For years, reconciliation has leaned on automation to deliver efficiency and scale. But as operational complexity deepens, settlement cycles compress and regulatory scrutiny intensifies, the conversation is moving on. The question is no longer simply ‘how do we automate more tasks?’, it’s ‘how do we build genuinely autonomous operating models?’.
That was the theme of AutoRek’s recent webinar in partnership with InvestOps, where Jim Sadler, Chief Transformation Officer at AutoRek, joined Mike Leader, Global Head of Investment Operations at CPP Investments, for a candid discussion on where reconciliation is heading and how firms can get there safely.
Why now?
If automation isn’t new, what has changed to make firms seriously consider autonomous models today? For Jim, it comes down to three converging forces. First, business context: the sheer growth in data sets, volumes and counterparties has begun to outpace human scalability, leaving firms with little choice but to look at what AI can do. Second, the technology itself has taken remarkable recent strides, with language models shifting from conversational tools to something genuinely “operationally usable” in a matter of months. And third, the quiet but critical progress in interoperability, such as the Model Context Protocol, which Jim likens to the convergence around a “USB-C plug” of the agent world: a standardised way for agents to interact with one another and with APIs, consistently, safely, and at scale.
From the operational front line, the picture is much the same. Firms have already automated the straightforward, repetitive tasks, he explained. What remains are the fragmented workflows, unstructured data and judgment-heavy exceptions, precisely the areas where incremental automation runs out of road. Add compressed timelines and pressure to scale without growing headcount, and the conversation naturally shifts from automating individual tasks to bringing entire workflows together more intelligently.
The operational reality
Organisations have expanded across asset classes, counterparties, fund structures and data providers, yet their operational architecture has evolved incrementally rather than being systematically redesigned. The result is fragmented data, inconsistent sources and a heavy reliance on spreadsheets and manual intervention, all squeezed by settlement compression that has removed any tolerance for delay. Mike specified that for him, it’s the collision between growing complexity and legacy operational design.
Jim proposed the underlying problem: in many teams, the humans have effectively become the middleware, stitching together disparate systems and contexts. Even a 99%+ match rate still leaves thousands of unmatched items and it’s here, he argued, that AI offers the chance to go “the last mile,” removing the low-value but highly human-dependent investigative effort that burdens teams and holds back scale.
What agentic AI looks like in practice
Both speakers were pragmatic about where the technology genuinely adds value today. We are firmly in the augmentation phase, AI agents assisting humans rather than replacing them.
The most realistic early use cases are those that reduce friction without immediately taking humans out of the process: intelligent exception prioritisation, automated investigation summaries, pattern recognition across recurring breaks, and recommending resolutions based on historical outcomes. Perhaps most compelling is AI’s ability to surface the unstructured operational knowledge locked away in emails, commentary and tickets. The context that teams currently spend so much energy chasing down.
There is, however, an important nuance: not all intelligence needs to be artificial. Rule-based workflows and deterministic matching engines already handle matching brilliantly and economically at massive scale. AI’s real contribution is to streamline the “human middleware”, shortening the lifecycle after a break occurs and enabling new reconciliations to be built quickly and autonomously, so that onboarding new trading partners, asset classes or counterparties no longer acts as a barrier to scale.
Defining true autonomy
So what separates a genuinely autonomous workflow from advanced automation with AI bolted on? The answer is goal orientation. Autonomous workflows are driven towards achieving a goal rather than following fixed rules, they can determine the best path, track their progress without losing context, validate when the goal is met, and, crucially, know when to ask for human help rather than getting stuck in a loop.
That’s a high bar, particularly in reconciliation, which Jim described as a deterministic discipline: “nearly right, is wrong in this world.” You must be able to prove what you did, how you did it and why you reached a conclusion. Working to high probabilities simply isn’t good enough. Mike reinforced the point: automation executes tasks, but autonomous systems manage decision flows within defined guardrails and autonomy should never be mistaken for uncontrolled independence.
Governance, trust and the human role
As systems begin acting more independently, governance and trust remain as critical as ever. Mike was clear that governance can no longer be treated as a separate compliance layer bolted on afterwards; it must be built into the architecture itself, with clear boundaries around what decisions AI can make, where escalation occurs, and where humans remain firmly in the loop. Explainability and auditability are non-negotiable in financial services.
Jim reframed trust as something profoundly human. We’ll learn to trust AI the same way we learn to trust any technology or each other: through transparency, predictable behaviour, graduated progress and continual reliability over time. Far from making people irrelevant, he argued, AI will make them more relevant, albeit shifting from operators to orchestrators. His analogy resonated: we don’t fly every plane manually, yet we still have pilots in the cockpit and air traffic control ready to intervene. Humans can’t divest accountability to AI, so the challenge is developing that “air traffic controller” perspective for operational workflows.
Looking ahead
In an industry moving in months, if not days, 3-5 year predictions prove almost impossible. Still, some themes stood out. Fully “lights-out” finance operations remain unlikely, but highly autonomous workflows and real-time exception management look increasingly probable as the industry moves towards T+0. There is even the prospect of ledgers and reconciliations merging, dissolving the siloed system boundaries that were only ever drawn to organise thousands of humans in the first place.
Reconciliation also looks set to shift from reactive to continuous and even predictive, with the function broadening to absorb operational risk, workflow management and data quality into a more integrated whole. Teams may become leaner, but roles will grow more complex, demanding a blend of analytical capability, process understanding, governance awareness and genuine technology fluency.
Where to begin
For firms starting this journey, Mike’s advice was begin with the operational pain points, not the technology. Invest early in standardising process and improving data quality, however unglamorous that work may be; and adopt a progressive approach, building organisational trust through small, meaningful wins. Above all, involve the people doing the work in the design process as they understand the friction better than anyone.
Start small, but never trivial. Take away the work no one enjoys, and you’ll earn a genuine vote of confidence. Get comfortable with a “try, fail, learn, retry” mindset borrowed from software engineering. And don’t discard the software already delivering value in pursuit of AI for its own sake. Perhaps most valuably: talk to your vendors about their R&D.
Reconciliation is evolving from a traditional control process into a genuinely intelligent, adaptive operational capability. The firms that thrive will be those that pair the right technology with thoughtful operational design and robust governance.