define how to run the delegate according to the following categories.

• On demand: A individual decides whenever to commencement the project and supplies the inputs.
• Event-triggered workflow: The project should commencement whenever item happens, specified as a form submission or document update.
• Scheduled workflow: The project should run at a set period or interval.

For any workflow, define the starting condition, the inputs passed to the agent, and anywhere its results go. With Agent Hub, teams can move and nexus workflow actions and agents without breaking the general logic.

Complete the preflight check.

Before publishing an agent, appearance through its setup alongside intended users. Check for:

• Instructions, tools, and citation matter are correct.
• Pilot users can complete the job, and reviewers can open its output and supporting sources.
• The starting conditions are met (records, triggers, events).
• Credit use per run and a monthly run bounds are set.
• An authorized individual owns the review.
• There is a fallback (decide who completes the project manually if an delegate fails).

Once the checks are complete, preserve and publish the delegate version.

Monitor the archetypal live run.

Run a small aviator first, and assessment all first result. Follow all project from commencement to end. In Agent Hub, teams can display for the delegate and inspect its runs. Users can inspect the status, output, sources, and estimated credits.

A Success position method the run completed without an implementation error. A individual motionless needs to judge whether the agent’s answer is helpful and accurate.

The workflow’s Run history helps teams examine whether the workflow started and anywhere an implementation failed. If a grave issue creeps in, intermission the workflow, acknowledge affected runs, and use the manual fallback during you fix and retest the workflow.

AI Agent Use Cases for Marketing, Sales, and Service

Curious what AI agents teams have built for distinct purposes? See how genuine marketers have used agents to simplify their workflows.

Marketing Use Case: Spot Influenced Deals Through Cross-References to LinkedIn Engagement

Not everyone who sees a LinkedIn ad clicks. But, that visibility may motionless power a deal. Filippo Irdi, a elder petition promotion director at Unmuted, built an delegate that cross-references LinkedIn ad belief data alongside HubSpot closed-won deals to acknowledge that influence. I asked Filippo to stroll me through a detailed delegate setup.

Here’s how you can recreate his work.

1. Pull LinkedIn belief data.

Using the LinkedIn API, drag belief data week by week for a particular timeframe. The data shows which companies were served your ads and how many impressions they received. LinkedIn doesn’t provision granular company-level belief data. Weekly data is adequate to acknowledge helpful patterns.

2. Match it alongside HubSpot deals.

“For the identical timeframe — for example, the final six months — drag agreement data from HubSpot. For all deal, grasp the company, whenever the agreement was created, whenever it changed stages, and whenever it closed. Then equivalent the two datasets,” Irdi suggests.

From there, you can now see whether a business was exposed to LinkedIn ads before entering the revenue pipeline or decision a deal. Use the following configurations to drag the deals.

3. Build a timeline.

Results volition be returned in two tables. I propose reformatting results into a timeline that’s easier to understand. Marketers can afterward perspective the timeline by week or month, making it easier to see the association between ad visibility and revenue activity. This is what Irdi’s timeline looks like.

4. Automate the analysis.

Using LinkedIn and HubSpot MCPs, an delegate can grip this procedure by pulling the two datasets, matching companies, and generating the timeline automatically.

“I run the delegate all duration to have a repeatable way to acknowledge overlap between LinkedIn ad visibility and closed deals, equal whenever those customers never clicked an ad,” shares Filippo.

A Few Agent Guardrails to Get You Started

Before Pulling Data

• Do not query item until the ad account, exact date range, period granularity, and output format are defined.
• Do not construe vague periods specified as “Q3” or “last quarter.” Convert them to definitive dates and verify them.
• Show the expected API disbursal of monthly vs. weekly granularity before operating the analysis.

Time Periods

• Define the complete period grid before pulling data.
• Mark partial weeks or months explicitly.
• Keep LinkedIn visibility bucketed by week/month, but keep agreement near dates on their exact dates.

LinkedIn Data

• Never use lone currently energetic campaigns for historic analysis.
• Never reuse one period’s run catalog for another period.
• If company-level data comes rear bare but run achievement shows impressions, treat it as a apt campaign-selection problem—not “no data.”
• Preserve LinkedIn institution IDs for deduplication.

HubSpot Data

• Retrieve the HubSpot portal timezone before handling near dates.
• Identify closed-won deals using hs_is_closed_won, not hardcoded phase IDs.
• Use home-currency values whenever calculating totals throughout currencies.

Sales Use Case: An AI Sales Compensation Agent

Commission disputes rarely stem from a sole bad month. They arrive from reps not being capable to see the math. AI can help.

Driven, a revenue commission platform, built an delegate for HubSpot that takes a inquiry or a preparedness document. The delegate reviews Driven’s compensation data and returns an clarification restricted in particular deals and targets.

Andres De Jonge, co-founder at Driven, explains how the delegate works.

• How reps use the agent: A rep asks, “Why is my payout lesser this month?” The delegate pulls their payout, the credited deals, and quota attainment. It afterward explains the gap, for example: “Two deals slipped to the next period, and attainment dropped below the 100% accelerator threshold.”

• How admins use the agent: An admin pastes a rule, uploads a scheme PDF or Excel file, or selects CRM fields. The delegate translates it into a organized outline plan, including formulas, tiers, quota references, and filters. It additionally runs a preview against genuine historic deals.

Inputs for this delegate are natural-language questions, scheme documents (PDF, Word, Excel), and CRM site selections from HubSpot or Salesforce, synced into Driven’s database. Outputs are payout explanations alongside connected deals, variance breakdowns between periods, and outline scheme configurations alongside calculation previews.

