How to read the numbers on this page
Most vendor case studies present a projection as if it were a result. This page does not. Each
figure carries one of three labels:
Measured We ran it and observed the number ourselves — on
the client's own data, not a benchmark set.
Baseline The client measured and reported it during
discovery. It describes the problem, not our result.
Modeled A projection built from the client's own stated
assumptions. It is an opportunity estimate, not an outcome.
The honest summary: most of these engagements are discovery and architecture
work — market research, workflow mapping, a scoped build proposal and an ROI model. Where a
system has been built and run, we say so and give the measured figure. Where it has not, the
number is modeled and labelled as such.
Vertical 01
Marketing and agency operations
Mid-market marketing agencies, martech platforms and multi-location marketing teams. The
recurring pattern: institutional knowledge fragmented across a CRM and a chat tool, delivery
capacity capped by headcount, and a growth model that cannot scale without proportional hiring.
Six engagements.
Marketing consulting & talent placement
Consultant-placement firm: winning enterprise placements on submission speed, with 20 years of project history locked in a CRM
~30 staff · 200+ consultant network · enterprise end-clients · placements routed through a managed-services procurement platform
The problem
A significant revenue stream ran through a managed-services procurement platform where
speed of submission decides who wins the placement. Twenty years of project
history — who did what, for whom, how well — sat fragmented across a CRM and Slack, unusable for
matching a consultant to a live requisition. Network signal (job moves, promotions, company news)
that should trigger outreach at the right moment was going unnoticed.
What we built
- Relationship nurture agent — monitors network changes and surfaces the moment to reach out
- Institutional knowledge base — 20 years of CRM and chat history made searchable for talent matching; this is the foundational data layer the other two workflows read from
- Case-study generator — sales collateral assembled from existing project data
The architectural finding that shaped the sequencing: the three workflows are not independent.
The knowledge base has to exist first or the other two have nothing to reason over.
| Phase | Focus | Duration |
|---|
| 1 | Signal monitoring + submission speed | ~3.5 weeks |
| 2 | Institutional knowledge base (CRM + chat archive) | ~4 weeks |
| 3 | Case-study generation | ~4.5 weeks |
Numbers
Baseline $100K+ in lost margin per missed placement signal — client-stated.
Modeled ~12-week delivery across three phases.
Status: in active delivery. This is a live build, not a
proposal. Measured outcomes will be published here when the phases complete.
Healthcare/legal marketing SaaS
Marketing SaaS platform: a 5-person onboarding team losing 64% of its working hours to manual admin
1,800+ clients · ~$7,800 average annual contract · ~200 employees
The problem
The onboarding team was the growth constraint. Each specialist spent 15 hours a
week chasing clients for missing inputs, hand-copying AI meeting summaries into the CRM,
relaying developer feedback between a design-review tool and clients, and writing agendas and
recaps by hand. The client had already deployed its own AI assistant, so anything we built had to
be complementary rather than redundant.
What we built
An AI operations agent that monitors the enterprise email stream, extracts structured data
into the CRM in real time, and puts human-in-the-loop decisions in front of a person as chat
cards — inside the tools the team already uses. No new dashboard, no replacement of any existing
system. A proxy integration path was designed around the design-review tool's API limits.
Email-stream extractionCRM write-backHITL chat cardsAdditive, no system replacement
Numbers
Baseline 64% of working time on manual admin · 15 hrs/week per specialist · 6,600 hours/year across the team.
| Phase | Investment | Annual value recovered | Payback |
|---|
| 1 — Email intelligence + follow-up | $72,000 / 12 weeks | $239,000 | 3.6 months |
| 2 — Feedback-loop automation | ~$38–48K | ~$130,000+ | ~4 months |
| Phases 1+2 | ~$110–120K | ~$370,000/yr | ~4 months |
Modeled Value figures assume 5 specialists at ~$75–85/hr loaded cost. Phase 1 recovers 1,340 hours directly and unlocks access to 3,400+ more.
Healthcare digital-marketing SaaS + services
Agency serving ~1,000 same-vertical practices: turning cross-client performance data into a moat no single client could build
~1,000 medical-aesthetic practice clients · hybrid SaaS-plus-services model · investor-owned, under board pressure to become a platform
The problem
A headcount-dependent agency being pushed to become a scalable technology business. Net
revenue retention was the board's concern — existing clients did not grow their spend. The
client arrived with two automation pilots already scoped: blog production and SEO workflow.
What we found that they had not scoped
The valuable asset was not either pilot. It was cross-client intelligence:
with ~1,000 clients in a single vertical, performance patterns can be synthesised across the
portfolio to improve outcomes for every client — a structural advantage no individual practice
could ever replicate, and one that directly addresses the retention problem.
| Workflow | Current state | AI opportunity |
|---|
| Blog content production | Manual, low throughput | Vertical-tailored drafts, scaled across the client base |
| SEO workflow | Human keyword research and optimisation | Automated pipelines with cross-client benchmarking |
| Cross-client intelligence | Siloed per-client data, no synthesis | Synthesis layer across the portfolio → competitive moat |
Numbers
Modeled Scale from ~1,000 to 5,000+ clients with no proportional headcount increase; agency-multiple to software-multiple re-rating modelled at $9–26M of enterprise value.
