How I'd build a self-running daily financial brief with AI
Most small business owners don't see their numbers until the accountant calls with bad news. Here's how I'd build an agent that hands them one honest paragraph every morning, and why the right way to build it depends entirely on the client.
Most small business owners don’t read their books. Not because they don’t care, but because reading the books means logging into two or three different systems, squinting at numbers that don’t mean much out of context, and trying to remember what normal looks like. So they don’t, until the accountant calls in March with news that was sitting there in October.
The fix isn’t a dashboard
It’s a short note that lands in their inbox before their first coffee: what changed since yesterday, what looks off, and what (if anything) actually needs their attention today. Something that reads like a quick message from a sharp friend who happens to be good with numbers, not one more screen they have to remember to open.
That’s a genuinely buildable thing right now, an agent that reads the accounts overnight and writes the brief while the owner is asleep. The interesting question isn’t whether it can be built. It’s how I’d build it, because the honest answer is: it depends on who I’m building it for.
Same idea, three different builds
The right stack for this isn’t a matter of taste or trend. It’s a matter of who the client is: how comfortable they are with new tools, what they’re already running, and how specific their situation actually is. My starting position is always the same: meet the owner where they are and hand them the simplest version of this they can run and own entirely on their own. The build only gets bigger when their specific situation genuinely won’t fit anything smaller. Here’s how that plays out for three different owners with the exact same problem.
The owner who’s never touched an AI tool
For someone with no technical background and no appetite for one, the answer isn’t to build them anything. It’s to configure something that already exists. ChatGPT now lets you connect your bank, card, and brokerage accounts directly (OpenAI partnered with Plaid in 2026, covering more than 12,000 institutions, including most of the major banks and brokerages an owner is already using), and it can run recurring tasks on a schedule through its workspace agent features, pulling data, summarizing it, and delivering the result automatically each morning. (OpenAI renamed these third-party integrations from “connectors” to “apps” in late 2025, though you’ll still hear both terms.)
For this client, the engagement is a consultation, not a build. Connect the right accounts, set up the recurring brief, write and tune the prompt until it actually sounds like something worth reading, and sit with the owner until they trust it and know how to adjust it themselves. No code gets written. Nothing needs to be hosted, maintained, or paid for beyond the subscription they probably already have.
The owner with some technical comfort, or a sharp person on staff
Here I’d lean toward Claude and a custom Agent Skill: a small, portable bundle of instructions, scripts, and reference material that Claude loads when it needs to do this specific job. Skills are just folders (a short instruction file plus whatever scripts and templates the task requires), and they work the same way across Claude’s chat interface, its developer tools, and its API. That makes them a good fit for a client who wants something more tailored than an off-the-shelf configuration, without taking on a cloud bill or an infrastructure team to support it.
Anthropic has started shipping official building blocks in this direction too: a Finance plugin that bundles accounting-flavored Skills (reconciliation, income statements, variance analysis) and connects to a business’s ERP, data warehouse, or spreadsheets through Claude’s connector framework, the same mechanism that lets Claude reach out to a live external system rather than work from what’s pasted into the chat. It’s a useful signal of where this is headed, though it’s built for finance teams running monthly close, not a single owner who wants one honest paragraph each morning. For this client, I’d take the same idea and scale it down: a lighter Skill, tuned to their business, wired to the one connector that actually points at where their books live, instead of standing up an institutional-grade suite they’ll never fully use.
The brief here can be more specific to how this business actually operates: which accounts matter most, which numbers the owner actually cares about, what “normal” looks like for their particular business, and how to flag when something drifts from it. It’s still lightweight. It’s still theirs to keep, move, or hand to someone else entirely.
The owner with a genuinely specific, demanding use case
Some businesses need more: stricter controls, deeper integration with existing systems, or a use case specific enough that no off-the-shelf configuration will cover it well. That’s when I’d reach for a custom build on AWS, using Bedrock and AgentCore (Amazon’s managed platform for running agents reliably in production, with proper session isolation and far more headroom than a typical serverless function), on top of the same proven, decade-tested primitives I’d use for any production system: object storage, queues, a real database. Full custom build, full ownership, on infrastructure with a long track record of not falling over.
This is the most expensive path to build and the most expensive to be wrong about, which is exactly why it should be reserved for the businesses whose situation actually calls for it, not treated as the default because it’s the most interesting thing to build.
The one thing that has to be true regardless of the path
Whichever of these three an owner ends up with, the brief is worthless if it isn’t trustworthy, and trust here means one specific thing: when the owner reads a number and asks “wait, where did that come from,” there has to be a real answer. Not a guess, not a plausible-sounding sentence, an actual pointer back to the actual transaction in the actual connected account. An agent that occasionally makes up a number is worse than no agent at all, because it spends down the one thing a financial tool can’t survive without: the owner’s confidence that what it’s telling them is true.
That’s the part of this that takes real judgment, regardless of which stack it’s built on: choosing data sources carefully, writing the prompt so it cites what it’s looking at, and testing it against weeks of real data before it ever lands in someone’s inbox unsupervised.
The point
The interesting decision was never “which AI.” It was always “which build serves this particular owner, with their particular comfort level and their particular business, without asking them to become something they’re not.” Sometimes that’s a configuration conversation that takes an afternoon. Sometimes it’s a custom system that takes months. Whichever it is, the goal is the same: hand them something they can take home, understand, and keep running without me, and reach for more only when their situation truly demands it. Most of the time, the right answer is whichever one the owner will actually still be using, owning, and trusting, a year from now.