Custom AI Agent Solutions
Build an AI agent that works the way your business actually runs.
Not an off-the-shelf chatbot. The agent is trained on your data, connects to the systems you already run and does the work — with guardrails and a handover to a real person.
- 10–30 daysA pilot with clear KPIs
- Multi-channelWeb · Zalo · WhatsApp
- Human-in-the-loopYou keep control throughout
- Overview
- Data
- Tasks
- Integrations
- Guardrails
YOUR BUSINESS AI AGENTSales & Service Agent
Test runAgent setup 80% complete
Primary tasksWhich work will the agent do?
3 tasks- Advise and identify the need
- Score the lead against criteria
- Book a meeting with a specialist
Knowledge sourcesThe data used to train the agent
12 sourcesConnected systemsLets the agent take action
4 connectionsGuardrails & HandoffControl and handover to a person
ReadyNever changes prices on its ownAlways requires approval
Test the agentRunning
I need a solution for a 20-person sales team.
I can help you work out which process fits. Which CRM does your team use?
Type a message…
ILLUSTRATIVE INTERFACE
From conversation to action
Not just answering.
The agent actually does the work.
A Custom AI Agent understands the intent, calls the right system and completes the task inside your process — with people in control at the steps that matter.
Sample scenarios
Flow runningCapture and qualify leads
Agent runningNACustomer · Website“We have 20 salespeople and use HubSpot. Can InCard integrate with it?”
- Step 1Understand the needIdentify team size, goals and the CRM in use. 0.8 s
- Step 2Look up the business dataCheck against the approved HubSpot integration documentation. 1.2 s
-
Step 3Update and score the lead in the CRMCreate the contact, record the need and apply the scoring rule.
HNguyen An · ABC CompanyHubSpot · Lead score 82/100Qualified
In progress - Step 4Propose a meeting with a specialistOnly after the customer confirms. Waiting
Guardrails are protecting the processThe agent never commits to a price or touches sensitive data without approval.Human approval
- What one conversation produces
- The lead is createdWith complete data in the CRM
- The need is classifiedScored against your own criteria
- Sales get the full contextReady to carry on
Flow runningBook a consultation
Agent runningTHCustomer · Zalo OA“I would like to see an adviser on Thursday afternoon — do you have a slot?”
- Step 1Establish the need and the time slotRecord the topic and the time range the customer proposed. 0.6 s
- Step 2Check the advisory team’s availabilityCheck against working hours and the configured slots. 1.0 s
-
Step 3Hold the slot and send a confirmationHolds the slot while waiting for the customer to confirm.
LThursday · 14:00 – 14:30Calendar · Held for 15 minutesAwaiting confirmation
In progress - Step 4Write the appointment into the CRMRecorded only after the customer confirms the slot. Waiting
Guardrails are protecting the processThe agent never books outside the configured hours and never accepts a booking on an adviser’s behalf.Human approval
- What one conversation produces
- The slot is heldNo clash with the advisory team’s calendar
- The customer gets a reminderOn the channel they are using
- The adviser has the contextKnows what the customer needs advice on
Flow runningCreate a quote
Agent runningMPCustomer · Website“Send me a quote for 20 users over 12 months.”
- Step 1Establish the plan and quantityRecord the number of users, the term and the items included. 0.7 s
- Step 2Fetch the current price listUses only the price list and terms approved for this channel. 1.1 s
-
Step 3Draft the quoteFills in items, term and conditions using your company template.
₫Draft quote · 20 users / 12 monthsAwaiting the owner’s approvalDraft
In progress - Step 4Send the quote to the customerSent only after the owner approves the content and the price. Waiting
Guardrails are protecting the processThe agent never sets a price, never applies a discount and never sends a quote before it is approved.Human approval
- What one conversation produces
- The quote matches the price listNo drift from the approved terms
- Discounts are a person’s decisionThe agent only prepares the draft
- Sales only review and sendCuts out the drafting time
Flow runningOpen a support ticket
Agent runningQDCustomer · Messenger“I have not been able to log in to my account since this morning.”
