The Medius source-to-pay agentic roadmap

This article provides a summary of a blog originally published on medius.com. To read the full-length blog, click here.

Most vendors announcing AI agents in 2026 are describing a destination. Medius is describing a journey it is already on.

Pick a random Monday in any AP team’s inbox. Supplier emails asking when an invoice will be paid. Mismatched purchase orders waiting for receipts that haven’t posted yet. Coding questions that need someone who knows the chart of accounts. Contract terms an approver wants to verify. Statements from suppliers that don’t reconcile. Anomaly flags from overnight processing. Somewhere in that pile, a genuine fraud attempt or a duplicate payment.

That scene plays out beyond AP too. Procurement teams have a parallel queue. Expense and payments teams have another. The detail differs but the pattern is the same: most of that work doesn’t need a human to start it, a lot of it doesn’t need a human to finish it, and almost all of it has been waiting for the kind of AI that can act on finance data rather than just talk about it.

That AI exists. The problem is that most of the conversation about it is misleading.

Why agent claims need a closer look

Vendors are announcing agents in 2026 the way they announced cloud in 2010: with a wave of marketing claims that obscure which capabilities are actually deployed, in which environments, with which accuracy, and against which audit trail.

The pattern that holds up under scrutiny is a maturity curve. Four stages, each one a precondition for the next, each one harder to skip than the marketing suggests.

Extract. Assist. Act. Orchestrate.

Most vendor agent announcements collapse all four into a single claim. The honest version describes what is live, what is in development, and what is on the roadmap, and ties each capability to the right stage of the curve.

Stage 1: Extract (in production)

Everything starts with extraction, and extraction is where most AI in finance fails quietly.

The reason is precision. An agent that takes action on incorrect extraction takes the wrong action. Every stage above this one inherits the accuracy of the data this stage produced. That dependency is why extraction isn’t a commodity problem to be solved with a general-purpose model. It’s the precondition everything else builds on.

Medius Capture runs at 96.3% touchless processing for top-performer customers on PO invoices. SmartFlow auto-fills coding, tax, and approver values at over 95% precision after learning from just two invoices from a new supplier. The pipeline runs roughly a thousand times faster and twenty-five times more cost-effectively than general-purpose large language models for the same extraction work, because extraction is a deterministic problem and deterministic problems are solved better and cheaper by purpose-built models.

The numbers matter less than the principle: extraction at this accuracy level is what makes the stages above it trustworthy. Without it, the agents act on bad data. With it, the data foundation is solid enough to build on.

Stage 2: Assist (in production)

Once the data is grounded, AI can be placed on top of it to help humans make better decisions. That is the assistant pattern, and it is the gating step between accurate extraction and genuine autonomy.

Medius Copilot is in production at more than 400 customers with over 3,300 users. It surfaces relevant invoice history, flags anomalies, summarizes context, and answers free-text questions from approvers. The large language model handles the language layer. The grounded finance data is what makes its answers correct rather than plausible.

This stage matters for a reason that goes beyond its immediate utility. Assistants build the organizational trust that finance teams need before they are willing to grant autonomy to agents. A finance leader who has watched their team use an assistant for six months has a much clearer sense of where the AI is reliable, where it isn’t, and what guardrails they want when they move toward action.

Skipping Assist means asking organizations to grant full autonomy on day one. Few will, and fewer should.

Stage 3: Act (in production and in development)

Act is where agents take real action on real work within bounded scope. Two agents are live in production today, and five more are in development across the source-to-pay lifecycle.

Supplier Conversations classifies the intent of vendor emails, pulls relevant context from the platform, and responds within defined guardrails. It currently handles around 13,000 supplier emails a month across 160 customers. AP teams using it report dropping from roughly eight hours a week managing supplier emails to thirty minutes. The thirty minutes left over is the small subset of supplier interactions that genuinely require human judgment, with routine work cleared out of the way.

Statement Reconciliation works alongside Supplier Conversations as a paired agent. When a supplier emails a statement, Supplier Conversations captures it. Statement Reconciliation extracts the data, matches each line against AP records, and categorizes the result: matched, exception, missing in statement, missing in system. Discrepancies route to a human within the workflow with a full audit trail attached. Two agents, one piece of supplier traffic, handled end to end before anyone touches it.

The five agents in development extend this pattern across the full source-to-pay lifecycle. PO Connect resolves invoice-to-PO mismatches without manual intervention. Payment Optimization tunes payment timing to capture discounts and balance supplier risk. Supplier Onboarding handles supplier data capture and validation through email. Contract Intelligence automates contract intake and extracts terms, obligations, and risk signals into structured data. Fraud Prevention stops fraudulent expense transactions before loss occurs rather than detecting them after.

The compounding effect shows up in invoice cycle time. Industry average sits at 9.2 days to process an invoice. Industry best-in-class is 3.1 days. Medius customers average 5.1 days on PO invoices. Medius top performers run at 1.4 days: 6.5 times faster than the industry average. Each stage of the maturity curve contributes a different compression to that number.

Stage 4: Orchestrate (on the roadmap)

Stage 4 is where individual agents become a coordinated system across the source-to-pay lifecycle. The Medius Agent Platform is on the roadmap, built around three pillars.

Agents Hub is where finance defines how work gets done. Policies, rules, and governance live here. When a controller updates an approval policy, the change propagates to every agent that needs to know.

Command Center is where finance supervises and controls outcomes. Decisions, approvals, and exception handling happen in one place. AI runs the work. Humans stay in control.

Mailbox Services is where agents intercept and act on requests directly from email, so work gets handled where it starts rather than requiring portals or manual handoffs.

Each pillar runs on the same data foundation, the same governance layer, and the same architectural pattern. Stage 4 is what binds the individual agents together. The agents do the work. The platform defines how, supervises the doing, and lets the work happen where it originates.

What this enables for the finance function

The maturity curve describes the agents. The wider question is what they enable for the finance function as a whole.

Touchless processing means most invoices, supplier interactions, and expense submissions clear without manual effort. The friction concentrates in genuine exceptions, and the platform routes those to a human only when risk, ambiguity, or policy exposure exceeds the threshold the controller defined.

Autonomous governance means policies enforce themselves. An unauthorized supplier bank-detail change isn’t an operational nuisance. It’s a balance-sheet risk. The platform monitors changes continuously, applies policy before execution, and pulls in a human only where verification is required.

Proactive optimization means the system models cash position, discount yield, and supplier risk, then acts within defined guardrails. Finance moves from operating the process to setting the policy that runs it.

This is what Spend Autopilot means in practice. Not replacing the finance team, but progressively embedding agent-driven capacity across the entire source-to-pay function so that the people running it are focused on judgment and strategy rather than administrative coordination.

Why the curve cannot be skipped

The reason most agent announcements in finance software in 2026 will not survive contact with production environments is not the model layer. It is the absence of the layers underneath.

An agent that acts on extraction it cannot trust takes the wrong action. An agent that calls a language model without a governance layer cannot pass an audit. An agent without an event-driven platform to act on can read state but cannot do work.

Stages 1 and 2 are the preconditions for Stage 3. Stage 3 is the precondition for Stage 4. The sequence is not a product philosophy. It is an operational reality.

Medius has solved each stage in sequence. Extraction has been in production for over a decade. Assistants have been in production for years. Two autonomous agents are live in AP today. Five more are in development across the lifecycle. The platform that orchestrates them is on the roadmap.

That is what makes the agent roadmap a credible extension of work already done, rather than a leap into something new.

Originally published on the Medius blog.