AI agents that prepare a lender’s back-office work. People own the decisions.
B:Side Capital, an SBA 504 and CDFI lender led by our founder Christopher Myers, needed to reduce repetitive preparation without giving up control of data or decisions. We built MARCUS around its process documents and existing tools: local document processing, filtered model access, and human approval of consequential actions.
Private by design. Auditable by default. Built from their own process. Published with permission.
- Client
- B:Side Capital
- Sector
- SBA 504 / CDFI lender
- Footprint
- 7 departments
- The team
- 14 AI agents
- Runs
- Local + hybrid
- Updated
What changed for the team.
Estimated staff preparation hours returned during June–August 2026, calculated from B:Side’s initial workflow studies. This is operating capacity, not measured payroll savings.
Reported weekly use by week six, across a staff of 45 employees. The percentage is rounded; an exact active-user count is not published.
Borrower identifiers reported as sent to outside models during the measurement window. Filtering is required; detection is not guaranteed to be perfect.
Of consequential actions approved by a person during the reported window, before sending, filing, posting, or paying.
Reporting period: June–August 2026. B:Side reports these results for the stated period. Preparation hours are estimates calculated from its initial workflow studies; operational activity is reported from its audit log. Returned capacity is not measured payroll savings. Annualized equivalents are not measured annual results.
- Type
- SBA 504 · CDFI
- Touches
- Borrower financials
- Stakes
- Regulated · auditable
B:Side Capital is a mission-driven lender, an SBA 504 and CDFI institution whose work touches sensitive borrower financials at every step: intake, eligibility, underwriting support, funding, portfolio monitoring, and the high-stakes work of servicing distressed loans. It’s a business where a single mishandled document isn’t an inconvenience; it’s a compliance event.
Sensitive documents, established processes, and accountable decisions shaped the build from the start.
Lending runs on preparation. That’s where the days disappear.
Before anyone can make a decision, someone has to gather the documents, pull the reports, run the checks, and draft the paperwork. The same steps, over and over, across seven departments. That preparation is where the bottlenecks form.
The team needed a defined boundary around borrower documents and a reviewable path from source records to prepared work. Eligibility, credit, and pricing remained with the people responsible for them. The system had to show its sources and record its actions.
The brief, in one line
Take the prep off our people’s plate, without ever taking them out of control.
Meet MARCUS.
MARCUS is not an app B:Side’s team had to learn. It’s a colleague that lives inside the tools they already use: Slack, their lending systems, the browser. It does the gathering, checking, and drafting, then hands finished, organized work to a person to approve.
B:Side’s process documentation defined the workflows and the sources each agent should use. That made the work reviewable against the institution’s own operating rules.
Four surfaces the team already knows
A chat front door
Ask a question, get findings back. The way the team already talks to each other.
In SlackAn approval queue
A simple list where a person says yes, sends it back, or escalates.
Human sign-offA browser assistant
Read-only. Answers questions about the screen in front of you, changes nothing.
On-screen, view-onlyThe systems of record
Once approved, work lands where it already belongs: their lending platform and tools.
Existing systems- AI agents
- 14
- one for each step of the lending lifecycle
- departments
- 7
- from intake to portfolio to distressed-asset servicing
- source documents
- ~840
- every action traced to a real B:Side process
We didn’t bring a template. We read everything.
Before we wrote a line of code, we reviewed roughly 840 files across every division of the business. From that, we mapped fourteen distinct agents, each one doing a specific, real job, each one citing the exact operations document it came from.
Every step in the blueprint cites the operations document it came from. That’s the difference between an AI demo and an AI system.
How it was built
We built it the way you’d actually want a system this consequential built: in order, one capability at a time, each piece proven on real files before the next one shipped.
Highest exposure, built last
The most delicate work, payment relief and loan resolution, was deliberately built last, once trust in the system was already established.
Local processing. Defined data boundaries.
MARCUS processes borrower documents on a secured machine in B:Side’s own server closet. A privacy filter scans prompts and documents before model processing. Selected tasks can use a commercial model on filtered text, as described in the results and architecture detail.
Local by default
Borrower-document processing runs on B:Side’s hardware. External model use is a separate routing decision for selected tasks on filtered text; local deployment does not make every downstream service local.
Identifiers filtered before model processing
Microsoft Presidio detects supported identifiers for removal before model processing. Detection is imperfect; seeded tests, routing limits, and review are part of the control.
Encrypted and tamper-evident
Stored documents and work products are encrypted. Actions are recorded in an append-only, hash-chained log so changes to recorded entries can be checked.
Borrower document
Local model
Encrypted store & audit chain
The diagram shows the local processing path. Read the security review for deployment choices, filter limitations, and the questions to ask about your own workflow.
MARCUS prepares the work. People decide.
Consequential sending, filing, posting, and paying require a person’s approval. MARCUS does not decide eligibility, credit, or price. Those numbers come from systems, never from memory. Every request arrives the same way.
It shows its work
Here’s what I did, here’s what I’m asking for, here’s the source I relied on.
A person chooses
Approve it, send it back with comments, or escalate it. While a request waits, nothing moves.
It escalates, never proceeds
If it waits too long, it escalates to a person automatically. It never proceeds on its own.
Numbers come from systems
Eligibility, credit, and price are looked up, never recalled from memory.
Every step is logged
Attributable and auditable, down to the source document.
It’s the design, not a limit
Humans in control isn’t a constraint we worked around. It’s the point.
B:Side’s people get their day back for the work only people can do.
With the preparation handled, the team is free for the relationships, the customer service, and the judgment calls that actually move a deal. The repetitive prep that used to create bottlenecks now arrives finished, organized, and named the way the team already does it, ready for a human to review.
The result is an institution that moves faster without giving up an ounce of control, compliance, or trust.
If you can document how your business works, we can build the machine that runs it.
MARCUS is what Main & Machine builds when the stakes are real: private, auditable, built from your own processes, and designed so your people stay in control the whole way through. Builds in this class are published on the price list as The Full Back Office — from $95,000, four taken per year. If that’s the kind of system your business needs, let’s talk.
The guarantee: If a scoped workflow is not live in your operation within 90 days, we keep building at no charge until it is.
The credit: Your audit fee can be credited toward a sprint signed within 60 days, up to 25% of the sprint price.
More client stories are clearing review as each client signs off on the numbers.
Read next The architecture MARCUS runs on, control by control: how we handle client data. The concept explained end to end, with these figures in context: private AI for small business. The reported scorecard: Results, workflow studies, and operating evidence.