Baseline Direct labour savings are modest here — offshore delivery staff are inexpensive. We said so. The case rests on scale enablement and the exit multiple, not on cost-out.
Pattern worth naming: in an agency with many clients in one
vertical, the automation ROI is usually not labour cost. It is the data asset the client is
sitting on and not using.
Digital marketing agency — healthcare vertical specialist
Specialist agency: the reported problem was dashboard reporting; the real gap was lead-quality attribution
Agency specialising in therapy practices and mental-health providers
The problem, and the correction
Discovery surfaced a claim that the client's reporting tool lacked historical trend
visualisation. Research showed that was false — the tool has extensive built-in
functionality for exactly that. We removed the claim from the proposal before the pitch. The
genuine gap was lead-quality tracking, which requires integrating marketing
platforms with healthcare practice-management systems — a materially harder problem with
regulatory implications, and one worth actually solving.
What we delivered
- Full discovery, questionnaire and analysis
- Interactive ROI calculator (web-based) so leadership could model their own scenario
- An AI dashboard preview — the future operational state, shown rather than described
- Every claim fact-checked against the discovery transcripts before the meeting
Why this one is here: the most useful thing we did was
delete a claim that would have been challenged in the room. A proposal that cannot survive
fact-checking is worth less than a smaller one that can.
Multicultural marketing agency
Social-media automation at scale: A/B testing across personal profiles versus company pages
Specialist brand agency
The work
Competitive research plus a 55-question discovery instrument covering content strategy, A/B
testing capability, personal-profile versus company-page strategy, competitor monitoring,
post-publication engagement and analytics integration.
The architectural finding
The major social platform APIs expose enough post-performance data on their own that
external UTM tracking is not required for this use case. That removed a whole
tracking layer from the proposed architecture — a simplification, found in discovery rather than
discovered mid-build.
Multi-location franchise retail
~140-location franchise: campaign execution coordinated by hand across every unit
~140 locations · hybrid corporate/franchise model · proprietary product line at ~60% margin
The problem
Data fragmented across units, franchise-level marketing coordinated manually, customer
feedback never aggregated, recruiting workflow inefficient — and a high-margin product line
materially under-exploited.
What we delivered
Full business and competitive research, a 100+ question internal discovery document across 9
domains, converted into an 85-question client-facing questionnaire. Notably, a dedicated
ROI baseline section capturing average ticket value, visit frequency, database
size, marketing hours per campaign, email performance, feedback volume, recruiting cost and
product attachment rate.
Pattern: capture the baseline metrics during discovery, in
writing, before anything is built. Without them there is no way to prove the result later — and
most engagements that cannot show a result simply never captured the "before".
Vertical 02
Commercial mortgage servicing and investment operations
Lending, servicing and investment firms running multi-billion-dollar portfolios on
spreadsheets, documents and email. The recurring pattern: high-value professionals doing document
assembly by hand, and a deliberately lean team that will reject any system requiring them to
change where they work. Two engagements.
Commercial real-estate mortgage banking
Commercial mortgage servicer, ~$10B portfolio: 30–90 page loan submission packages assembled by hand over days
$9.56–10.5B commercial loan servicing portfolio · 1,600+ loans · 47 states · member of a national network of independent servicers
The problem
Core operations ran on spreadsheets, word processing and manual assembly. A single loan
submission package — 30 to 90 pages — took days to build. Insurance compliance
monitoring across 1,600+ loans, investor reporting on the industry standard templates, and
document chasing were all manual.
| Workflow | Current state | AI opportunity |
|---|
| Loan submission package assembly | 30–90 pages assembled manually over days | Document classification, extraction and template assembly → hours |
| Financial statement spreading | Dozens of rent-roll formats from different property systems → industry standard | Extraction at 90–99% accuracy (established vendor tooling) |
| Document collection / chasing | Phone and email follow-up for certificates, returns, financials | Automated portals and reminders → 70–80% less follow-up time |
| Insurance compliance monitoring | Manual policy comparison across 1,600+ loans | Rules engine + policy scanning → fully automated |
| Investor reporting | Monthly/quarterly cycles taking days | Template-driven automation → minutes |
Numbers
Baseline ~$12.5–25M/year of servicing fee income on the current portfolio (12.5–25bps on ~$10B). Loans originate at $500K–$500M each.
Modeled Originators freed from document assembly write more loans; the same solution deploys across the other member firms in the network by configuration rather than rebuild, multiplying addressable value roughly 4×.
The repeatability argument, stated plainly: this vertical's
firms are structurally near-identical. The variation between them is configuration — different
templates, thresholds and filing portals — not architecture. That is the specific condition under
which one build becomes a product rather than a one-off.
