By Curtis Socha, Principal Product Manager, and Seth Webster, Senior AI Engineer, Emburse
Most people do not return from a business trip excited to complete an expense report.
They return to a calendar full of follow-ups, a phone full of receipts, several card charges they vaguely remember, and a policy they hope they followed.
Then the reconstruction begins.
Find the receipt. Match the transaction. Split the hotel folio. Add the attendees. Explain the outlier. Submit the report. Wait for questions.
Each task is small. Together, they create a process that consumes employee time, slows reimbursements, and forces finance teams to correct problems long after the spending occurred.
Automation has improved parts of this process. Receipt capture can extract amounts. Rules can identify missing fields. Workflows can route reports to the right approver.
But completing individual tasks is not the same as coordinating the full outcome.
That is where agentic AI changes the conversation.
Agentic AI can gather information, understand context, interact with connected systems, recommend a next step, and take permitted actions on a person’s behalf. In travel and expense, that creates an opportunity to replace fragmented administrative tasks with an experience that actively helps employees and finance teams move work forward.
The goal is not to remove people from the process. It is to remove the unnecessary work surrounding the decisions only people should make.
Automation completes tasks. Agentic AI coordinates them.
The distinction between automation and agentic AI can sound technical. In practice, it is straightforward.
Automation generally follows a predefined path. Give a system a receipt, and it extracts the merchant, date, and amount. Submit an expense, and a rule checks whether the amount exceeds a threshold.
These functions are valuable. They also tend to be narrow and deterministic. The system has been instructed to complete a particular task in a particular way.
An agentic system works toward a broader outcome.
Instead of simply extracting information from a receipt, it might help an employee complete the expenses associated with an entire trip. That requires more than one automation. It may need to:
- Read receipts and hotel folios
- Match card charges with supporting documents
- Connect expenses to an itinerary
- Separate lodging, taxes, parking, meals, and incidentals
- Apply relevant company policies
- Identify missing context
- Suggest corrections
- Ask the employee for a decision when judgment is required
Automation provides many of the building blocks. Agentic AI coordinates them.
It also adapts to what it finds. A missing receipt, unusual charge, policy exception, or partial reimbursement may change what the system does next. Rather than forcing the employee through the same fixed sequence every time, the experience can respond to the actual situation.
That is a significant shift. The objective is no longer to make every form slightly faster. It is to reduce how much of the process needs to become a form in the first place.
The best agentic experience may not look like a chatbot
When people hear “agentic AI,” they often picture a conversational assistant waiting for a question.
Conversation can be part of the experience, but it should not define the experience.
A chatbot still requires the employee to recognize a problem, formulate a question, and initiate the interaction. A more useful agentic experience can work proactively.
Imagine returning from a customer event. Your itinerary, card transactions, hotel folio, and uploaded receipts already provide much of the necessary information.
Instead of presenting a blank expense report, the system prepares the trip for review. It connects the relevant transactions, extracts line items, suggests categories, and checks the expenses against company policy.
It then tells you what it completed and asks only for the information it could not determine.
Perhaps several expenses are missing a business purpose. Rather than making you open and edit each transaction, the system recognizes that they are connected to the same event and asks one clear question. Your answer can then complete the relevant expenses together.
That is the experience agentic AI should create: less searching, less clicking, and fewer interruptions.
The employee stays involved, but the surface area of interaction becomes smaller. The system handles the mechanical work. The person supplies intent, context, and judgment.
“Agentic AI is not about removing people from the process. It is about removing the unnecessary work around the process, so employees, approvers, and finance teams can focus on the decisions that actually matter.”
Curtis Socha, Principal Product Manager, Emburse
Control moves closer to the moment of spend
Traditional expense management often applies control after the activity is over.
An employee submits a report. An approver finds an issue. Finance returns it. The employee tries to remember what happened several weeks earlier.
Every handoff adds delay. Every delay increases the likelihood that context will be lost.
Agentic experiences can move guidance upstream.
If a receipt is incomplete, the employee can be notified while still near the merchant. If an expense appears inconsistent with policy, the system can ask for clarification before submission. If required information is missing, it can explain what is needed and why.
This is more than a better employee experience. It improves the quality of the information finance eventually receives.
Cleaner submissions mean fewer reports moving backward through the workflow. Approvers have better context. Auditors can spend less time investigating clerical errors. Employees receive fewer unexpected rejections and may be reimbursed faster.
The same principle applies to written policy.
No expense system perfectly represents every sentence, exception, regional variation, and special allowance contained in a company’s policy documents. Some rules fit neatly into configured thresholds. Others depend on context.
Consider an employee submitting only the reimbursable portion of a larger personal mobile phone bill. A basic check may flag the difference between the receipt total and the amount claimed. An agentic layer could also examine the written policy, recognize the company’s monthly allowance, and determine that the expense may be legitimate.
Instead of automatically sending another questionable item to an auditor, the system can add context, explain the discrepancy, and reduce unnecessary review.
