AI for financial services
Patterns, no client yetAdvisory firms, lenders, and finance teams where senior people spend their day gathering what they need to think about.
In a finance business the expensive work is judgment, and most of the day is not judgment. It is pulling documents together, reconciling numbers that should already agree, rebuilding the same memo, and chasing a client for a statement. That gathering is the part worth taking back.
No client yet
Not in this category
Code does math
A model never computes a dollar
No advice
Nothing we build recommends
Where the time goes
We have not built a system for a firm in this category. What we have built is a construction billing engine where code owns every dollar, and a commercial real estate underwriting check that argues back. The patterns below are where those two overlap with your work.
Nothing starts until the statements, the returns, and the schedules arrive. Getting them means asking a client repeatedly for things they said they sent. It is not analysis, it is not billable in any satisfying way, and it is most of the elapsed time on a file.
The file waits on a chase
Two systems hold the same figure and report it differently. Finding out which one is right takes a person tracing it back through both. It happens every cycle, on the same accounts, and it is the same work every time.
The same reconciliation, every cycle
The thinking took an hour. The document took four: pulling the figures, formatting the exhibits, and writing the sections that are the same on every memo you have ever produced. The senior person did both, at the same rate.
Four hours of assembly per hour of thinking
An assumption gets set early, everything downstream inherits it, and nobody goes back to check whether the source documents support it. The person who would catch it is the same person who set it, reading their own work.
The assumption inherits, unchecked
Who is past due, what was already sent, who is disputing something. If your system's automatic reminders are switched off, it is because they were context-blind and chased somebody who had already paid. So chasing went back to manual, which means inconsistent.
Context-blind reminders get switched off
Client financial information, recordkeeping obligations, and rules about what may be communicated and how. Every idea above touches at least one of those, which is a real constraint on the design rather than a paragraph at the end of it.
The rules decide what is buildable
The machine
The operational surfaces around the analysis. The analysis itself is yours.
01
02
03
04
The split
Two hard rules here, and they come from work we actually did. First: every dollar is computed by code, never by a model. We learned this building a construction billing engine where a wrong number reaches a general contractor, and the stakes in your category are not lower. Second: nothing we build gives advice. It gathers, reconciles, drafts, and flags. A person recommends, and a person signs.
What we build
01
The outstanding list, tracked, with follow-up on a schedule. It is the least interesting system on this page and it usually removes the most elapsed time, because the file was not slow, it was waiting.
02
Reading the documents into the shape your process needs, rather than a person keying them and a second person checking the keying. The model reads. Code computes. Those are different jobs and they do not get mixed.
03
Not a system that reconciles for you. One that shows you only the places two sources disagree, with both figures and where each came from. The exceptions were always the work. Everything else was finding them.
04
The standard sections and the exhibits drafted from figures that code computed, so the senior person edits the thinking rather than assembling the document. The hour of judgment stays expensive. The four hours around it do not have to.
05
It reads the model against the source documents and reports which assumptions are not supported, citing the line it disagreed from. We built this shape for commercial real estate underwriting. A checker that agrees with you is a mirror, and a mirror is not a control.
06
The past-due list, ranked, knowing what was already sent, who replied, and what is disputed. This is the same system we build for contractors, and the reason is identical: your platform's built-in reminders are context-blind, which is why somebody turned them off.
Under the hood
Accounting and ledger
Client and pipeline
Documents
Your ledger is the system of record and nothing about that changes. As with every regulated category, the first question is what client financial information may move and where, and that answer constrains the design rather than getting appended to it. We would rather find out a workflow is not buildable in the roadmap than after you paid for it.
Principles
This is the oldest rule in our build doctrine and it came out of a construction billing engine where a wrong figure lands in front of a general contractor. A model reading a messy statement into structured data is a good use of a model. A model adding numbers up is a bad one, and no amount of accuracy in testing makes it a good one.
It gathers, reconciles, drafts, and flags what looks unsupported. It does not recommend, and a person signs everything. If you want a vendor who will tell you their system can make the call, they exist, and you should think carefully about what they are actually offering you.
A checker that agrees with your model has removed the only reason you built it. We learned this from a commercial real estate broker whose sharpest critique of AI was that it is a yes-man. So the assumption checker disagrees by default and shows the source line, which makes the disagreement something you can act on or dismiss on evidence.
The vocabulary on this page is borrowed honestly from the construction and commercial real estate work, which is where our real numbers live. There are no stat chips here pretending to be results. When there is a financial services system worth pointing at, it will be on the work page with a name attached.
Same approach, different language
Commercial real estate
Commercial brokers working multifamily, retail, and hospitality, whose day still runs on a cold-call list.
We speak the language →Insurance
Medicare brokers, independent agencies, and the licensed agents who live and die by enrollment windows.
Live system →Professional services
Agencies, consultancies, and expert firms where the product is judgment and the constraint is the founder's calendar.
We speak the language →Questions
No. The closest real work we have is a construction billing engine where code computes every dollar and the export reconciles to the cent, and a commercial real estate underwriting check that reads a rent roll against a model. Both of those overlap with your work, and both are described on their own pages with the actual numbers. This page borrows the vocabulary and does not borrow the proof.
No, and this is not a soft boundary. Nothing we would build recommends anything. It gathers documents, reads them into structured data, surfaces where two sources disagree, drafts the standard parts of a memo, and flags assumptions the source documents do not support. Every recommendation is a person's, and a person signs everything that reaches a client.
No, and we design as though it cannot. Every dollar is computed by code, in a tested function, never by a model. The model's job is reading a messy input into a structured one, which it is genuinely good at. Arithmetic is a solved problem and handing it to a probabilistic system is a choice nobody has to make. We learned this on a build where a wrong figure lands in front of a general contractor, and your stakes are not lower.
It constrains the design rather than getting appended to it. What client financial information may move, where it may go, and what has to be retained is your firm's determination and your compliance people's call. Our position is that if a workflow only works by putting client information somewhere it should not go, we will tell you not to build it. We are not going to give you a regulatory opinion in a sales conversation.
A system that reads your model against your source documents and reports which assumptions are unsupported, citing the specific line it disagreed from. We designed this shape for commercial real estate underwriting, where a projected net operating income can quietly diverge from what the rent roll actually supports. It does not value anything and it does not decide anything. It disagrees, with a citation, and you decide whether it is right.
If they are on and working, you do not. In our experience they get switched off, because they are context-blind: they will chase a client who paid yesterday or one who is mid-dispute, and each of those costs a relationship. Collections needs to know what was sent, who replied, and what is contested. That is a judgment problem, which is a different kind of software.
A readiness workshop. One session, one real workflow mapped end to end, and you keep the map and the plan whether or not you build anything with us. In a category where we have no case study, paying us to guess would be the wrong trade for you, and the workshop is designed so you can walk away from it with something useful.
Read next
Explainer
What an AI-legible business actually looks like, the five-point readiness checklist to score your own operation, why one person is usually compensating for every broken handoff, and the 90-day path from messy to legible.
Operations
How to reduce admin overhead without adding staff by redesigning intake, follow-up, and reporting workflows into systems that run without constant manual attention.
Strategy
Why business operators should build portable markdown knowledge bases before relying too heavily on any one AI model, platform, or vendor memory layer.
Where to start
Find the bottleneck, price the fix, build the system, then keep compounding it. The four engagements are how you buy it, and you can start on any rung.
See the four engagements →Get started
Book a 30 minute call. We find the one place your work waits on you, and you get a straight answer about whether AI is the fix. No pitch, no deck.