Morgan Trowland works at a small architecture and construction startup, a team of six or seven people running the whole operation. The firm had always kept its statutory accounts and tax returns in order, reconciling the books at least once a quarter. The catch was the time it took. Getting to an accurate financial picture meant a heavy, recurring effort of receipt entry and reconciliation, so the firm only saw its real numbers a few times a year rather than continuously.
That is the gap Morgan set out to close. The books always got done. What he wanted back was the time they swallowed, and a way for the firm to see where it actually stood on any given morning instead of a few times a year.
He is not a software person. He says so himself. But over one month of experimenting, he turned the most time-consuming, judgment-heavy job in the back office into something that runs before he has finished his coffee.
Quick overview
- Morgan Trowland works at a small architecture and construction startup of six or seven people, where the books were always kept current but the routine accounting ate up significant time.
- He built an AI assistant he named Douglas to handle the routine accounting: entering receipts, reconciling them against bank payments, keeping the books current.
- He found Vellum after a week of hand-coding his own bots, when a Claude web search finally surfaced a tool that could do what he had been struggling to build.
- Douglas now runs the bookkeeping every morning for well under fifty cents a session, often twenty to thirty, giving the firm a clear daily financial picture while cutting the time spent on routine accounting.
- Morgan chose Vellum because it is open source, which means the firm can self-host it, own everything, and avoid being locked into a corporate platform.
A job no software fully solved
Bookkeeping for a small firm is deceptively hard. The work itself is dull, but it resists automation because so much of it is judgment. "Taking in all the receipts and making sure they're entered properly in the system, and then trying to reconcile them against bank payments, all this boring, very particular stuff," is how Morgan described it.
He had tried the obvious route. "Even with all the package products, SaaS products on the market, there's still just loads of judgment manual work that has to be done." The software handled the easy parts and left the messy middle for a human, and for a firm with no full-time finance person, that middle is exactly where things fell behind.
Morgan wanted the daily picture without the heavy recurring time it used to take. He wanted to prove to himself and colleagues that the routine work could be handled unquestionably well, every morning, so the firm could see its own numbers in real time. The solution was a Vellum assistant.
Morgan chose Vellum
Morgan did not start with bookkeeping. He started with badminton.
He and a group of friends book courts that get released at ten o'clock at night on the dot, a scarce thing that a lot of people are racing to grab. So he spent about a week hand-coding his own bots, wiring in a WhatsApp connection, fighting with Cursor and Claude Code. It was painful. "I'm not a software person," he said, "but it's so difficult to get Claude Code and Cursor to make a thing that would interact with websites."
He went looking for something better. After a lot of searching, a Claude web scout finally surfaced Vellum. He tried it for half an hour. "I tried that for half an hour and it was just smashing it," he said. The web-page interaction that had been so painful to hand-code, Vellum just did. "It seemed very robust. It just went and found stuff and got halfway along my badminton booking task process within like twenty minutes."
That was enough to make him think about work. He weighed his options, considered Agent Zero and other open-source stacks, and landed on Vellum for two reasons that mattered more than the demo.
Morgan found his solution in Vellum:
- A skills system that let the assistant write new skills live and apply them straight away, which he called "a secret sauce" and "incredibly useful."
- Reliable web-page interaction that worked through finicky HTML without getting stuck, the exact thing his hand-coded bots kept failing at.
- An open-source foundation he could self-host and fully own, which for his firm was a hard requirement.
That last one is worth sitting with. For Morgan, the deciding factor was sovereignty over both data and capability. "It's sovereignty of not just data, but capability," he said. "We really don't want to be further enmeshed into corporate tech things with the lock-in." As more and more capability gets tied to these assistants, he did not want his firm dependent on a private company that could one day be sold to something larger and less benign. "Your one being an open source thing is very critical. Wouldn't have looked at it otherwise."
Meet his assistant Douglas
Morgan named his assistant Douglas, and Douglas now has a morning routine.
