Case Studies — Meet the Agents

Case studies

Meet the agents.

Ten AI employees, built by AI Builders and measured on real work. Client names and identifying details are withheld; the numbers are real, and where a figure is modeled rather than audited, we say so. Every agent below runs with a human approval gate — nothing acts alone.

Full write-ups

Each one, answered as a question.

Every agent below has its own page — the situation, what we built, the numbers, and an honest note on where it stops working.

Can AI do bookkeeping and bank reconciliation?

Marge — AI Controller. A six-month backlog cleared in under a day; ~$110,000 in phantom cash found.

What does an AI executive assistant actually do?

Quinton — AI Executive Assistant. Zero money-on-the-line items missed; ~280 hours a year absorbed.

Can AI run Meta ad campaigns without a media buyer?

AdPilot — AI Media Buyer. 3,500+ clicks on ~$835 of spend, and no dark days in a 13-day window.

Can AI run social media posting across five platforms?

Lumberyard — AI Social Media Producer. Five platforms from one approval screen; ~16 hours a month back.

Can AI find government contracts a small business can actually bid?

Threshold — Government Opportunity Scanner. Eleven sources daily at $0 in data costs; the valuable output is the no.

Can AI write grant applications for a nonprofit?

Marlowe — AI Development Director. Funders no database lists, and a compliance stop before a liability.

Can AI build a WordPress website?

Sitewright — AI Website Builder. 62 pages built, migrated and self-audited in two working days.

Can AI reposition a company’s brand and pricing?

Mercer — AI Repositioning Officer. Six specialist roles in sequence, seven days, one continuous engagement.

Can AI replace a strategy consultant?

Cassandra — AI Strategy Bench. Five personas in deliberate tension; its best output was a correction.

Can AI conduct an interview?

Sawyer — AI Interview Agent. A journalist’s behavioural spec, three weeks of spare 20-minute blocks.

AI Controller

Marge

The client: A closely held construction & design group — three legal entities, ~199 ledger accounts, twelve live bank and card feeds.
The problem: Six months behind on reconciliation while the founder acted as his own CFO in the middle of a capital raise.

Marge was given governed read-and-draft access to the accounting stack and told to actually do the work: query the ledger, trace the discrepancies, draft the corrections — and post them only on explicit owner approval. She worked through thousands of transaction lines across twelve feeds, cleared $27,388.88 of stale prior-period payables in a single approved batch, and rebuilt a five-loan debt register that surfaced ~$8,358/week of debt service nobody was tracking in one place. Every action she took is in a permanent, auditable run log. Nothing in the closed fiscal year was touched.

~$110,000
phantom cash found and corrected
Under a day
to clear the six-month reconciliation backlog
$95k–$170k
modeled first-year economic impact
Straight talk: Marge posts nothing without human approval — zero unapproved entries, by design. The impact figures are modeled, not audited.

AI Development Director

Marlowe

The client: A startup nonprofit building a trades high school — no development staff, no grant history.
The problem: Grant funding stood between the mission and the doors opening, and hiring a development office was out of reach.

Marlowe runs like a configured development office: a research protocol, a scoring rubric, a compliance guardrail layer, and institutional memory, all pointed at one organization. Instead of listing famous funders that reject startups, it reverse-engineered a peer organization’s public tax filings to find the quiet family foundations that actually fund groups like this — four of the nine funders it scored can’t be found in any grant database. It checks every prospect against legal eligibility gates before a dollar of effort is spent, and it caught three compliance exposures before any became a liability — including stopping solicitation entirely until the organization was legally cleared to ask.

~70% less
human labor for a year-one development program (modeled 981 → 297 hours)
~4.5×
more grant submissions possible from the same founder hours
$32k–$84k/yr
modeled cost of the human equivalent
Straight talk: No grants have been won yet — the organization hasn’t submitted. What Marlowe changed is what’s possible per founder-hour, and its first big save was saying “stop” at the right moment.

AI Executive Assistant

Quinton

The client: A founder and chief executive running several companies at once.
The problem: An inbox that never stopped, and money-on-the-line items buried inside it.

