Leadership1 distinct publisher3 min readPublished
One owner's account of an AI rollout at a nonprofit SBA lender argues that order of operations, not phrasing, decides adoption. The human-owned list has to exist before the purchase order, or it reads as damage control.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
The two acts are the same whichever order you run them in. You sort the work by judgment, and you buy the software. Run the sort second and the list of things a machine will never touch describes a tool you have already paid for, which is why staff hear it as reassurance about a decision rather than a constraint on one. Run it first and the list screens the shortlist. That is the whole mechanism, and it is why this is a sequencing question rather than a communications question.
Run the survey figure through a headcount before you plan the meeting. At 52% [3], a 50-person all-hands contains roughly 26 people who are already worried when they sit down [14]. The hand that went up at B:Side reflected the room's median, the typical reaction rather than an outlier. Any announcement built for the enthusiast is therefore built for the minority, and the author's efficiency deck died in the room for a reason he watched land on people's faces: told a tool will make everyone more productive, staff hear that the company will need fewer of them [13].
The author says the commitments cost him nothing to say [8], a claim that holds only on the day it is made and stops holding once the promise is tested against real decisions afterward. Once a workforce knows that a person makes every credit decision and that no customer will discuss hardship with a machine [7], the vendor tier that automates underwriting is off the table for as long as the promise stands. The trade is adoption speed now against an option on automation later, and it is worth naming as a trade rather than as a best practice.
The list alone does not prove commitment, since any owner can recite a set of sacred tasks and then quietly erode them. The answer is in the cost of reversal. A list published before procurement constrains what you sign; a list published after it can be revised without anyone noticing, because nothing was ever foreclosed. The author's own framing is the useful one: a machine can hold knowledge but cannot hold responsibility, and a borrower calling because a business is failing wants a person who can own an answer [12]. That is a claim about accountability. Better models do not settle it, because the gap is not one of capability.
This quarter's record supports something modest and cheap. You can do the three-bucket sort, where automate means a mistake is cheap and fixable, assist means the machine drafts and a person decides, and human-owned never moves [6]. You can point the first build at the work nobody will miss, as B:Side did when it killed its planned flagship for document intake [9]. The stronger claim goes further than the record can back up. The author's observation that the fastest adopters have the owner who said out loud what would stay human, not the best software [11], comes from his own client base at Main & Machine [2], unquantified and unaudited.
The pressure over the next several years will not sit on the automate line. It will sit between assist and human-owned, where a good enough draft turns a decider into a signer. Naming the human-owned bucket costs a meeting and one line in a contract; holding it costs whatever the automated tier would have saved, and that bill arrives later than the applause does.
Ranked by verification strength, evidence, and original report placement.
At an all-staff meeting announcing that B:Side Capital was adopting AI, the author arrived with a deck about efficiency and the future of work, and the first hand up ignored it and asked, 'Is this how the layoffs start?'
The author runs B:Side Capital, a nonprofit lender specializing in Small Business Administration loans, and started Main & Machine, a company that builds AI systems for small businesses, describing himself as both the owner buying the technology and the builder shipping it.
Before looking at a single vendor, B:Side sorted its work by judgment rather than by task: the machine never acts alone on a credit decision, never talks to a borrower about hardship, and never commits the company to anything.
The author's sorting method uses three buckets: automate, for anything where a mistake is cheap and fixable; assist, where the machine drafts and a person decides; and human-owned, where nothing ever moves and everyone on the team should know the contents by heart.
The revised announcement led with a plain list of what would not change: a person makes every credit decision, no customer ever discusses hardship with a machine, and nobody is punished for leaning into the new tools while people who learn them get rewarded.
The team killed its first instinct to build an impressive flagship and pointed the machine at document intake instead: the sorting, checking and transcribing that everyone dreaded.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 31, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
leadership
A CDFI lender's CEO briefed his board on Monday using rules written before the crisis1 distinct publisher
product
Anthropic's usage policy says no explicit content. Opus 4.6 said yes 10 times out of 10.1 distinct publisher
invest
BayFirst's $41.5M cleanup shows what a bad lending niche costs to exit1 distinct publisher
invest
Meta's $17 billion settlement puts a cash number on the tech backlash1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One account, from an interested party
The meeting, the three-to-four-hour loan file, the near-total uptake in a quarter, what MARCUS catches — every detail the argument stands on traces to the same man, who built the software and runs the lender that bought it, writing under Entrepreneur's contributor banner. The lone outside data point, Pew's 52%, arrives with no date, sample or link. Nothing has been checked by anyone who wasn't in the room.
One real deployment, owner-scored
This is not vaporware: a named system is working real loan files at a real lender, and the reported uptake — nearly everyone within a quarter, unprompted — is concrete enough that someone could contradict it. But the sample is one company, the scorekeeper is the owner, and the businesses Main & Machine serves elsewhere are gestured at without a single second deployment described.
A case of one, stated as a rule
The prescription is modest; the generalizations are not. 'The team decides first, every time' and the claim that fast-adopting teams never have the best software are drawn from a single shop plus unnamed clients. The advice — name your human-only work before you buy anything — costs nothing and may well be right; it simply hasn't been demonstrated at the width it is asserted, and the productivity figure closing the argument is the author's own math on his own product.
The builder writing up his own build
Main & Machine sells AI systems to small businesses, and this account names the product it sold to the author's own lender. The disclosure is early and honest, which counts for something, but the shape stays intact: a trust-building sequence whose final step is a purchase order, published on a platform where founders write their own copy. The only account of how MARCUS performs is the account of the firm that ships it.
Clear on the setup, blind on cause
What this story is, we can read off the page: one interested witness, one deployment, disclosed self-interest. Where we cannot follow it is causation. The sequencing may have earned the adoption, or handing a team the transcription work it already resented would have landed under any framing at all — and no one outside B:Side has tried the experiment either way.