Policy stops being something employees consult after a problem. It becomes guidance delivered in the flow of work.
Agentic does not mean unconstrained
The moment an AI system can take action, trust becomes essential.
But trust should not be based on confidence alone. It should be built into the architecture.
A trustworthy agentic system needs clear boundaries around what it may access and what it may do. Those boundaries should be enforced outside the AI model itself.
The model can recommend an action. It should not get to decide whether its own permissions apply.
At a practical level, this means an AI agent should operate within the permissions of the person using it. If an employee cannot access another department’s financial information, the agent acting for that employee should not be able to access it either. That restriction should be enforced at the system and API level, regardless of what the agent attempts.
The same principle applies to actions.
Some actions are low-risk and reversible, such as suggesting an expense category. Others have greater consequences, such as approving a large transaction, deleting information, or initiating a payment.
Those actions require stronger controls. Depending on the organization’s policies, that may include explicit confirmation, additional approval, or a mandatory human decision.
Agentic AI does not have to mean autonomous AI. It can operate within carefully defined levels of authority.
“Trustworthy agentic AI needs more than a capable model. Permissions, guardrails, and verification must be enforced outside the agent, so people remain in control even when the system makes a mistake.”
- Seth Webster, Senior AI Engineer, Emburse
Trust requires verification and explanation
Boundaries prevent an agent from moving outside its permitted area. They do not, by themselves, show that every decision inside that area is correct.
People also need a way to verify what happened.
A useful agentic experience should make its work observable. Employees, approvers, and finance teams should be able to see what the system did, what information it considered, and why it reached a recommendation.
If a receipt is flagged, the system should not simply announce that something looks wrong. It should identify the concern.
Was the receipt potentially altered? Did the amount differ from the card transaction? Was the merchant unusual for the expense category? Was the transaction duplicated? Did the expense occur at an unexpected time?
An explanation allows the reviewer to make a faster, more informed decision. It also gives the organization a way to evaluate the AI itself.
Finance teams should be able to monitor accuracy, examine edge cases, and determine where the system performs consistently. As confidence grows, they may decide to automate more low-risk activity. Where performance is less certain, they can keep stronger review requirements in place.
This creates graduated trust.
Organizations do not need to choose between reviewing everything and automating everything. They can decide how much authority to give the system based on the action, the risk, and the evidence they have observed.
Human in the loop does not mean human in every step
Keeping people in control does not require inserting a person into every routine action.
That would preserve the very burden agentic AI is intended to reduce.
The better model is human accountability supported by intelligent intervention.
AI can gather information, identify patterns, prepare work, and surface signals. People should remain directly involved when intent, accountability, unusual circumstances, or material risk require judgment.
An AI system may recognize that a charge is unusual. It may not know that an employee made an emergency purchase after a flight cancellation. It may identify a receipt inconsistency without understanding the business circumstances behind it.
Humans understand why spending happened.
The system’s responsibility is to make the issue visible, explain the available evidence, and guide the right person toward a decision. The employee, approver, auditor, or finance leader remains accountable for that decision.
Over time, repeated human decisions can also inform how the workflow is configured. If finance consistently determines that a particular pattern is legitimate, the organization may allow future examples to move through automatically while retaining a warning or audit trail.
Human judgment becomes an input to better orchestration, not a bottleneck placed in front of every transaction.
T&E is becoming a coordinated system
Travel and expense work rarely lives in one place.
A single trip can generate information across an itinerary, booking system, card program, email inbox, receipt image, expense platform, corporate policy, and approval workflow.
Historically, the employee has acted as the integration layer. People move between systems, interpret the information, and manually connect the pieces.
Agentic AI can change that relationship.
Because modern agentic systems can work with both structured data and human-readable information, they can interact with the documents, policies, descriptions, and systems organizations have already built for people.
That allows the employee or finance leader to operate at a higher conceptual level.
A finance leader should not need to know which database contains a particular field, which system owns the itinerary, or how several datasets must be joined to answer a business question.
They should be able to ask the question.
The agentic system can determine which permitted sources are relevant, retrieve the necessary information, connect it, and return an answer with supporting context.
That is the larger promise of agentic AI. It is not merely faster expense entry. It is a force multiplier that makes complex systems and information accessible to more people.
Less work around the work
The future of travel and expense will not be defined by a more impressive chatbot or an expense report with fewer fields.
It will be defined by how effectively technology coordinates the work.
Emburse Assurance already demonstrates part of this shift by bringing AI-powered guidance and contextual review into the expense journey before and after submission. Our work toward more reimagined, agentic experiences extends the same principle across capture, expense creation, policy, compliance, and review.
The system should prepare what it can, explain what it did, and bring people in when their judgment matters.
- For employees, that means less time reconstructing a trip.
- For approvers and auditors, it means better context and fewer routine distractions.
- For finance leaders, it means stronger control without creating more friction.
Agentic AI should not make people wonder whether they are still in control. It should give them a better way to exercise it.