Fourteen agents. Seven departments. Zero files that ever left the building.
A year ago, B:Side Capital — an SBA 504 and CDFI lender — asked us to wire what they already had into something that could be asked questions: on a machine they own, behind a filter they control. MARCUS went live in waves through 2026 and passed its year-end go/no-go review. These are the measured numbers from the first 90 days of full-fleet operation.
- Client
- B:Side Capital
- Window
- First 90 days of full-fleet operation
- Published
- Verified by
- Audit log
Four numbers the board asked for. All four came off the log.
Staff hours of preparation returned in 90 days — a ~5,000-hour annual run rate, roughly 2.4 full-time people of pure prep work.
Of staff using MARCUS weekly by week six — with no new app, no new login, and no training deck.
Borrower identifiers sent to any outside model. Every prompt and document cleared the privacy filter first. No exceptions, because no bypass exists.
Of consequential actions approved by a person before anything sent, filed, posted, or paid. Zero unapproved actions — verified by the audit log, not asserted.
Measurement window: the first 90 days of full-fleet operation. Every figure on this page is reconstructable from the tamper-evident audit log — that is the point of having one.
We didn’t sell them software. We wired what they had into something that can be asked questions.
Seven pieces, built in order, each proven on real files before the next one shipped.
B:Side already had the data, the front door, and the books — process documents by the hundred, a team living in Slack, systems of record for loans and accounting. What they didn’t have was a way to ask any of it a question, or a way to hand it the preparation work, without borrower data leaving a controlled environment.
So we built the seven things below, in order, each one proven on real files before the next shipped. The highest-exposure work — funds movement, distressed assets — was deliberately built last, behind the strictest gates, once trust was established. Here is what each piece did.
Every other AI vendor rents you access to a machine in someone else’s building. We put the machine in theirs.
One box, in a locked room, with B:Side’s name on the invoice.
Specified, procured, installed, and configured to run open-source models locally — bought in B:Side’s name, their property from day one, at cost, with zero markup. An owner who can’t evaluate a model architecture can absolutely evaluate a box in a locked room. It also settled the data question before it was asked.
In November we swapped the local model for a newer one — better answers, same workflows, nobody retrained. Once the machine is yours and the filter is in place, the model is just a part you replace.
Uptime since commissioning in August — one planned maintenance window, zero unplanned outages.
Model calls processed entirely on-premises in 90 days. In normal operation, none left the building.
Hardware markup. Bought at cost, in the client's name. That we do not mark it up is the point.
Mid-flight model upgrade, completed in an evening, with zero workflow changes for staff.
How it’s sold: hardware at cost in the client’s name · fixed configuration fee by workload tier · monthly care for patching, updates, and the replacement path.
Their people were already pasting customer data into a chatbot. This is how that stopped being a problem.
Every prompt, every document, scrubbed before any model reads it.
The filter sits between every person and every model. Prompts and attachments are scanned, personal identifiers stripped, and only then forwarded — to the local model or a commercial one, B:Side’s choice per task. Files get read, scrubbed, and passed through as clean text. There is no route around it, which is why the compliance answer is a design fact rather than a policy hope.
We don’t take the filter’s accuracy on faith. A seeded set of documents carrying known, planted identifiers runs against it every month — and every miss becomes a new test case.
Once it was in place, “model of your choice” became a feature instead of a risk: routine work routes to the local model, heavier reasoning to a commercial one, on scrubbed text only.
Of prompts and attachments cleared the filter before reaching any model. No bypass path exists in the architecture.
Personal identifiers detected and stripped — names, SSNs, account numbers — across ~38,000 documents processed.
Detection recall on the monthly seeded red-team test — measured, not claimed, and improving each cycle.
PII incidents since the filter went live. The number the board actually asks about.
How it’s sold: fixed install + monthly. It closes like a security purchase, not a transformation project — and it sells standalone to companies with no interest in agents at all.
The company remembers more than any one person in it. This is where that memory went so it could be asked questions.
Everything B:Side knows, in one place, in a form the machine can use.
Structured loan records on one side; process documents, correspondence, and policies on the other — joined so a single question can cross both. A lender has loans and findings the way a contractor has jobs and change orders: same shape, different nouns.
We didn’t design their memory from scratch. B:Side answered the knowledge-base questionnaire — core entity, event stream, document sources, access rules — in one working session, and that answer sheet generated the tables, indexes, ingestion pipeline, and read tools. It is what turned a six-week build into two.
From completed questionnaire to a queryable knowledge base — generated from the intake, not hand-designed.
Process documents ingested, joined with thousands of loan records and a decade of institutional correspondence.
Questions asked by staff by month three — the questions that used to wait for whoever just knows.
Of answers delivered with a citation to the source document or record. The rest say they do not know, and escalate to a person.
How it’s sold: fixed build fee, tiered by source-system and entity count. The anchor of the line.
The best interface is the one their people were already staring at.
No new app. No new login. The first question came eleven minutes after go-live.
