How to proceed with AI agent introduction | Determination criteria for introduction support, costs, and PoC

Those who are considering support from organizing the target business to prototyping and system coordination can check the implementation flow at ZETTAi AI agent construction support.

ZETTAi AI agent construction supporthttps://zettai-ai-agent-site-render.onrender.com/

When a company introduces an AI agent, it proceeds in the following order: Select one task → Prototype (PoC) to check effectiveness and failure cases → Operate within a limited range. Before requesting implementation support, decide which tasks will be delegated to what extent and where people will approve them. Once you know the necessary collaboration partners, information arrangement, and evaluation methods, you can specifically consult with the support company about the scope and cost you are looking for.

When choosing a support company, focus on how to stop and fix the process in case of failure, usage fees, and the scope of responsibility after handover, rather than the completed prototype screen. Please select one target business, provide a sample including normal cases and exceptions, and prepare the assumptions for the estimate.

We will check in order how to select a business, the details of the request to the implementation support company, the conditions for passing the PoC, the cost and the operation system. If you would like to consult about how it applies to your own business, you can check the scope of support at ZETTAi's AI agent development and implementation support.

Three stages of AI agent introduction. Select one target business, verify accuracy and exceptions with a prototype, and perform limited operation with human approval.

*Introduction flow. At each stage, decide what to check before proceeding to the next step, and if the results of the prototype do not meet the passing conditions, review the scope. *

Does the work require an AI agent in the first place?

Simply saying "I want to introduce AI" does not allow you to choose the system you need. If your work involves referring to multiple materials or systems and determining the next process according to a step-by-step process, you may want to consider an AI agent. On the other hand, if you only need to draft a fixed phrase, an existing chat AI or template may be sufficient.

What you want to doHow to try it firstConditions for considering implementation support
Document summaries and draftsSmall trials using existing generation AI toolsReference authority and source management for in-house materials is required
Collate multiple documents and make a proposal to the person in chargeDesign a business flow with approvalRequire multiple connection destinations such as customer information and schedule
Updating information and sending it externallyStart with a prototype that requires human confirmationIt is necessary to design update permissions, execution logs, and stop procedures

By first determining whether existing tools are sufficient or whether separate development/coordination is required, it will be easier to avoid estimates that include unnecessary functions. Microsoft's Agent AI Implementation Guide (https://learn.microsoft.com/ja-jp/agents/adopt-overview), which serves as a resource for checking internal readiness, organizes aspects such as business value, governance, technology/data, and organization. This is not for the purpose of adopting a specific product, but can be used as a reference for identifying conditions that are lacking within your company.

Which job should you choose first?

Divide the current work into reception, information gathering, judgment, drafting, approval, and execution. For each, write down the materials used, the person in charge, the number of cases per month, and any problems you are having.

For example, when preparing for sales, the first tasks are checking the schedule,'' gathering customer information,'' and ``compiling negotiation notes,'' and the person in charge is responsible for contacting the customer. Limiting the scope makes it easier to identify causes and measure effects.

If there are multiple candidates, select the job that has a large number of candidates, can explain the input and expected output, and is easy for people to check the results. For operations that have a large impact on misprocessing, such as transferring money or automatically sending messages to customers, decide on a policy that requires approval from the first attempt rather than automatically executing them.

In the example of sales preparation, the input is visit schedule, customer materials, and past negotiation notes'', the output is preparation memo with sources'', and the person in charge's job is ``confirming facts and deciding on the next proposal''. If you don't include the part where the AI ​​contacts the customer all at once, measure both the time it takes to make preparation notes and the time it takes to confirm, you can judge whether the burden has really been reduced. This is a hypothetical example for illustrative purposes only and is not an actual value.

In which four stages should we proceed with implementation support?

StagesThings to doConditions for next step
OrganizeDecide on target operations, data, and responsible personInput/output and confirmer are clear
PrototypeTry the process with a sampleYou can check normal cases and exceptions
Trial operationPractical use by a limited number of personnelAcceptable quality and confirmation time
EstablishmentPrepare procedure manuals, monitoring, and improvement personnelContact and stop operations in the event of a problem

Separating the deliverables and costs for each stage will make it easier to decide if you need to change targets midway through.

Determine the passing conditions for the prototype first

Just because it "works" does not mean you can judge whether it can be used in practice. Before PoC (proof of concept), prepare multiple requests of the same type and record correctness, confirmation time, and behavior in case of exceptions. Compare the manual work before implementation and the total time after humans check the AI ​​output under the same conditions.

Confirmation itemsConfirmation examples during sales preparationExamples of failure
Correctness of informationCustomer name and negotiation date match the original documentInformation from another customer is mixed
Tracking the sourceOpen the original material from the summaryUnable to confirm the basis
Human approvalReps see content before contacting customerSend without confirmation
Exception handlingWaiting for confirmation when there is no materialGuessing and continuing processing
EffectCompare the combined time for creation and confirmationThe burden of confirmation exceeds the original work

Standards for acceptable errors and confirmation time vary depending on the job. The person in charge will decide before the trial, and if it doesn't match, the target range and procedures will be reviewed.

The evaluation looks not only at the final answer, but also what materials and tools were used to arrive at that result. Google Cloud's Official material for AI agent evaluation also shows the idea of ​​evaluating the final answer and the path of tool usage separately. Regardless of the product you use, it will serve as a reference when creating a practical test sheet. For details on the decision to move from PoC to production, please see Causes and evaluation criteria for AI PoC failure.

Official material for AI agent evaluationhttps://docs.cloud.google.com/gemini-enterprise-agent-platform/models/evaluation-agents

Include “troublesome requests” in the test

If you only try general requests, you will miss exceptions in practice. Let's prepare a case where the materials cannot be found, there are multiple clients with the same name, the person in charge has left the company, or the requests are contradictory.