- Step 1Identify the issue and its priorityClassify it against the support team’s own criteria. 0.5 s
- Step 2Check the account historyReads earlier requests so the customer is not asked everything again. 0.9 s
-
Step 3Create the ticket and assign the right teamWrites the full conversation context into the ticket before handing it on.
TTicket · Cannot log inHigh priority · Engineering teamAssigned
In progress - Step 4Hand over to the agent on dutyA person takes it on and decides how to resolve it. Waiting
Guardrails are protecting the processThe agent never closes a ticket, never issues a refund and never promises a fix deadline.Human approval
- What one conversation produces
- The ticket has full contextThe customer does not repeat themselves
- Priority set correctlyAgainst the team’s criteria
- A person picks it up immediatelyNothing gets lost
These scenarios illustrate the process. The systems connected and the actions the agent is allowed to take are defined per project.
Enterprise-ready by design
Powerful enough to act.
Safe enough to scale.
Every agent is built on one consistent control layer — from the data it may use, to the actions that need approval, to how it hands over to a person.
Answers only from sources the business trusts
Connects your website, documents, CRM and internal knowledge base. Access is controlled per agent.
Data sources are controlled per agentClear boundaries on every action
Define restricted content, sensitive data fields and the tasks that require approval.
The agent knows when to stop
Hands over to the right specialist with the full history, intent and data collected.
The right person, the right permissions
Role-based access, a change history, and every agent action tracked.
One agent, consistent across every channel
The same knowledge, rules and context wherever the customer starts.
Measured by operational KPIs
Track response time, task completion rate, handovers and conversation quality.
- Built-in guardrails
- Human-in-the-loop
- Role permissions and an action history
- Operational KPI tracking
A pilot with a clear roadmap
From a real problem to
an AI agent running a pilot in 10–30 days.
We do not start with a large project. We pick one process that matters, agree the KPIs and run a pilot small enough to learn fast — and real enough to measure.
- 10–30 daysExpected pilot duration
- 01 processSuggested starting scope
Choosing the right problem to start with, together.
The workshop focuses on your current process, its bottlenecks and the business result to improve. Both sides agree the pilot scope before anything is built.
- ICThe InCard teamSolution · AI · Integration
- DNYour businessProcess owner · IT · Business
Solution workshop90 minutes · Online or on siteSelecting the pilot scope
InputYour current process
- The start and end points
- The systems in use
- The steps needing human approval
Defined togetherUse case & KPI
- The tasks the agent must complete
- The data it may access
- The metrics that judge the pilot
Delivered after the workshopPilot rollout blueprint
AI Agent Pilot BlueprintScope · Workflow · KPI · TimelinePDF
Checkpoint #1The business confirms scope, data and KPIs before anything is built.Business approval
Designing the flow and the limits before building.
The blueprint becomes a conversation flow, a list of tasks the agent may perform, and guardrails for each sensitive step.
- ICThe InCard teamSolution · AI · Prompt
- DNYour businessProcess owner · Compliance
Agent designConversation flow · Tasks · GuardrailsFinalising guardrails
InputBlueprint approved
- The agreed pilot scope
- The KPIs and how they are measured
- The systems involved
Defined togetherFlow and limits
- The main conversation scripts
- The tasks the agent may perform
- The steps requiring approval
Delivered after designAgent and guardrail specification
Agent Design SpecFlow · Tasks · PermissionsPDF
Checkpoint #2The business approves the flow, the guardrails and the permissions before systems are connected.Business approval
Connecting real data, training on permitted sources.
Connect the systems within the approved scope, load internal documents and tune the agent in a sandbox separated from production data.
