Accounts payable offers agentic AI an unusually clear progression. A system can begin by answering questions, advance to recommending the next action, and eventually execute bounded parts of invoice processing under established finance controls.
Basware is following that path by adding agents to its invoice-lifecycle platform. The company’s long-term ambition is “100% automated, 100% compliant, and 100% protected” processing, but its first releases are more measured.
The AP Business Agent gives contextual guidance to people handling invoices, while the AP Data Agent lets employees question system data in natural language. Future supplier and professional agents are intended to manage disputes, payment enquiries, and processing questions.
Guidance and Data Access Remove Everyday Friction
Invoice work contains many small interruptions. An employee needs to know why an item is delayed, which approval comes next, or whether a supplier offered an early-payment discount. Answering often requires navigating reports, contacting colleagues, or interpreting process status.
The Business Agent recommends a next step using the invoice’s context. The Data Agent can answer questions such as which invoices await approval in a particular jurisdiction. This brings information into the workflow instead of forcing a user to leave it for a separate reporting tool.
These functions are useful because they target routine coordination rather than pretending to replace the complete finance department. Reduced follow-up can give accounts-payable teams more time for supplier issues, cash opportunities, controls, and analysis.
Performance should be measured through resolution time, query accuracy, approval delay, invoice quality, and the amount of manual searching removed. Natural language is only an improvement if the answer is grounded in authoritative transaction data.
Autonomy Gates Translate Policy Into Action
Finance teams will not delegate work unless authorisation, compliance, and auditability remain intact. Basware says proposed agent actions pass through a central policy engine that applies business rules, risk thresholds, and compliance requirements. The company calls these controls autonomy gates.
That architecture is more important than the conversational interface. An agent may propose contacting a supplier, changing a status, or resolving an exception, but the policy layer determines whether it is permitted to proceed.
Each gate needs a clear owner and test. High-value invoices, unusual suppliers, conflicting records, or regulated jurisdictions may require human approval. Lower-risk actions can proceed automatically once the system has shown stable performance.
Explainability also needs to be practical. Staff should see the records, rule, and confidence behind an action, not merely a polished explanation generated after the decision.
Basware cites paper manufacturer Billerud, which reports improved invoice quality and tangible savings. Other organisations should establish their own baselines rather than assume those results will transfer unchanged.
Supplier Conversations Raise the Standard
A planned Supplier Agent would contact suppliers about disputes and payment questions, summarise the conversation, and identify next steps. That is a larger delegation than answering an internal query because the agent represents the company to an external party.
Organisations will need controls for identity, approved language, recording, escalation, and commitments. A supplier should know when it is speaking with AI and have access to a person when the issue is ambiguous or consequential.
The quality of invoice data remains foundational. No policy engine can fully compensate for inconsistent purchase orders, duplicate supplier records, or missing approval rules. Agent deployment should therefore expose and improve the process beneath it.
A survey for Basware found that 61% of organisations had deployed agents experimentally, while one quarter of finance leaders did not fully understand what an agent looks like in practice. The company’s product direction offers a concrete answer: specialised software that receives context, proposes or performs work, and remains inside finance controls.
The path to touchless processing will not be one dramatic release. It will be a series of earned permissions. Basware’s agents become valuable when every new action removes friction without removing accountability.
