A Framework For Finding The AI Automation Candidates Inside Your Treasury Workflow

Most treasury teams already have between five and fifteen workflows that are strong candidates for AI agents. They usually know it intuitively. The work that repeats every period, slips first when the calendar fills, and produces the same shape of deliverable cycle after cycle is the work that fits the agent pattern.

The harder question is which candidate to build first, and how to scope it without spending a quarter on discovery.

This article walks through a five-step framework for finding those candidates inside your own workflow, ranking them, and picking the first one to build. It is the same framework Trovata's solutions team uses when scoping an agent with a customer, structured so you can run it on yourself today with a blank document and the work that is on your calendar this week.

Why treasury is ready for this now

The agent pattern became practical only recently, and not because the idea is new. Three things had to line up first.

Cash data had to become normalized at the source. For most of the last three decades, balances lived in fragmented bank portals, each with its own format, and cleaning that data consumed the first hour of every analyst's day. Direct API and SFTP connections now standardize cash data on the way in, so the analysis above it runs against one structured layer instead of a manual consolidation.

Language models had to grow capable of applying judgment to structured financial data, rather than only summarizing text. And cloud architecture had to mature to the point where those models could operate inside a controlled set of tools with an auditable runtime.

Those conditions are now in place, and treasury teams are acting on them. The share of treasury practitioners expecting AI to improve cash forecasting rose from 65% in 2024 to 76% in 2025, according to Strategic Treasurer. The question has shifted from whether to use AI in treasury to what to build first.

What makes a treasury workflow a good AI agent candidate?

A treasury workflow is a good AI agent candidate when it repeats on a schedule or trigger, reads from connected data, and produces a defined deliverable a treasurer reviews.

That is the shape of an agent: it runs when something happens, reads the data in scope, applies a written prompt that captures how a senior practitioner would do the work, and returns a finished output. Work that depends on context, judgment, or unstructured inputs belongs in a different category. The framework below finds the work that fits the pattern and ranks it.

The five-step framework

Five steps move a candidate from idea to first production run: inventory, score, qualify, sequence, and operate. The whole exercise fits inside an hour and produces a ranked list of candidates plus a clear pick for the first one to build.

Step 1: Inventory the work that repeats

List the recurring work your team performs over a typical month, across daily, weekly, monthly close, and quarterly cadences, and do not filter yet. A useful prompt is to ask what a new analyst would need to learn in their first 90 days that gets repeated, because that question surfaces the work that has standardized enough to be teachable.

Aim for ten to twenty-five items. Teams that claim they only have three or four candidates have usually missed the daily and weekly work, including bank reconciliations, position assembly, variance attribution, sweep reviews, and concentration checks. None of it shows up on a project plan because it just happens, and all of it fits the agent pattern.

Step 2: Score every task on four dimensions

Rate each task one to five on four dimensions:

• Frequency. How often the work runs across a year. Daily work compounds fast, while quarterly work runs fewer cycles though the hours per cycle may be higher.

• Time cost. Analyst hours per cycle, which becomes an annual number when multiplied by frequency. Thirty minutes weekly is 26 hours a year; five hours monthly is 60. Both are worth recovering.

• Error risk. What fails when the work is rushed or skipped. Missed disputes are dollars left with the bank, late variance commentary slows the CFO's read, and broken reconciliations create downstream problems at close.

• Data availability. Whether the input already lives in connected, structured form. Tasks fed by bank APIs, normalized AFP codes, or ERP feeds are ready to scope; tasks fed by desktop spreadsheets need a data connection built first.

The goal is not the single highest score. It is the cluster of tasks that score high across several dimensions, because those are the candidates worth scoping further.

Step 3: Qualify against three tests

A task is a strong agent candidate when it passes three tests:

• It has a clear trigger, either a schedule, such as every Monday at 7am or the first business day of the month, or an event, such as a balance crossing a threshold or a forecast version being approved.

• It has a defined output, such as a position narrative, a findings report, a draft email, or a journal entry. Work whose output is "an analyst thinks about it and decides what to do" is not ready as a scheduled agent yet.

• Its input data is connected and structured. Bank balances flowing through APIs, fee statements ingesting against AFP codes, and forecast versions stored in your platform all qualify.

A task that fails one test can still be a strong candidate later. It is just not the first one to build.