We evaluate not only whether the answer is correct, but also whether the process can wait for confirmation and not continue processing on its own. It's easier to track improvements when you can see the execution logs and the original documentation together.

When connecting to external services, separate viewing, updating, and sending permissions to avoid giving unnecessary permissions to the prototype. Decide how to stop erroneous processing and who will be responsible for checking records before the trial run.

When checking internal regulations and the division of roles with subcontractors, you can refer to IPA (Information-technology Promotion Agency)'s Guidebook for Safely Utilizing Generated AI and AI Agents (PDF). Be sure to record which items have been confirmed regarding your company's input information, users, contractors, and approvers. The approval design for external transmission and data updates is organized in detail in AI agent approval flow and authority management.

Guidebook for Safely Utilizing Generated AI and AI Agents (PDF)https://www.ipa.go.jp/jinzai/ics/core_human_resource/final_project/2026/rcu1hd0000018kkn-att/rcu1hd0000019kwj.pdf

What should I compare quotes from support companies?

Check the number of connections included in the construction cost, preparation of materials, explanations to employees, and number of revisions after implementation. Also decide who will respond to changes in external service specifications or changes in personnel.

Continuation of operations is affected by who manages accounts, source code, and configuration information. If you want to avoid a situation where you cannot change anything without asking the company in charge, please discuss the scope of transfer materials before signing the contract.

In the estimate, we will separately check not only the initial construction cost, but also the usage fee for the AI ​​service, maintenance, monitoring, and change response for external collaboration. AI agent introduction cost | Initial/monthly breakdown and how to compare estimates organizes the items that affect costs.

Items to compareQuestions to check with estimates
Scope of prototypingTo what extent does it include target operations, input materials, collaboration partners, and exception tests
Initial costCan you separate design, implementation, verification, and explanation to employees
Monthly costWhat are the upper limits for AI usage fees, monitoring, maintenance, and additional modifications
HandoverWho manages accounts, settings, logs, and termination procedures
Additional workWhat is the unit price and approval method when the number of target operations and collaboration partners increases?

Even if the amount is the same, the total burden will change if document preparation and operational transfer are not included. Compare quotes that match your conditions.

What should I tell them before requesting implementation support?

At your first consultation, you can consult multiple companies with the same conditions by summarizing the following six items on one sheet. Do not fill in the undecided fields with guesses, but write "Undecided."

ItemExample of entry when targeting sales preparation
Target businessGather customer materials and past business negotiation notes before visiting
Current problemsInformation is in multiple locations and preparation takes time
Input and link destinationCalendar, customer management, internally shared materials. Connection availability is unconfirmed
Expected OutputPreparation notes with sources. Not sent to customers
Passing conditionsNo mix-ups and comparison of required time including confirmation
Restrictions and persons in chargeSpecify handling of customer information, viewing privileges, approvers, and suspension decision-makers

It is important to separate and communicate work to be entrusted to a support company'' and work to be decided internally.'' Internal personnel are responsible for determining operational correctness and usable information. With ZETTAi implementation support service, you can check the scope of support from consultation to prototyping and operation.

Test table to determine pass/fail of prototype

If you're an agent helping with sales preparation, you might want to enter something like the following to see if it works: This is an example of evaluation design. Please decide the number of cases and allowable conditions according to your company's business.

Situation to testExpected behavior
Identify customers and related materialsCreate preliminary notes with sources
There are multiple customers with the same nameAsk about the target and do not mix information from other companies
Necessary information is missing from the reference materialMark as unknown and do not supplement the facts
User does not have permission to viewDo not display unauthorized materials
Connection destination does not respondNotifies failure and does not treat it as completed

Failures found during trial operations will be retained as confirmation data for repeated attempts even after corrections have been made. This is because we do not judge based on specific success stories.

What you will receive before moving to full-scale operation

In addition to usage instructions, you will receive a list of linkages and privileges, how to check usage fees, termination procedures, and contact information in the event of a failure. If an in-house person in charge actually attempts to stop and restart the system while receiving explanations, they can confirm any deficiencies that cannot be detected from documentation alone.

When consulting, please prepare a sample document that replaces the above request memo, business flow, and personal information, and draw the line between to what extent should it be automated'' and where should a human approve it?''

FAQ

Can I install it even if I don't have in-house engineers?

You can get support, but you need someone in-house to judge the correctness of your work. Let's separate the technical and operational responsibilities.

Is it better to implement it for all departments at once?

Although there are advantages to designing a common platform, it will be easier to verify if the initial demonstration is focused on a single task. Please also check the differences in materials and permissions for each department.

What should I give to the implementation support company in order to proceed with the consultation?

Please organize the procedures for the target work, the types of materials used, the number of cases per month, the amount of time it currently takes, and operations that require approval. Do not share actual data, including customer information, until you have confirmed the handling conditions, and use a sample with the content replaced first.

Related articles

Official documents referenced in this article

The following are actual public documents. The tables and sales preparation examples in the article are not reprinted from these materials, but are examples created to help readers organize their company's conditions.

Consult according to your company's situation

If you have decided on the target business, please see the details of implementation support at ZETTAi AI agent construction support.

ZETTAi AI agent construction supporthttps://zettai-ai-agent-site-render.onrender.com/

If you want to organize it yourself, please check What is an AI agent and Breakdown of implementation costs. If you have decided on the target work, you can check the scope of support at ZETTAi AI agent development and implementation support.

For individual consultations, please write the details of the above request memo in the "Inquiry details" section of the Corporate inquiry form. When first contacting them, instead of sending the client's name or confidential information, it will be easier to organize the next points to confirm if you tell them the outline of the business, the problem you are having, and the scope of support you would like.