The AI Gap in Asset Management
Most asset managers run AI as a series of isolated projects, and I say that with sympathy because it is the natural way these things start. A team has a question, someone builds a model to answer it, the model works, and it goes into production as a one-off. Then the next question arrives and the cycle repeats. After a few years you have a handful of models, each with its own story about how it gets trained, where it lives, and which engineer is afraid to touch it. No shared infrastructure connects them, so the team relearns on the fourth model the lessons it paid for on the first.
Poor AI infrastructure looks fine right up until it doesn't. A team builds a model, validates it, deploys it, and on day one it is good. The trouble arrives months later. Ask which version is running in production and you get two different answers. Ask what data went into training it and someone has to go looking. Retraining is a project, so it happens once a year if at all, and as client behaviour moves on the predictions get worse. There is no crash and no alert. The predictions drift from useful to plausible to misleading while the confidence scores stay high.
In asset management this costs more than it does elsewhere, because model outputs drive commercial decisions: which clients your sales team calls this week, which relationships you escalate as at risk, and whether two records arriving from different platforms are the same client. A model that fails loudly gets fixed. A model that fails quietly reallocates months of human attention before anyone notices. And in a business built on evidence, a head of distribution will not act on an insight they cannot trace.
"A model that fails loudly gets fixed. A model that fails quietly reallocates months of human attention before anyone notices."
What drew me to Aiviq was the chance to build the whole loop, from training to monitoring, instead of inheriting fragments of one, and, of course, the data. Aiviq had solved the hard, unglamorous problem of ingesting and conforming AUM and flow data from a long list of platforms and transfer agents. A clean, longitudinal dataset like that is the best starting point machine learning can have, and it deserved better infrastructure than a folder of model files.
What We Built: Aiviq's AI Infrastructure
We built the machinery around the models. Today it serves a suite of production models, each answering a different commercial question about a client relationship, and each one runs on the same rails. Training runs as a reproducible pipeline inside each client's environment and we log each run in full. A model registry records which version is in production, and a promotion gate decides whether a new candidate earns its place. A nightly pipeline scores fresh data, and a monitoring layer watches the lot. We could rebuild any one of the models in a quarter; the lifecycle machinery is where the durable value sits. We host all of it ourselves in dedicated, locked-down environments, one per client, so no client's data shares infrastructure with another's. We ruled out SaaS machine learning platforms on day one, because client relationship data does not leave those environments.
How We Train Our Models, and Keep Them Sharp
The models train on the same conformed AUM and flow data that powers the rest of the platform, so there is no separate "data science copy" of the truth drifting away from what clients see. On top of it we engineer behavioural, transactional and structural signals for each client: flow momentum over several windows, streaks of consecutive redemptions, the breadth of strategies held, and flags for outsized redemptions. A single model can draw on around two dozen of these. An engineer starts a training run with one action, the job executes inside the cluster, and we log it in full so we can reproduce any result.
Continuous training in practice means two loops. The scoring loop runs itself each night: the pipeline pulls fresh data and writes fresh predictions without anyone touching it. The training loop is deliberate but cheap. We retrain when there is a reason to, whether that is new data or a dip in performance, and because the whole path is automated, we do it often. A training run cannot touch production. It produces a candidate, a registered version that still has to prove itself.
A candidate then faces the current production model in a straight contest. The gate builds a fresh evaluation set, scores champion and challenger on the same rows, and judges them on the one metric that matters for that model's job. The candidate that wins takes over production; the one that loses stays a candidate. If a promoted model misbehaves, one action rolls back to the previous champion, and the system logs each promotion and rollback with who triggered it and why. Users can also flag individual predictions they disagree with, which shows us where the models are weakest.
"Most of the industry retrains on a calendar and promotes on faith."
We made retraining cheap, so it happens often, and we made promotion depend on evidence, so it is safe. An engineer decides when a new model is worth considering. The gate decides whether it ships.
