PromenautFrom assistant to workforce
The first wave of AI in financial services put a cleverer assistant next to an individual: a chatbot or a copilot that drafts, summarises or answers. Useful, but it reaches a thin slice of the organisation and it waits to be asked.
Agentic AI is a different shape. An agent plans a series of steps, calls the tools and data it needs, checks its work and returns a finished result. Compose several agents into a team, give the team a declared piece of work, and you have something closer to a digital workforce than a chatbot: breadth across functions, real outputs, run after run.
Where agentic AI already works in financial services
The useful question is not what an agent could do but where it does real, repeatable work today. Several areas of a bank fit, because the work is structured, document-heavy and high-volume:
Business case and investment evaluation
Examining a submission through a consistent set of lenses, surfacing gaps and assumptions, so a committee sees a structured assessment rather than a stack of decks.
Requirements and delivery
Reading a requirements set, scoring and tracing it, finding gaps and duplicates, and drafting user stories and acceptance criteria, through to the steps before coding starts.
Client and market intelligence
Turning scattered client, market and performance signals into usable insight on a schedule, rather than a one-off analyst scramble.
Risk, compliance and model governance
Drafting and checking documents against policy and controls, and keeping the evidence as the work happens.
Operations
Reconciliations, triage, and the document-heavy processes that absorb capacity without adding judgement.
These are not hypotheticals.
Promenaut is running in production with leading industry partners, including a global tier 1 bank.
Why it is hard in a regulated bank: the control problem
This is the part that stalls most agentic AI programmes in a bank, somewhere between a promising pilot and production. A firm can license the cleverest agent on the market and still be unable to run it, because it has no answer to the questions its own risk function, and its regulator, will ask: who did this agent act as, and could it reach more than that person may? Can you show what it did, for any run, including the ones that failed? Can you cap what it spends, and know the cost before it runs? Can you stop it cleanly?
Agentic AI is not mainly a model problem. It is a control problem, and the institutions that pull ahead will be the ones that can put a whole workforce of agents to work under the disciplines they already apply to their people and their models.
The scarce thing in a regulated enterprise is control that holds while the work scales.
How to run agentic AI safely in financial services
Running agents in a regulated setting works best when control is produced by the work itself, rather than reviewed or reported afterwards. Five properties matter:
Authority comes from a person
An agent acts as a deputy of the person who started the work, within that person’s permissions. An agent with its own broad credentials is a privileged account nobody owns.
People decide where it matters
Sensitive steps wait for a person, and a sensitive change is held for a checker who is not the person who started it, so no one approves their own work.
Evidence comes from the work
Each run leaves its own record: which agents ran against those approved, what each did against what it declared, and which model answered.
Cost is known before a run
Spend is a decision, not a discovery: a run costed before it starts and drawn against a budget.
The data stays where the bank chooses
A bank rarely wants to send its data to a public model service to make an agent useful. The workforce should run where the bank chooses, against the models it chooses.
How to evaluate an agentic AI platform for a bank
Does the work come shipped, and can you build your own?
A framework hands you parts to assemble; a workforce already does the work and still lets your engineers build.
Where does your data live?
In your estate, or inside a vendor’s walls where the agents only work while your data stays with them.
Is it neutral to the model and the estate?
Or does it tie you to one cloud or one suite.
Is governance produced by the work, or added on?
Audit logs assembled afterwards are not the same as evidence produced by each run.
Can you show it to your second line and your regulator?
Without being asked to take the vendor’s word for it.
Where Promenaut fits
Promenaut is an operating system for a governed digital workforce: teams of AI agents that do real work across a bank, held for a person where the workflow says, costed before a run, with a verdict from each workflow run.
It can run in the bank’s own cloud account, on the model providers it chooses, and the bank’s teams build their own agents and workflows on it.
Promenaut is running in production with leading industry partners, including a global tier 1 bank.
Common questions
What is agentic AI in banking?
The use of AI agents that carry out multi-step work rather than only answering questions: they plan, call systems and data, and return a finished result, and they can be composed into teams that run whole pieces of work across the bank.
What can AI agents do in a bank today?
Real, repeatable, document-heavy work: evaluating business cases, analysing requirements and drafting delivery artefacts, producing client and market intelligence, drafting and checking compliance documents, and operational processes such as reconciliations and triage.
Is agentic AI safe for a regulated bank?
It can be, if control is built into how the work is done: agents acting within the authority of the person who launched them, people deciding sensitive steps, maker and checker kept apart, cost known before a run, evidence from each run, and the data kept where the bank chooses.
How is agentic AI governed in financial services?
By the disciplines a bank already applies to models and people, model risk management, audit, operational risk and financial control, applied to agents, with the evidence to show what happened and the ability to stop work cleanly.