Here is the shape of it. Each morning Douglas does the bookkeeping: it takes in the receipts, makes sure they are entered properly in the system, then reconciles them against the bank payments. Morgan reviews what it did, catches the spots where it did not quite think something through, and corrects it. "I'm looking at what it's done and saying, you didn't quite do that," he said. The corrections are training. Over a couple of weeks the routine has tightened to where Douglas handles the bulk of the work and Morgan handles the judgment calls on top.
The cost of a morning run is small. Morgan has been tracking it by hand, and a routine run now comes in under fifty cents, often twenty to thirty, depending on how many bills and invoices came through that day. The goal he is driving toward is simple and, for a company his size, genuinely valuable: a clear financial picture every single morning. "If I can get it to a point where we can get an accurate financial picture every morning, that would be phenomenal."
There was a more dramatic first act, too. In what Morgan called his "exciting phase," he set the assistant loose on a massive accounting backlog, six months of work the firm had not dealt with, and let it run on auto top-up without watching the spend too closely. It cost more than he planned. It also cleared the entire backlog. "It's fantastic," he said.
A daily picture they never had
The reclaimed time is not the whole story. The bigger shift is visibility.
A firm that used to see its real numbers a few times a year now has a path to seeing them every morning, off a job that used to take significant time or money to do well. That is the kind of operational hygiene that small companies rarely get to do daily, and Douglas is making it close to free.
Morgan's roadmap from here is partly about cost and partly about conviction. His firm runs on a principle: use open-source technology and migrate to it wherever they can, because that is how you avoid getting locked into a single corporate platform. Vellum fits that principle exactly, since the firm can self-host it, run it through an ethically owned provider, or export everything and own it outright. "We could self-host it and run it ourselves," he said, "or have it run by whatever ethically owned and operated service provider we might choose." Bookkeeping was the wedge: the unquestionably valuable task that proves the case to his colleagues. What comes after it is the rest of the back office, on infrastructure the firm controls.
How to build this yourself
You do not need to be a software person to set up a version of this. Morgan was not. Here is the pattern, with his stack as the worked example.
- Pick the routine task that resists off-the-shelf software. Lead with the job that is dull, repetitive, and full of small judgment calls that SaaS products leave for a human. For Morgan it was bookkeeping: receipt entry and bank reconciliation.
- Give your assistant the context it needs. Point it at where the receipts live and the system where entries belong, the same way you would brief a new hire on day one.
- Let it write its own skills. Rather than hand-coding each step, let the assistant build the skill for the task and apply it live. This was the capability that sold Morgan over the alternatives he considered.
- Run it once, then review every line. Let the assistant do a full pass, then check its work and correct what it got wrong. Each correction is training for the next run.
- Turn it into a daily routine. Once the output is reliable, schedule it. Morgan's runs every morning and costs under fifty cents a session, often twenty to thirty.
- Track the spend and set a ceiling. Watch what each run costs and set your own top-up limits so the bill stays predictable.
Done this way, it is less a software project and more a workflow you assemble in an afternoon.
Main takeaway
The pattern travels well beyond construction accounting. Any small team has a job like Morgan's: necessary, repetitive, judgment-heavy, and chronically behind because no one has time and the off-the-shelf tools only get you halfway. It is one more example of what a personal AI assistant can actually do once you point it at the right job. An AI assistant that learns the task, runs it on a schedule, and improves with correction turns that job from a quarterly scramble into a daily habit.
For Morgan, the payoff is a firm that can finally see its own numbers every morning, on software it owns and controls. The quiet promise underneath is bigger than bookkeeping: the back office of a small company, handled, without giving up sovereignty to do it.
Morgan found this in Vellum. If a routine job is running behind in your business, hatch your own assistant at vellum.ai.
Frequently asked questions
What is AI bookkeeping?