If you try replicating this agent, you may encounter a bump since the delegate example never does the math. Payouts are calculated by Driven’s equation motor alongside decimal.js for financial-grade rounding. The agent’s job is to fetch the calculation trace and explain it, or to propose a configuration that the motor afterward executes.

Service Use Case: Find Recurring Support Issues With HubSpot and ChatGPT

Your assistance squad may answer the identical inquiry dozens of times without realizing how frequently it comes up. Connecting HubSpot to ChatGPT lets you examine recurring ticket topics and decide which matter deserves improved records or a nearer appearance from the merchandise team.

Here’s how it works.

1. Connect HubSpot to ChatGPT.

Ask your HubSpot Super Admin to endorse the connector and aid admission to the ticket data you’ll review. Then, in ChatGPT, open Settings > Apps > HubSpot > Connect. Follow the prompts to choose your HubSpot document and authorize the connection.

2. Choose the tickets to investigate.

Start alongside one assistance pipeline and an exact date range. Open Deep research in ChatGPT and allow HubSpot under Apps. Then modify this prompt:

Review HubSpot tickets created in [pipeline name] between [start date] and [end date]. Use the ticket names and descriptions you can access. Group tickets by the underlying client problem, figure all group, and nexus to supporting tickets. Keep unclear cases separate. State how many tickets you analyzed and any retrieval limits. Don’t alter records.

3. Turn the findings into recommendations.

Once the study is ready, ask:

For the three largest groups, propose what the assistance squad should examine next. Distinguish questions that power advantage from clearer instructions from reported merchandise problems. Explain your reasoning using connected tickets. Flag anywhere the evidence is insufficient.

Treat these as suggestions for review. Remember, seeing a indication doesn’t established its cause.

4. Check the evidence and allocate the next step.

Open the connected tickets and inspect that they pertain in their groups. Compare the reported total alongside the identical filters in HubSpot. Then, examine any mismatch before presenting the findings as a complete picture.

From there, provision an approved advice to the suitable owner. The deliverable is a short, originated matter report. Judge its usefulness by whether the groupings are accurate, if the recommendations last review, and if the procedure takes small period than your manual review.

When to Consider Models, Protocols, or Multiple Agents

Your archetypal delegate is up and running. As you provision it additional work, you may discover breaking points. Before adding complexity, acknowledge whenever the scheme breaks. That volition assistance you decide whether to alter the model, nexus another tool, or distinct the activity among agents.

Choosing a Model

The example is the particular AI program that interprets instructions, evaluates information, and chooses the agent’s next step. Its capabilities power the quality, speed, and disbursal of the work. Examples contain models from OpenAI’s GPT, Anthropic’s Claude, and Google’s Gemini.

The tools you use may have a default AI model. If your phase lets you choose models, difference outputs using the identical test cases. Pick the smallest costly choice that meets norm and response-time requirements.

What Agent Protocols Mean

A protocol is how application exchanges requests and results. For agents, there are two protocols:

• Model Context Protocol (MCP). This connects AI applications to tools and data sources. For example, an delegate could use an MCP association to hunt an approved document library.
• Agent2Agent (A2A). This allows autonomous agents to toggle tasks and results, including agents built on distinct platforms.

In Agent Hub, teams can nexus external tools to HubSpot through MCP. This lets agents admission live endeavor data and obtain act throughout supported applications without tradition integrations or code.

When Multiple Agents Make Sense

A multi-agent setup divides activity among agents alongside distinct responsibilities. One power collect research, during another uses that investigation to prepared a brief. Remember: additional agents average additional outputs and handoffs to check. Add an delegate lone whenever the betterment in norm or speed justifies the additional disbursal and maintenance.

Consider splitting the activity among agents when:

• Different tasks need distinct tools, instructions, or admission permissions.
• Independent investigation tasks could run at the identical time.
• Testing shows that one delegate often mixes up responsibilities notwithstanding a clear, narrow setup.

Pro tip: Always difference the blended outcome alongside your single-agent setup using the identical tests.

Frequently Asked Questions About How to Build an AI Agent

Do I need a developer to build my archetypal agent?

No, developers are not necessary whenever construction AI agents. Bring in a developer whenever the project requires a tradition integration or controls that your builder doesn’t support. You can build dozens of marketing, social media, and revenue agents in a no-code builder akin HubSpot Agent Builder.

How do I enhance an AI delegate without training a model?

Use unsuccessful tests to acknowledge what needs changing. That may contain unclear instructions, missing citation material, incorrect inputs, or unsuitable tools. Fix the applicable part and rerun your test set.

How should I grip delicate actions akin refunds or data deletion?

Set AI guardrails for high-impact actions. Have the delegate prepared the projected act for an authorized individual to review. Show exactly what would happen, including the customer, amount, or records affected. Enforce endorsement through the workflow or connected tool, and keep implementation blocked until endorsement arrives.

When should I move from one delegate to multiple agents?

Build multi-agent systems whenever evaluation shows a apparent advantage from separating responsibilities. For example, distinct tasks power need distinct admission permissions. If you add another agent, define what all one receives and returns, afterward difference quality, speed, and total disbursal alongside your first setup.

What data can a no-code AI delegate use?

Access depends on all builder, its supported connections, and the permissions you grant. Common sources contain CRM records, uploaded documents, assistance articles, spreadsheets, and data supplied alongside all request. HubSpot’s knowledge vaults assistance files, images, selected HubSpot content, and CRM segments.

Start alongside one delegate your squad can depend on.

When construction your archetypal agent, commencement alongside one project your squad knows well. Give the delegate the data and tools it needs, afterward test its activity against apparent expectations. And, before you measure an agent, assessment results carefully.

Already a HubSpot user? Agent Builder lets you create agents in plain language, no coding required. So get started, and see what you can build.