Your Outsourced AI Department | Affine Works

Your Outsourced AI Department

Why financial service firms need one, what it does, and how Affine Works runs it.

Today, every chat searches and re-reads the firm's raw systems on its own, so the work repeats and answers drift. With Affine Works, documents are read once into a shared, sourced foundation that every tool and agent draws on. Figures are illustrative.

Most financial service firms already pay for ChatGPT or Claude. Seats are available to the firm's employees, and the employees are individually tasked with finding ways to use the technology. These tools are very empowering out of the box: producing first drafts in seconds, boiling down a long document, expanding web-based research capabilities, and much more. We use these tools every day in a multitude of ways and know how powerful they are when leveraged correctly.

What most firms miss is that AI's value only compounds when someone owns it for the whole firm. When responsibility for getting value out of these tools is decentralized and left to each employee, that is roughly where it stays. The people who put in the hours get a lot out of the products, but everyone else gets a faster search engine. Questions about the firm itself can be answered, but only by someone willing to find the right files, move them into the context, and explain what the firm means by its own terms. The next person has to do it all over again. Nobody's job is to make that easier for everyone else, so it never gets easier. A few months into the rollout, most firms find that AI has made a few people faster and left the firm's aggregate efficiency only marginally better. It's often hard to find the direct ROI in rolling out these general tools when a plan, infrastructure, and accountability for utilization are not in place.

Projects and connectors

People who are fluent with ChatGPT or Claude usually get the most out of them by creating a Project. A Project is a workspace that holds the files, notes, and instructions for one client, deal, or workflow, so the tool has the right context every time the user revisits or continues working. It works, but it is not a scalable way to manage context, and learnings across users are independent and do not compound. Someone must create each Project by hand, keep it current, and rebuild it when things change. It tends to reflect that person's view of which files matter and what they mean. A firm where dozens of people each keep their own Projects ends up with dozens of slightly different versions of the truth, each a little more out of date or biased than its owner realizes.

Connectors are the other common step. They let the tool reach directly into your email, shared drive, and other systems, and they are great for quick and low-stakes lookups. Ask where a document lives or what a client said in last week's email, and a connector will usually find it and analyze in seconds to minutes. Bigger and broader questions, especially involving high-stakes repeatable work, are where complications begin. With a connector, all searching, file parsing, record matching, and data joins happen at runtime in the moments directly after someone asks. The model must find the right files, read them, match the same client across systems that call it different things, and decide which system to trust when they conflict. That process is probabilistic, so it can pull different documents into context on different runs, and the same question can come back with a different answer.

That work is also slow and costly, and financial documents make it especially hard, because the numbers that matter sit in PDF tables, scanned statements, and spreadsheet formulas. With a connector, it is thrown away as soon as the answer is delivered, so the next question pays for it again, even though the facts in a document only need to be worked out once.

The bigger limit is what each workflow can see. A connector works within the permissions of whoever is using it, and that is the right control, but it means every chatbot workflow is built around what one person can pull into one conversation. The alternative is to design workflows at the organizational level, each working from role-based context built specifically for the question at hand, so it has everything that question needs and nothing it doesn't.

None of this makes connectors a bad idea, and your team will likely reach our work through a version of one. What matters is what sits on the other end: raw systems that have to be searched and read from scratch, or a copy of the firm's information that has already been gathered, checked, permissioned, and organized. We'll go deeper on this in a separate post.

Doing the work once

Affine Works does that work once for the whole firm, before a question ever reaches an AI model. We connect your systems, learn how your business runs, and engineer the data so it is gathered, related, and kept current without anyone moving files around. In practice, this moves the work from the moment someone asks a question to the moment the information arrives. When a new statement lands in an inbox or a spreadsheet changes on the shared drive, the system reads it once and works out which client, company, or entity it belongs to. It pulls out what matters, checks it, and files it alongside everything else the firm knows about that relationship. By the time anyone asks, the reading is done. The result works like a ChatGPT or Claude Project for nearly any question you could ask about your business. Think of it as every question you could have about your firm already having the correct context assembled, Project-style, before it is asked. Your team can use it through our portal or plug the ChatGPT or Claude they already know straight into it.

Trust and auditability

Finance has always held its work to a few plain rules: cite your sources, never invent a figure, and flag anything that doesn't reconcile. We hold the system to the same rules, and most of the effort goes into making sure it works from the right information. In a financial firm, that phrase is more specific than it sounds. The right information is current, so the answer comes from the latest statement and not from the file someone named "Final" two versions ago. It is consistent with how the firm talks about itself. Words like "active," "committed," and "overdue" mean something particular at every firm, so we define them once, with the people who use them, and write them into code so they mean the same thing in every answer. It is connected, so a client, a company, or a legal entity resolves to one place no matter what each system calls it, and the "Bill Smith" in one system and the "William F. Smith III" on a signed agreement end up as the same person.

It also has to show where it came from. Every figure and fact carries a link back to the source document, page, CRM record, or spreadsheet cell it came from, along with the date it was taken, so any answer can be traced to its source. When two documents or sources disagree, the system shows both and names the conflict instead of choosing one itself. A contradiction between a contract and an insurance certificate is often the most useful thing a firm learns that week. When a source changes, anything built on it is refreshed or marked out of date. And numbers are checked before they reach a report, so a figure that fails a check points straight back to the cell that caused it instead of turning up for the first time in front of a client.

This matters more in finance than almost anywhere else. Auditors, regulators, investment committees, and clients all eventually ask the same question: where did that number come from? An answer you can't trace to its source is an opinion, and a firm can't put its name on an opinion.

What sits on top

That foundation is where the rest of the work starts. With your team, we build agents to take on the work that repeats. An agent is an AI model given a job, clear instructions, and access to only the data and tools that job requires. Some run on a schedule, some run when something happens, such as a document arriving or a deadline approaching, and some run when someone asks. They deliver their work where people already are, in email or a spreadsheet, and anything that leaves the firm waits for a person to approve it. Our portal shows every run: what the agent did, what it found, which sources it used, and what is waiting for approval.

We watch how the agents and the data are used, then fix, tune, and expand them over time. Every run is recorded, along with how long it took and what it cost. When an answer is wrong, we trace it to the step that caused it and fix the cause. When a better model is released, we test it against the firm's real work and switch only if it helps. Leadership gets a clear view of where the tools are helping and where they aren't shown as patterns across the firm.

Working with your people

A partner from Affine Works meets with your leadership regularly, trains your people, and can even take over the administration of your ChatGPT or Claude accounts, so keeping up with AI stops being one more thing on everyone's plate. That relationship is also how the work grows. The sessions with your team are where we learn what people still do by hand every week, and those chores become the next things we build. Over time it feeds itself: better data makes the tools more useful, people use them more, and what they still do by hand tells us what to build next.

Whose job it is

The problem we started with was that making AI work across the firm was nobody's job. Done properly, it is several people's jobs. Someone engineers the data, someone builds and maintains the agents, someone watches quality and cost, and someone teaches and supports the people using it all. Most financial services firms can't justify the payroll and admin overhead for even one of those roles, let alone all four. An outsourced AI department makes it someone's job anyway. That is what Affine Works is. We want to be the ones financial services firms trust to be accountable for ensuring their business leverages this amazing technology. We built it after years working inside financial firms and building real AI systems for them, so we know the work and what it takes to automate it well.

See how we work ← Back to the blog

Ready to Make AI Work for Your Business?

Bring one recurring workflow. On the intro call, we will map what it takes to build it, train your team on it and run it.

Book an intro call