- ICThe InCard teamAI · Integration
- DNYour businessIT · Data owner
Test environmentA sandbox separated from production dataTuning the answers
InputApproved data sources
- Internal documents and FAQs
- The accounts connecting the systems
- Role-based permissions
Tested togetherAnswer quality
- The test question set
- The cases the agent must refuse
- The handover threshold
Delivered after trainingSandbox agent build
Agent Sandbox BuildConnected · Trained · LoggedPDF
Checkpoint #3The business signs off answer quality on a test question set before going live.Business approval
Running for real with a limited group of users.
The agent handles real work inside the agreed scope, with someone on duty watching and stepping in when needed. Every adjustment is logged.
- ICThe InCard teamAI · Support
- DNYour businessThe team running it day to day
Pilot runningLimited user group · Someone on dutyMonitored daily
InputLive scope
- The agreed channels and hours
- The user group taking part
- The handover rules
Monitored dailyItems needing attention
- Conversations needing a person
- Questions the agent cannot yet answer
- Guardrail adjustments as they arise
Delivered during the pilotPilot operations log
Pilot Operations LogConversations · Tasks · Handover pointsPDF
Checkpoint #4Both sides review the logs weekly and agree adjustments before continuing.Weekly review
Checking the KPIs before deciding to expand.
Compare the pilot results with the KPIs agreed in stage 01. Expanding is a decision backed by data, not a default next step.
- ICThe InCard teamSolution · AI
- DNYour businessProcess owner · Leadership
End-of-pilot reportCompared with the KPIs agreed in stage 01Awaiting the expansion decision
InputPilot data
- The conversation and task log
- Feedback from internal users
- The handover points
Decided togetherNext step
- Keep, adjust or stop
- The next process if you expand
- How it runs long term
Delivered at the end of the pilotPilot report and expansion proposal
Pilot Review & Rollout PlanKPIs · Lessons · RoadmapPDF
Checkpoint #5The business decides to expand based on the KPIs; continuing is not the default.Go / No-go
The stages can overlap. 10–30 days is the expected pilot duration for one process, not a commitment that applies to every project.
- Rollout principles
- 01Start smallOne process with clear value
- 02Measure from day oneKPIs agreed before the pilot
- 03Expand under controlScale only once the results are confirmed
Frequently asked questions
Before you start building an agent.The things worth clearing up.
What the agent is allowed to do, what data to prepare, how long it takes and how far compliance goes — the four things businesses usually settle first.
Next step
Prototype an agent on your own process.
The workshop looks at the process costing you the most people, works out which data and systems to connect, then sketches the scope of a first pilot.
-
An off-the-shelf chatbot answers from a fixed script. This agent is trained on your own data, connects to the systems you already run and performs tasks — lookups, creating records, booking — with guardrails and a handover to a real person where it matters.
The difference is that it does the work, not just the answering. -
As a reference, 10–30 days for a pilot in three phases: days 1–7 connecting data and systems; days 7–14 training the agent, setting guardrails and testing with your team; days 14–30 running the pilot with real users and tuning against the agreed KPIs.
-
The usual candidates are your CRM, knowledge base, helpdesk, calendars and internal operational systems. How deep each connection goes depends on the APIs, permissions and existing data architecture — final scope is agreed after a technical review, not promised up front.
-
Role-based access, an action log, encryption in transit and at rest, plus NDA and SLA. On the legal side, support for GDPR, PDPA and Vietnam’s Decree 13, along with per-channel consent management. The specific arrangement is written into the rollout agreement.
RBAC · action log · encryption · GDPR, PDPA and Decree 13. -
No. Actions such as changing prices, touching personal data or anything outside the configured scope require approval or a handover to a person. The section “Powerful enough to act, safe enough to scale” above describes this control layer.
People keep the decisions at the steps that matter. -
Ecommerce and retail, education, real estate, hospitality and travel, beauty and wellness, plus services and SaaS. Each industry has its own scripts and handover criteria — see the industry pages for details.
-
Yes. Several businesses have deployed it and shared their results in the case studies. During the workshop, InCard also prototypes a scenario using your own data and process so you can see the agent running in your context.