Step 4: Sequence by impact against build effort

Rank the qualified candidates by impact against build effort, and make the first agent high impact and low effort: high time cost, high skip rate, simple data inputs, and well-understood ranking logic. Bank fee analysis is the canonical first agent for most teams, with forecast variance commentary a strong second.

Save the more complex candidates for the second and third agents. Once one agent runs and the team trusts it, the muscle exists to scope a harder one, and the compounding value of agentic treasury tends to come from the agents that follow the first. Starting with a hard candidate risks the whole approach if the prompt does not converge cleanly.

Step 5: Operate with human-in-the-loop checkpoints

Define the checkpoints before the first run, not after. Anything that touches a bank or external counterparty stays human, including disputes, hedge executions, and payment initiations. Anything that changes policy stays human, including investment policy, hedging ratios, and concentration limits. Anything that needs context the agent does not have stays human, such as a verbal commitment that pre-dated billing.

Run the agent in parallel with the manual analysis for the first cycles, compare the outputs, and tighten the prompt against the gaps between them. By the time the team trusts the output, the analyst's job has shifted from production to review and exception handling. The methodology that emerges through that tuning is what makes the agent specific to your team rather than generic to the platform.

A worked example: bank fee analysis end to end

The framework applied to bank fee analysis shows how the five steps connect.

1. Inventory. A treasury team lists monthly fee reconciliation, at four to five hours per relationship across six relationships, roughly 30 hours a month, often skipped when the calendar tightens.

2. Score. Frequency monthly, time cost about 360 hours a year, error risk high because every missed dispute is dollars left with the bank, and data availability already in place through fee statements ingested via SFTP and normalized against AFP codes.

3. Qualify. Trigger is the first business day of each month, output is a findings report with dollar impact and dispute drafts, and the data is connected and normalized. All three tests pass.

4. Sequence. Highest combined score in the inventory, so it gets built first.

5. Operate. Early cycles run in parallel with the manual audit, the prompt is tightened against the gaps, disputes go out under treasurer signature, and ECR target decisions stay with the treasurer.

The team moved from running fee analysis once a quarter when time allowed to a full audit every month, on every relationship, with no incremental manual work above the review.

What the exercise gives you

An hour with this framework produces three artifacts. The first is a written list of every recurring task on your team's calendar, which most teams live with daily but have never written down. The second is a ranked list of agent candidates with the math behind the ranking, which any solutions conversation will need anyway. The third is a clear pick for the first agent to scope, with its qualifying tests passed and its impact-against-effort math done.

Run it on a Monday morning before the calendar fills. The framework is built to produce something useful even if you stop after Step 2 and go back to your day.

Where to go deeper

Pick one task on your calendar this week, and the next time it comes up, run the five steps against it. The point is not to scope a complete agent yet. It is to feel how the framework changes the way you look at recurring work. After the first pass, you start spotting candidates everywhere.

The full AI Agents Playbook for Treasury Teams covers this framework in more depth, with bank fee analysis worked from inventory through first production run, plus a library of ten agent patterns Trovata customers deploy today, each with the prompt that drives it and a representative output.

Read the AI Agents Playbook for Treasury Teams to scope your first agent with the full reference open on your second monitor.

Frequently asked questions

What is a treasury AI agent? A treasury AI agent is a scheduled or event-triggered process that reads from connected treasury data, applies a written prompt modeled on senior practitioner judgment, and produces a finished deliverable the treasurer reviews. It is not a chatbot or a dashboard with alerts; it is shaped around a specific recurring task.

How many workflows can a treasury team realistically automate? Most treasury teams have between five and fifteen workflows that qualify as agent candidates, spread across daily, weekly, close, and quarterly cadences. The framework surfaces them and ranks them so the team builds in priority order rather than all at once.

Which treasury workflow should we automate first? Bank fee analysis is the canonical first agent for most teams, because it is high time cost, frequently skipped, fed by structured data, and built on well-understood ranking logic. Forecast variance commentary is a common strong second.

What work should stay with a human? Anything that touches a bank or counterparty, changes policy, or requires context the agent does not have stays human. The agent assembles, analyzes, and drafts; the treasurer reviews, approves, and decides.

Do we need new data infrastructure before building an agent? A task only qualifies when its input data is connected and structured, so tasks fed by bank APIs, normalized AFP codes, or ERP feeds are ready, while tasks fed by desktop spreadsheets need the data connection built first. That connection is a precursor, not a reason to abandon the task.