Agentic Finance Portal

Product vision

2026

Product Designer

I designed the next stage of Finance Portal: AI agents that prepare invoices and chase missing details, while Finance keeps every financial decision.

Yellow Flower

Agentic Finance Portal

Product vision

2026

Product Designer

I designed the next stage of Finance Portal: AI agents that prepare invoices and chase missing details, while Finance keeps every financial decision.

Yellow Flower
Yellow Flower

Agentic Finance Portal

I designed the next stage of Finance Portal: AI agents that prepare invoices and chase missing details, while Finance keeps every financial decision.

Product vision

2026

Product Designer

Yellow Flower
Yellow Flower

Project summary

Scope

I owned the concept end to end: reframing the problem, mapping the work, the model for splitting it between people, rules and agents, a working prototype in Cursor, and a phased rollout.

Challenge

The Builder redesign cut invoice creation from about 2 minutes to 20 seconds, but Finance's week didn't get shorter. The time had moved to finding information and chasing other teams across seven tools.

Solution

Agents do the preparation and chasing. The existing billing rules calculate every amount. Finance reviews, approves and sends. How much an agent can do alone is set per skill and capped by Finance policy, and every action shows up live in Watchtower.

Reframing the ask

Finance asked for search. The real need was less searching.

Finance asked for a global search. I treated that as a signal, not a spec. The usage data showed 898 Create clicks in 30 days, up to 7 tools per invoice and 15 or more Entities per person. Building an invoice was already fast. Getting everything it needed was not.

The clicks come from a preproduction sample, so they show how repetitive the work is, not how many invoices were sent.

Decision

I didn't design search. Search tells you where something is, and you still have to go and get it. I proposed that agents do the finding and chasing instead. I framed it as a hypothesis to test with them.

Outcome

It set the brief for everything after: give agents more of the work without taking any control away from Finance.

Mapping the real work

Only 2 of 11 steps need Finance's judgement

I followed serval invoices from due to sent. It took roughly 11 steps across seven tools: Finance Portal, the client's access plan, Client Control, Delivery, Slack, usage reports and Gmail. When a PO was missing, Finance asked CST on Slack, waited hours or days, then had to find the invoice again.

Decision

I sorted each step by the kind of work it was. Most follow rules Finance already stores, or are chasing and waiting on other teams. Only two need judgement: reviewing warnings on the invoices, and sending.

Outcome

This drew the line the rest of the design depends on: what an agent can take on, and what stays with Finance.

Deciding who does what

Agents prepare, billing rules calculate, Finance approves

Any agent touching invoices had to answer two concerns: accuracy, and who stays in control of what clients receive. One general assistant that "does billing" would be hard to trust and hard to audit.

Decision

Agents prepare and chase. The billing rules calculate every amount. Finance reviews, approves and sends. Each agent has one job, like resolving missing POs, and only starts when an event triggers it. Every action it takes is logged and can be undone. I prototyped it in Cursor so the handovers could be clicked through, not described.

Outcome

Agents can't change an amount, so Finance can hand over preparation without handing over financial decisions. Each agent's settings read like a job description: when it starts, what it works on and what it's allowed to do.

Choosing how much each agent can do alone

Agents can attach and draft, but only Finance sends

The real question wasn't whether to use agents, but how much each one should do alone. Attaching a usage report is low risk. Sending an invoice to a client is not, even when the same agent does both.

Decision

Each skill gets one of three levels: Flag, Prepare or Execute. Finance policy sets the highest level a skill can reach. Attaching the usage report and drafting the email can run on Execute, but sending is capped at Prepare.

Outcome

Finance can start a skill on Flag and move it up to Execute once they trust the results, for example attaching usage reports. Sending an invoice can never go past Prepare.

Letting Finance watch and step in

Finance sees every agent action and can take over

If agents arrive before the rules they depend on are reliable, one bad invoice could cost Finance's trust in the whole idea. People also need to see an agent working before they'll let it work alone.

Decision

I designed Watchtower so Finance can see every agent's work as it happens. Anything that needs a decision waits under Approval needed, and Finance can pause an agent for one invoice or take over completely. On the invoice itself, the Watchtower panel shows each step, where the agent stopped and why, and a full activity log, newest first.

Outcome

The concept builds on work already in place. Finance Portal stores invoice notes, billing cadence and invoice settings as structured data, which are the rules any agent would need to work from.

Measuring whether it works

Three measures, plus one guardrail

I set these measures before anything is built.

Cash in sooner. We'd track the days from billing date to invoice sent. Invoices that go out earlier get paid earlier.

More capacity. We'd track the hours Finance spends each cycle preparing and chasing. Fewer hours means the team can take on more Entities as the business grows.

Fewer errors. We'd track how many invoices are right first time, with no missing PO and no later correction. That means fewer rejected invoices, credit notes and chases to CST.

Baselines come from Phase 1, and targets are agreed with Finance. One guardrail runs throughout: the share of agent actions Finance undoes.