Most conversations about AI inside an investment firm start somewhere in the middle. There is usually a pilot that has been running for a few months, a small group of analysts who have quietly built their own habits around ChatGPT, and a set of internal expectations that nobody has written down. The question on the table is what to do next, and the useful answer depends almost entirely on which stage of adoption the firm is actually in.
Across dozens of deployments at asset managers, mutual funds, and hedge funds, the sequence turns out to be remarkably consistent. In most cases it begins with a specific person getting tired of a specific piece of work, and the organisation then slowly reorganising itself around what happens next.
What follows is that sequence, in the order it tends to occur, followed by the single habit that determines whether a firm gets through it or stalls somewhere in the middle.
Successful deployments almost always start because somebody inside the firm has grown tired of a workflow that no longer makes sense. Sometimes it is a head of research watching thirty analysts lose another earnings season to repetitive work. Sometimes it is an operations team buried under DDQ responses. Sometimes it is a portfolio manager spending more time searching for information than making decisions.
It begins, in other words, with a specific pain and a named owner rather than an AI strategy document. That distinction matters more than it sounds like it should. A strategy document has no natural stopping point and no obvious way to be judged, so it tends to expand until it covers everything and commits to nothing. A single workflow with somebody's name attached to it has both.
The step that gets skipped most often here is measurement. Before anything changes, the team should write down what the workflow currently costs. How many hours does it take? How many people touch it? At which step do mistakes usually happen, and what happens downstream when they do? Those numbers become the baseline, and they are worth capturing even when they feel imprecise, because once the pilot begins, every discussion about whether it worked eventually comes back to those measurements.
Firms have lost genuinely good deployments at this stage, not because the system failed but because nobody could say what it had improved. The analysts liked it, the output looked better, the hours seemed lower, and none of that survived contact with a procurement review. Ten minutes of counting at the start is usually the difference.
A regulated asset manager runs on information boundaries that have existed for decades. Every piece of sensitive data carries strict rules about who is allowed to read it, where it is allowed to live, and what has to be recorded when somebody looks at it. When new technology enters that environment, those access rules end up dictating the entire build.
Before anyone discusses prompts or models, compliance teams need to verify where the system runs and whether internal data stays inside the firm's walled garden, so to speak. Turning those regulatory requirements into software is mostly an upstream architectural problem. The deployment may need to be hosted inside the firm's own cloud environment. It may need permissioning that mirrors the access controls that already exist, so that the system answers a junior analyst differently from the CIO and never carries a restricted document across a wall. In some cases the constraint is physical, and the deployment has to sit on hardware inside the firm's own data room.
These are structural decisions. They narrow the field of possible vendors long before anyone compares model quality, which is why bringing compliance in during scoping tends to save so much time. The constraints get settled while they are still cheap to accommodate.
Groups that leave this approval until the end usually discover that their architecture needs something like three to six months of rework to meet basic standards. That is the part worth flagging most directly. While a fast prototype tends to generate real excitement, the reality of data residency, permissioning, and audit trails still has to be solved before anything reaches a production-grade rollout, and solving it afterwards means solving it twice.
Once the pilot goes live, the system leaves the controlled environment in which it was built and starts encountering real work. A single team uses it every week. Edge cases appear almost immediately. A table gets parsed wrong. A segment that was renamed mid-year gets treated as continuous. An analyst points out that the system is missing a piece of context that everyone on the desk carries in their head and nobody has ever written down.
The builders then spend the first few months refining retrieval, prompts, and workflow logic. That period can feel chaotic, particularly to a sponsor who was hoping the hard part was over at signing. It tends to be a healthy stage of the deployment. The complaints mean people are actually trying to rely on the system rather than treating it as a demonstration, and a system nobody complains about after launch is usually a system nobody is using.
The part worth calling out here is what to watch during those months. The number that matters is how quickly the workflow starts paying people back, measured against the baseline from stage one, which tells you far more than an abstract accuracy figure would. A system that gets 70% of the work done reliably, with the analyst reviewing rather than rewriting, is further along than one that occasionally produces something brilliant and needs to be checked line by line every time.
Deployments usually begin with a single team and one painful workflow, and a year later the same deployment is often serving desks nobody had even met at the initial signing. Adjacent teams see the output and start surfacing their own requests. Sales wants meeting preparation. Compliance wants disclosure checks. Strategy wants competitor tracking. Client servicing wants a faster way to answer the same forty questions.
That expansion usually comes from proximity rather than from any internal push. Folks see another team using the system every day, hear that it saves time, and begin asking whether it can help with their own work. One workflow becomes two, two become five, and gradually the platform grows into something much larger than the original deployment.
This is also why measuring the first workflow matters as much as it does. Every conversation about expanding the system eventually comes back to the same question, which is what happened the first time. A clear before-and-after answer gives the next team the confidence to try, and gives whoever controls the budget something to point at. Without it, each new workflow has to be argued from scratch as though the firm had learned nothing.
The deployments that stick usually pair a senior research sponsor, someone who owns both the workflow and the outcome, with one technical builder and one person from data or infrastructure. That small group tends to outperform a standalone AI department, because every decision stays anchored to work somebody in the business actually cares about.
Hiring a Head of AI before mapping any workflow tends to create a role with no operational surface to work on. The person arrives, spends a quarter surveying the vendor market, and produces a plan that the desks have no particular reason to adopt. The sequence that seems to work better starts with one or two workflows, automates one of them, and lets the results justify the additional headcount organically.