Quinton runs a 7:00 AM morning brief every weekday — calendar, overnight triage, drafts ready for review — then makes roughly ten inbox passes a day, sorting everything and pinging the executive only when something is genuinely time-sensitive. In its first two weeks it caught, without a single miss: a lien threat, a past-due loan notice, legal demand letters, and a live five-figure financing deal moving through multiple lender swaps — all while the executive was on the road. It also ran a full multi-round redline on a multi-million-dollar construction contract (with a change log and draw-schedule workbook), built a podium-ready public-comment statement from a city’s own public records, and produced a six-figure client event proposal — and it files tasks and delegations directly into the team’s tools instead of leaving notes about them.

~280 hrs/yr
of recurring inbox and admin work absorbed
~25 hrs
of drafting and research absorbed across three projects in two weeks
Zero
missed money-on-the-line items
Straight talk: Quinton drafts and flags; the executive approves everything that goes out. Savings scale with whose hour it’s protecting — $7k/yr at admin rates, several times that at an executive’s.

AI Website Builder

Sitewright

The client: An award-winning custom home builder with an aging website.
The problem: A ground-up rebuild the agency market prices at $38,000–$55,000 and schedules across 10–20 weeks.

Sitewright works directly inside WordPress and the CRM — not beside them. It planned the architecture, designed the system, wrote 16 new pages from scratch, migrated 43 legacy items with zero loss, wired the conversion and CRM infrastructure, built custom software along the way, then audited its own work and caught issues a human team under deadline would likely have shipped. A human stood as the approval gate at every step; nothing published itself.

62 pages
planned, written, built, migrated, and audited
2 working days
of agent execution time
$38k–$55k
what the 2026 agency market prices for the same scope, over 10–20 weeks
Straight talk: The two days are agent execution time inside a ten-day engagement window — human review and approval time is real and isn’t hidden in the number.

AI Repositioning Officer

Mercer

The client: A technology company whose products had outgrown its positioning.
The problem: Market position, product line, pricing, financial model, and website all needed rebuilding — conventionally a ~$52,000, 3–4 month engagement across multiple vendors.

Mercer did the jobs of a market-intelligence analyst, a positioning strategist, a pricing consultant, a financial modeler, a brand copywriter, and a web team — in sequence, in one continuous engagement. It produced the market research, defined the category, rebuilt the product architecture and pricing, constructed the sales argument and financing rails, wrote the positioning brief and brand voice, and built an eleven-page website — in seven days.

7 days
from raw situation to launch-ready reposition
11 pages
of website built, plus pricing, financial model, and go-to-market
~$52,000
the conventional multi-vendor price for the same scope, over 3–4 months
Straight talk: The dollar figure is a modeled market-rate comparison, not an audited invoice — and revision rounds that cost weeks with agencies cost minutes here, which is half the point.

AI Strategy Bench

Cassandra

The client: Two ventures: a statewide property-data play and a technology company mid-pivot.
The problem: Strategy-consulting scope — market analysis, data architecture, pricing, go-to-market — without the budget or the appetite for four to six separate vendors.

Cassandra is five named analytical personas held in deliberate tension under a written rule set and a persistent knowledge base — a standing adversarial review board. Pointed at two ventures, it took both from raw thesis to execution-ready, delegation-ready specification. Its single most valuable output wasn’t a document at all: it was a correction — catching that a go-to-market plan led with cold outreach while a warm network sat unused, before capital was committed to the wrong sequence.

2 ventures
taken from concept to build-ready specification
$63k–$170k
modeled conventional consulting scope displaced
4–6 vendors
not hired, and 4–7 months not waited
Straight talk: These are modeled figures against published market rates — and the framework’s own documents attack their numbers on purpose. The honest claim: analysis that would never have been commissioned at any price got done, and it changed decisions.

AI Social Media Producer

Lumberyard

The client: A multi-brand media operation publishing across three brands.
The problem: Turning approved video into published posts meant an employee bouncing between five apps for every single post.

Lumberyard connects the whole pipeline: an approved video automatically becomes a transcript, the transcript becomes platform-specific captions in each brand’s voice — LinkedIn, Instagram, Facebook, YouTube Shorts, TikTok — with the right hashtags applied by rule. The team gets one approval screen with bulk scheduling, posts publish directly through the CRM, failures explain themselves in plain English with one-click retry, and real performance numbers flow back in automatically. It also replaced a paid scheduling tool outright.