Automated Monitoring: Reliability Without the Overhead
The monitoring layer works at three levels. Pipeline health tells us whether the nightly scoring ran, how long it took, and whether training jobs finished. We classify failures: a job killed for exceeding its memory allocation shows up as a memory problem, so the on-call engineer does not have to decode a generic error. Model health covers the metrics from each training run, the gate's evaluation results, and walk-forward back-tests against naïve baselines, because a model that cannot beat a simple rule of thumb has no business in production. The third level, governance health, is the one few teams build: an automated audit checks that the model in production went through the promotion gate, and flags any version that someone promoted by hand.
Before this existed, keeping models healthy meant checking pipelines by hand, reading logs to work out why a job died, asking around to find out which version was live, and regression-testing each candidate yourself. We automated all four. The hours we got back matter, but the bigger change is in behaviour: releasing a new model used to be the part of the job you dreaded, and now it is a routine action with a safety net, so you improve things you would once have left alone.
My favourite catches are the ones that would have slipped past a human reviewer. The gate refuses to evaluate a model on a defective data window: if the evaluation slice is too small, or one class of outcome is missing from it, the gate stops and reports the problem, so a data glitch cannot wave a bad model through. A corrupted metric raises an error before the gate compares it with the incumbent. And failure classification once turned what would have been an afternoon of log forensics into a one-line diagnosis that pointed straight at the job's memory limit.
Each number in the product now arrives with its provenance: the model version that produced it and the gate decision that let it into production. We can show a sceptical user the chain of evidence behind an insight, and people act on a number far sooner when they can see where it came from.
Why Aiviq Intelligence Stands Out
"None of our tools are secret."
The hard part to copy is the combination of the data and the discipline. Aiviq holds conformed distribution data from a fragmented set of providers, clean and consistent over years, and you cannot buy that off the shelf. Our models keep proving themselves against that data through the gating loop, and each new model or use case lands on rails we have laid. A competitor can stand up a model registry in a week. Reaching the point where retraining is routine and the promotion loop has run long enough to trust takes years.
Asset management shaped most of our design choices. Distribution data arrives in daily and monthly batches, so we built a batch-first architecture and skipped the fragility of live model serving. And because this business has to evidence its decisions as well as make them, we designed the audit trail and promotion records in from the first week. A compliance reviewer can read the full history of any model without us assembling it after the fact.
For a client, the difference shows up in timing and in trust. Aiviq Intelligence risk models look three and six months ahead and the pipeline rescores them each night, so a relationship manager hears that an account is weakening while there are still months to act on it, instead of reading about it in a quarterly review after the assets have left. The signals stay sharp, because new candidates keep challenging the production models. And a head of distribution can take a recommendation into a meeting knowing the trail behind it: which model produced it and why the gate trusted it.
What Excites Me Most: What's Next
The things I am most excited about is widening what the system can learn from, and widening what it can do with what it knows.
- The first thread is unstructured data. Much of asset management's commercial reality lives in distribution agreements and rebate letters, so we are building document extraction that turns those PDFs into structured terms, with a confidence score on each field and a citation back to the clause it came from.
- The second thread is feedback. Users can tell us today when a prediction looks wrong; the next step is to feed realised outcomes back into training and gating, such as whether a flagged client did redeem or a flagged opportunity did convert, so the models learn from their own track record. The training loop also gets more autonomous, with retraining on a schedule and the gate running the routine checks, which leaves our engineers' judgement for the cases where it changes the outcome.
All of this points towards agentic AI. The next generation is an agent that takes a signal, assembles the evidence behind it from the client's flows and from what their agreement says, and hands a salesperson an explanation and a proposed next step. The human still makes the call wherever the decision matters; they start from prepared evidence instead of a raw score. Over the next three months we plan to move document extraction from pilot into the product, wire realised outcomes into training, and put the first agentic workflows in front of users. The industry has not closed the gap between prediction and action. Aiviq owns the data, the infrastructure and, soon, the terms, and that is what closing it takes.
.png)
* All model training described here takes place inside each client's dedicated, secure Aiviq environment, and client data never leaves it.