MARCUS lives inside B:Side’s Slack. Work arrives as messages; approvals happen where the conversation already is. Every owner who has been burned by software was burned by adoption, not capability — so we killed the adoption objection before it could be raised.
The entire rollout, per division, was a twenty-minute introduction. There was no training deck, because there was nothing new to learn.
From go-live to the first real staff question — no training session had happened yet. None was needed.
Weekly active adoption by week six, across all seven divisions — including the skeptics.
Of all MARCUS interactions happen in Slack — finished work delivered to the channel where the team already works.
How it’s sold: fixed and thin, bundled with the knowledge base by default. Mostly configuration — its job is to make the knowledge base feel real in week one. It did.
A chat tool for the whole company — except the company owns it, and nothing leaves the building.
A branded workspace for everyone who just wants to ask something.
Slack is for teams who want finished work delivered to them. The chat workspace is for the other moments — drafting, summarizing, looking something up — connected to the knowledge base, running on B:Side’s server, behind B:Side’s filter. Staff opened it the way they would open any chat tool, because that is exactly what it looks like.
And the pattern we predicted happened on schedule: in week two, someone asked it about last year’s loans. Before the knowledge base was connected, it couldn’t answer. That single unanswered question did more to sell the knowledge base internally than any pitch we could have made.
Conversations per week by month three — policy lookups, document summaries, and first drafts leading the mix.
Of chat traffic routed through the privacy filter — the paste-into-a-chatbot habit, made safe instead of banned.
When the first question about loan history arrived — the knowledge base selling itself, in the wild.
How it’s sold: fixed setup + monthly care. Paired with the privacy filter, it is the most natural first purchase in the line.
The books already knew the answer. This is what it took to ask.
Reconciliation prepared. Invoices staged. Receivables chased. Nothing posted until a person approves.
The accounting connector let MARCUS see the books: month-end reconciliation arrives prepared with exceptions flagged, vendor invoices arrive pre-coded with the source attached, aging receivables get follow-ups drafted the day they cross the threshold. A person approves every entry before anything posts — that is not a limitation, it is the design.
The connector code is a durable internal library; the deployment is the billable event. Build it once, deploy it against every client.
Bar lengths are proportional to time. “After” figures are human review time — the preparation itself happens before anyone sits down.
Vendor invoices staged in 90 days, each pre-coded with source attached — every one approved by a person before posting.
Reconciliation mismatches surfaced automatically before month-end — found by the system, decided by a person.
Improvement in average days-to-collect on fee invoices, from follow-ups drafted the day an invoice ages past threshold.
How it’s sold: fixed price per connector, from a published catalog — ordering parts, not commissioning software. See the build catalog.
Anyone can demo an AI. We show ours failing, on purpose, on a schedule.
Every break gets found by the tests, not by their staff.
Every night, a stress harness runs against B:Side’s live configuration — eligibility edge cases, malformed documents, planted identifiers, adversarial prompts — and posts the results to a dashboard B:Side’s own team reads. When something breaks, the harness finds it before a staff member does. That is not a promise; it is a count we can show you.
Behind every action sits the audit trail: an append-only, hash-chained log. When a question comes up about what MARCUS did and why, the answer is a lookup, not an investigation.
Automated assertions run nightly against the production configuration — plus a monthly deliberate failure drill.
Regressions caught by the harness before any staff member encountered them. Zero found by staff first.
Entries on the tamper-evident audit log — every action traceable to its trigger, its approver, and its source document.
Of the harness is generic assertion-running; the remainder is client-specific fixtures. Which means it deploys anywhere — including on AI systems somebody else built.
How it’s sold: bundled with every build, never optional — it is the differentiation. Plus a standalone fixed-fee assessment for systems built elsewhere: the only item in the line sellable to someone who is not a client yet.
MARCUS prepares the work. People decide.
Nothing sends, files, posts, or pays until a person approves it. While a request waits, nothing moves.
MARCUS never decides eligibility, credit, or price — those numbers come from systems, never from memory. Across the measurement window, every consequential action passed through a human approval: approve, send back, or escalate.
Escalation is automatic, including for anything that has waited too long. Silence never means yes.
The work moves forward — approved on first presentation, most within the hour.
Returned with comments. Every send-back becomes training signal and, where it generalizes, a new test case.
Handed to the right person, plus every request that waited too long. Escalation is automatic; silence never means yes.
You already have the data, the front door, and the books. We wire them into something you can ask — on a machine you own, behind a filter you control, with your people approving every consequential action. Every path here starts with the same free 30-minute assessment, and we reply within 24 hours.
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: 100% of your audit fee credits toward a sprint signed within 60 days, up to 25% of the sprint price. See the price list →
If a scoped workflow is not live in your operation within 90 days, we keep building at no charge until it is. And 100% of your audit fee credits toward a sprint signed within 60 days, up to 25% of the sprint price.
A fixed price, quoted in writing before we start, and a straight answer within 24 hours.
— Christopher Myers, Founder