Skip to content
Promenaut

Promenaut joins forces with HSBC as strategic investor and design partner.Read More

Agentic workforce

Every AI run should make the next one smarter. Does yours?

Every AI run should make the next one smarter. Most don't. Here's why, and what it takes to fix it.

Stephen Murphy
Stephen Murphy
12 November 20265 min read

The organisations that pull ahead won't be the ones with the most AI agents. They'll be the ones whose workflows get measurably smarter every time they run: where each case handled makes the next one faster, cheaper, and less dependent on a specific person being available.

A large bank handles the same work thousands of times a year: onboarding a client, resolving a trade break, implementing a software change request. Many of those workflows now include AI. Yet ask what the bank knows after the thousandth client onboarding that it didn't know after the first, and the answer is usually very little. What it does know lives in someone's head, a post-mortem nobody reads, or a Confluence page last updated two years ago. The AI agent that ran step three has no idea what the agent running step seven learned. Neither does the person who picks up the exception.

The raw material for learning is already there. Every case produces it.

What every case produces

Every case produces signal that most workflows throw away. An exception reveals a scenario that was not designed for. An override shows where an experienced person set the rule aside, and why. A long resolution time exposes a handoff nobody owns. A human correction to an AI output is, technically, a labelled training example, specific to that bank's processes, that asset class, and that regulatory context. Together these are the most accurate and proprietary picture of how work actually runs. Almost no workflow is built to retain it in a form the next case can use.

Say an engineer resolves a production incident at 2am: a batch job fails whenever an upstream feed arrives late. The fix goes into a post-mortem. The runbook doesn't change. The knowledge lives in that one engineer's memory and a document nobody finds under pressure. Months later, the next engineer, or the next AI agent, hits the same failure and solves it from scratch. The resolution path, the edge case, the judgement call: none of it transferred. The bank keeps meeting the same problem for the first time, and paying to solve it again.

The unit that learns has to be the whole workflow, including the handoffs where most lessons are lost. Each time the workflow handles a case, it should record what it found, what a person decided, and why, in a place the next case can read.

Why AI agents on their own don't get you there

Point solutions don't fix this. An AI agent optimises the step it owns. It has no view of the workflow upstream or downstream, no mechanism to share what it learned with the agent in the next function, and no way to record its reasoning in a form that survives a model update. Ten AI agents deployed across ten functions leave a bank with ten isolated productivity gains, ten disconnected memory stores, and no compounding. The whole is worth no more than the sum of its parts.

Learning needs governance

Compounding works in both directions. A workflow that records the wrong lesson repeats it faster. In a regulated institution, an AI output that improves over time without an approval trail is a model-risk and audit liability, because a regulator will ask what changed, who approved it, and what it was based on. That means every lesson the workflow captures has to be treated as a controlled change: reviewed, approved, traceable to its source case, and visible only to the people and agents with entitlement to see it. The governance layer isn't a constraint on learning. It's what makes the learning usable.

Keep the learning above the AI model

AI models will keep changing, and banks will want the option to switch between providers. What doesn't change is what the organisation knows about its own work: its process patterns, its exception logic, its regulatory interpretations, its hard-won edge cases. If that knowledge lives inside a single model's context window or a single vendor's product, it leaves when the model does. The right architecture keeps institutional knowledge in a governed layer the bank owns: a knowledge graph that any model, any agent, and any person can draw on, with appropriate controls. The model is a consumer of that knowledge. It is not the place where knowledge lives.

Where the advantage sits

When every bank has access to the same foundation models, model access stops being a differentiator. The advantage moves to data and memory: specifically, to who has built the richest, most governed record of how their work actually runs. That record can't be bought off the shelf. It accumulates case by case, workflow by workflow, correction by correction. Banks that build it now are building something competitors can't replicate by signing a new vendor contract. Banks that don't are funding an AI estate that depreciates every time the model changes.

That is what Promenaut is built for. People and AI agents working as one governed workforce, where every case handled adds to a knowledge graph the bank owns, and where every future case draws from it. Not a point solution. Not a wrapper on a foundation model. An operating layer that makes the whole workflow smarter, with the audit trail a regulated institution can stand behind.

Bring us one workflow you own and we'll show you what it looks like on the platform: hello@promenaut.ai.

One question to take back

If your bank handled its most important workflow a thousand more times, would the thousandth case be handled measurably better than the first? Faster, more accurate, with fewer exceptions, less escalation? If not, the learning isn't in the system yet.

Related reading: Constraint Migration: Why Faster People Don't Make Faster Companies · Tribal knowledge: the AI unlock, not the hidden cost

#agentic workforce#knowledge graph#enterprise AI#institutional knowledge#governance#banking
Stephen Murphy
Stephen Murphy
CEO and Founder

Entrepreneur, technologist and founder. My background combines deep technical roots with real-world operational leadership. I’ve held senior and C-level roles at Goldman Sachs, Merrill Lynch, HSBC, and BTG Pactual, and operated across the world’s key financial centers -New York, São Paulo, Hong Kong, and London. Now focused on advising, investing in, and launching new ventures - particularly where AI, developer productivity, and financial innovation intersect. I bring a builder’s mindset, proven execution across multiple markets, and a strong global network of investors, founders, and enterprise leaders.

Rather see it than read about it?

No self-serve trial. Every engagement starts with a scoped conversation.