AI bookkeeping is using an AI assistant to handle routine accounting work such as entering receipts, categorizing transactions, and reconciling them against bank payments. Instead of a person doing the repetitive data entry, an assistant runs the task on a schedule and a human reviews the output. The judgment-heavy parts that traditional bookkeeping software leaves unfinished are exactly what a capable assistant can take on.
Can AI really do bookkeeping for a small business?
Yes, for the routine and reconciliation work. Morgan Trowland's firm uses an assistant to enter receipts and reconcile them against bank payments every morning, work that used to be done in larger batches a few times a year and took significant time. The assistant handles the bulk of it and a person reviews and corrects, which also trains it to do better next time. It does not replace an accountant for complex tax or advisory work, but it keeps the day-to-day books current.
How much does it cost to run an AI bookkeeping assistant?
It depends on how much work you ask it to do. Morgan tracked his daily bookkeeping runs by hand and found a routine morning came in under fifty cents, often twenty to thirty, varying with how many bills and invoices arrived that day. Costs rise with the volume and complexity of the tasks, which is why setting your own spending limits matters. Most assistant platforms let you cap spend so the monthly bill stays predictable.
How is an AI assistant different from accounting software like QuickBooks?
Accounting software automates the structured parts of bookkeeping and leaves the judgment calls to you. An AI assistant, like Vellum, can take on that messy middle: reading receipts, deciding where ambiguous entries belong, and reconciling them against the bank. It works alongside your existing tools rather than replacing them, filling the gap that SaaS products leave open.
Do I need to know how to code to set this up?
No. Morgan describes himself as "not a software person." The shift that made it work for him was an assistant that writes its own skills and applies them live, so he did not have to hand-code each step the way he had been struggling to with Cursor. You describe the task in plain language and review the results.
What does it mean for an AI assistant to be open source?
An open-source assistant is one whose code is open and that you can run yourself rather than only through a vendor's servers. In practice it means you can self-host it, export your setup, and own everything end to end. For Morgan this was the deciding factor: his firm wanted to avoid being locked into a single corporate platform, and open source was the only way to guarantee that.
Can I self-host an AI bookkeeping assistant?
Yes, with an open-source assistant. Morgan's firm specifically chose a tool it could self-host and run on its own machines, or have run by an ethically owned service provider it trusts. Self-hosting keeps sensitive financial data and the assistant's capability under the firm's own control rather than a third party's.
How do I keep an AI assistant's spending under control?
Set top-up limits. Most assistant platforms let you cap how much an assistant can spend, and when it reaches the limit it simply stops until you raise it. Morgan learned this the practical way: during an early backlog cleanup he left auto top-up on and spent more than planned, so now he tracks each run and works within a ceiling.
What kind of tasks can a small business AI assistant handle besides bookkeeping?
A general-purpose assistant can take on most repetitive, rules-based work: scheduling and bookings, web research, data entry, monitoring, and routine correspondence. Morgan first built his to grab scarce badminton court bookings the moment they opened, then generalized it to friends' gym and swimming reservations before pointing it at accounting at work. The bookkeeping job was the wedge, not the ceiling.
How long does it take to set up an AI bookkeeping routine?
The first useful result can come in an afternoon, but reliability builds over a couple of weeks. Morgan had the assistant doing real bookkeeping quickly, then spent the following weeks reviewing its output each morning and correcting it until the routine was solid. Treat the first runs as training rather than finished automation.
Is it safe to give an AI assistant access to financial data?
It depends on the deployment. With an open-source assistant you can self-host, which keeps the data on your own infrastructure rather than a vendor's. That was central to Morgan's decision: his firm wanted sovereignty over both its data and the capability handling it. Reviewing the assistant's work each run, as Morgan does, adds a second layer of control.
Extra resources
How a Solo Shopify Owner Automated Customer Support With an AI Assistant →
The Best Personal AI Assistants in 2026 →
The Best Private Personal AI Assistants →