~16 hrs/mo
of staff time saved at current volume
5 platforms
published from one approval screen
~$7,650/yr
estimated combined labor + software savings
Straight talk: Humans still approve every post before it goes live — the automation is the middle of the pipeline, not the judgment.

AI Media Buyer

AdPilot

The client: Two sister small businesses — a marketing agency and a general contracting firm — competing in a regional readers’-choice awards vote.
The problem: A fixed 13-day voting window, three live ad campaigns on Meta, and no media buyer on staff — while outside management for the same scope runs $1,500–$3,000 a month.

AdPilot started from zero and became the day-to-day media buyer for both ad accounts: it built the Meta Business Manager plumbing a developer would normally set up, designed and launched three campaigns, and produced a 13.5-second vertical video ad — scenes, motion, synced audio — with no outside editor. Then it ran the flight: an automated 8 AM daily recap, live budget forecasting against the deadline, and seven data-driven budget moves that kept dollars flowing to the strongest performer. When a brand-new video ad got just 2 impressions in 30 hours — quietly benched by the platform’s own auction — AdPilot caught it and restructured within a day, protecting about $120 of budget. Two account-level billing blockers that would have gone dark on delivery were caught and cleared the same day they appeared.

3,500+ clicks
delivered on ~$835 of managed spend — $0.21 CPC on the strongest campaign
Same day
every delivery problem and billing blocker caught and fixed — no dark days in a 13-day window
$1,320–$3,630
modeled labor equivalent for the flight (media buyer + video editor + daily management)
Straight talk: The dollar values are modeled from published 2026 market rates, not an invoice — and a human approved every launch and budget change. What AdPilot really removes is the days between a problem starting and someone noticing; that lag is where most wasted ad spend lives.

Government Opportunity Scanner

Threshold

The client: A small technology services firm entering the public-sector market with no past performance and no certifications.
The problem: The conventional entry path is a five-figure market-intelligence subscription plus a capture consultant on retainer.

Threshold scans eleven federal, state, and local sources every day — all free or keyless, $0 in data costs — and scores every notice against a written profile of what the firm can and cannot legally bid. Each morning it delivers a ranked brief with a human-readable reason for every verdict. On its first run it triaged 26 notices: 10 worth real bid hours, 5 structurally gated — flagged before anyone wasted a week on them. One first-week finding redirected the firm’s entire capability-investment plan; another flagged a compliance exposure before it became a filing.

11 sources
scanned daily, at $0 in data costs
~$2,400/yr
to run — 79–98% below the conventional intelligence-plus-consultant stack
1 working day
from nothing to its first ranked brief
Straight talk: Threshold’s real product is the no: telling you which opportunities you’re structurally barred from before you spend a single bid hour.

AI Interview Agent

Sawyer

The client: A chief executive running several companies — interviewed in 20–40 minute blocks, mostly from a car.
The problem: The expensive category of conversation that can’t be a form: the value is in the follow-up question. Ghostwriter discovery, brand-voice development, and deep intake all price accordingly.

Sawyer conducts long-form, on-the-record interviews with the behavioral specification of a professional journalist: one question per turn, a pre-loaded ledger of contradictions from the source documents, a tuned protocol for naming a dodge exactly once, adversarial on numbers and gentle on family. It surfaced factual conflicts the paper record held, accounts that differed from the public story, and operating rules the subject had never articulated about himself — then produced a structured five-part voice-and-decision profile other AI systems can now work from. The same skeleton redeploys as client intake, discovery calls, exit interviews, customer research, and succession knowledge capture.

Under $1,000
vs. the $22k–$45k human equivalent (ghostwriter discovery + voice development)
~3 weeks
in spare 20-minute blocks, vs. 8–12 weeks conventionally
$40k–$70k/yr
modeled saving when the same skeleton runs client intake vs. an in-house seat
Straight talk: It doesn’t save the subject’s own time — six hours of talking is six hours — and everything it captures is an unverified assertion until a human checks it.

Which one does your business need first?

That’s exactly what the audit answers — your operation, mapped, with the top five opportunities ranked by return.

Book Your AI Opportunity Audit — $1,000, credited toward your build

Free — no email required

The AI Answer Desk

Straight answers to what owners actually search at 11pm: what an AI employee is, what it costs, whether it’s safe anywhere near your books, and what happens when it gets something wrong. No gate, no download form, no drip sequence.