Firms also tend to underestimate how much of the required talent already sits inside the building. The analyst who has spent years maintaining sector models in Python often understands the firm's workflows and proprietary data better than most external hires would after a year, and that person is frequently the natural builder of these systems. One needs to look for them before writing a job description.
Reading back through the sequence, every stage quietly depended on somebody being able to see a job as a series of steps. Stage one required identifying a workflow and measuring it. Stage two required knowing which steps touch sensitive data. Stage three was step-level debugging. Stage four was the same exercise repeated for a new team. Stage five was staffing against it.
Underneath all of that sits a single skill, best described as thinking in workflows. Of everything in this piece, it is the habit most worth taking away.
Thinking in workflows means taking a job the team already performs, writing down the steps it actually breaks into, and then deciding at each step whether software, a model, or a person should own it. That is the entire exercise. It sounds almost too simple to be called a skill, and it is still the clearest dividing line between the teams that get AI into production and the teams that stay stuck in the pilot phase.
Part of the reason it works is that it changes what the firm is buying. A firm that has not done this exercise goes to the market asking for an AI product, which means evaluating vendors on demos and claims. A firm that has done it goes to the market with six specific steps and a view on which of them it wants automated, which is a much easier thing to test, price, and hold somebody to.
Earnings coverage is a useful example because almost every public-markets team spends a meaningful part of the year there, and because the workflow is familiar enough that the exercise is easy to follow.
Written down as steps, the job usually breaks into something like the following.
Once the workflow is written down, assigning ownership becomes much easier. Software handles steps one and six, the pulling and the routing, because they are deterministic and there is no reason for a model to be involved. Models handle steps two, three, and five, the extraction, the comparison, and the first draft. The analyst keeps part of step four and owns everything that touches an actual investment view, which is to say the judgment about what the changes mean.
One large problem has quietly become six much smaller ones, and each of them can now be measured, improved, and evaluated independently. That last point is the practical payoff. When the system produces a bad note, nobody is debugging "the AI." The question becomes which of six steps degraded, and usually the answer is one of them.
The same exercise works just as well for morning meeting briefs, DDQ responses, portfolio commentary, factsheet production, or post-call synthesis. The steps change, the method does not.
Investment teams are better positioned for this than they tend to assume, because the industry already thinks in process. An idea gets screened, the screen feeds diligence, diligence becomes a memo, the memo goes to committee, and the position gets monitored until exit. The habit is already there. It has simply not been pointed at the question of which steps a machine should own.
Pick one process the team runs every week. Write the steps down on a single page. Mark each one as software, model, or human, and note next to each step roughly how long it currently takes and who touches it.
The exercise takes an afternoon. Once that map exists, most of the decisions that follow, including which workflow to automate first, what to ask a vendor, and what compliance needs to look at, become much easier to make.
Pascal AI Labs builds the context graph and agent infrastructure that investment teams run these workflows on, deployed inside your own cloud environment, VPC, or data room depending on where your constraints sit. We work with firms managing several hundred billion dollars in assets, and most engagements start exactly where this piece ends, with one workflow written down on a single page. Book a demo
Bring a process your team runs every week and we will map it with you, step by step, and show you what a production deployment of it looks like.
For a single, well-scoped workflow, something like eight to sixteen weeks from scoping to a team using it every week is a reasonable expectation, assuming compliance is involved from the start. The timeline stretches mainly for architectural reasons rather than model reasons. Firms that settle data residency, hosting, and permissioning during scoping tend to land at the shorter end, while firms that leave compliance approval until the end often add three to six months of rework.
The common causes are organisational rather than technical. Compliance arrives late and the architecture needs rebuilding to satisfy data residency and audit requirements. No baseline was captured, so nobody can demonstrate improvement. The pilot was scoped as a general capability rather than a specific workflow, leaving no clear definition of success. Or the sponsor sat outside the team doing the work, so nothing was ever depended on day to day.
That decision is usually driven by where inference is allowed to happen rather than by benchmark performance. Closed API models offer frontier capability with no infrastructure to manage, but prompts and documents leave the firm's environment during the model call. Open-weight models hosted inside the firm's own cloud or data centre keep inference within controlled infrastructure and behave consistently until the firm chooses to replace them. Many firms end up with a hybrid, using frontier models for hard reasoning and self-hosted open-weight models for high-volume extraction and classification.
In a workflow, the sequence of steps is fixed in advance by whoever built the system. In an agent, the model decides the sequence itself, choosing what to do next in pursuit of a goal. Agents also act on the world through tool calling, editing files or sending drafts onward, which makes their mistakes harder to undo. Most production research systems today are workflows with agentic steps inside them rather than fully autonomous agents.
Usually not as a first move. A senior research sponsor who owns the workflow, one technical builder, and one person from data or infrastructure tends to make faster progress, because every decision stays tied to real work. Dedicated AI leadership makes more sense once several workflows are running and the coordination burden is real. Firms also tend to underestimate the talent already inside the building, particularly analysts who have maintained sector models in code for years.
Contact
USA
228 Park Ave S PMB 55152, New York
UK
86-90 Paul Street, EC2A 4NE, London
India
Indiqube Edge, ORR, Bangalore
Copyright © 2026 | All Rights Reserved
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.