Family office / investment management
Family office, 50+ portfolio companies: two CRM implementations already abandoned because they required data entry
8–10 person team · 4 investment professionals · 50+ investments across private equity, real estate, venture, energy, aviation and agriculture · $500M+ AUM
The problem — and the constraint that determined the architecture
The team had already implemented and abandoned two separate systems, both for
the same reason: they demanded manual data entry. The client stated the constraint directly:
"a system that lives outside email is extra work and will therefore not be used." Deal
activity was tracked by asking; "what's the latest?" emails were consuming investment-team time.
What we built
A zero-interface agent. No new login, no dashboard, no data entry. It reads
the existing email flow, structures deal activity automatically, and returns a weekly digest by
email. Human-in-the-loop decisions arrive as chat cards in the tool the team already has open.
The design principle was that the system must be invisible to the user.
Numbers
Baseline 10–25 deal opportunities evaluated monthly, 0–3 closed annually. Team estimates 15–20% of working hours go to manual status tracking and data hygiene.
Modeled 4 professionals × $200–250/hr loaded × 15% of 2,080 hrs = $250–310K/year of recoverable capacity. Proof-of-concept scoped at $10–15K.
Pattern worth naming: where a team has already rejected two
tools, the binding constraint is adoption, not capability. The correct architecture is the one
that requires no behaviour change at all — which usually means the agent goes to where the work
already happens rather than asking the work to move.
Vertical 03
Regulated document and compliance workflows
Where a regulator dictates the document, every firm in the vertical has the same problem in
the same shape — and the existing software is workflow digitisation with no extraction or
reasoning layer. Two engagements, including the one measured technical result on this page.
Healthcare regulatory analytics
Regulatory compliance data provider: 10,000+ documents a year at $15 each, across 46 different state formats
~6,000 facilities served · 13% penetration of a 45,000+ facility addressable market · data used by lenders and investors in multi-million-dollar decisions
The problem
Processing 10,000+ regulatory compliance documents annually at $15 per document,
a cost that scaled linearly with market expansion. The growth constraint, in the client's words:
"we can't scale into new states without scaling the team." Federally-regulated facilities
file on one standardised federal form; state-regulated facilities file on 46 different
formats across 46 states. Accuracy requirements were medical-grade — bad data could
affect $30M+ investment decisions.
What we built — before writing the proposal
We built a working extraction pipeline against the client's own sample documents prior
to proposing the engagement, and measured it.
- Confidence-based routing — high-confidence extractions process automatically; uncertain documents route to a human reviewer. There is no gap period and no wrong data, by construction.
- Dual pipeline — the standardised federal form is extracted programmatically from its structured tables; the 46 state formats are converted and handled by a reasoning layer.
- Format change handling — a new state is a planned schema update measured in days, not a retraining cycle. Unplanned format changes are caught by a confidence drop and routed to review.
Numbers
Measured 95.9% raw extraction accuracy across 26 test documents from three facilities, on the client's real documents. This is an observed result, not a benchmark figure.
Baseline 10,000+ docs/year × $15/doc = $150,000/year in manual processing cost.
Modeled 80–90% reduction → $110–130K/year saved; ~10-month payback on a $100K production build. The larger effect is structural: a headcount-constrained growth model becomes a software-scalable one.
Why the 46 formats are the moat, not the obstacle: the
format variance that makes this hard to build is the same thing that makes it hard for a new
entrant to copy. In regulated document workflows, the messy part is usually the defensible part.
Environmental compliance consulting
Stormwater compliance: 20–25 hours of manual labour per permit document, in a market with zero AI tooling
10–19 person consultancy · under $5M revenue · state agency and municipal clients · operating under the most stringent construction stormwater permit regime in the United States
The problem
A single pollution-prevention plan takes 20–25 hours of manual labour to
produce. The governing state permit is the strictest in the country and carries multi-million
dollar enforcement penalties. Despite that, market research found no AI-powered tools in
the category at all — the dozen-plus incumbent vendors offer workflow digitisation only.
Why we treated it as a product rather than a services engagement
The signals that distinguish a productisable pattern from a bespoke build were all present:
a regulator-mandated document format identical for every firm in the vertical, no AI-capable
incumbent, and proven feasibility in adjacent regulated-document markets (construction permitting
automation is a funded, working category). So the proposal was a joint venture,
not a services contract: domain expertise and municipal relationships from the partner, platform
from us.
Numbers
Baseline 20–25 hours of skilled labour per plan document.
Modeled Four exit scenarios over a five-year horizon, from an early-traction trade sale to a category-leading position — a wide range, built from M&A comparables in adjacent regulated-document software. Treat these as scenario planning, not forecast.
Vertical 04
Platform strategy and partner evaluations
Engagements where the deliverable was a market and architecture decision rather than an
automation build. Included for completeness — the shape of the work is different. Three engagements.
Workplace wellness AI
Coaching platform: which of ten proposed AI agents are actually uncontested?
Founder-led consultancy moving from solo practice to platform · $5.34B global coaching market growing ~9% CAGR; AI coaching segment projected to $8.2B by 2032 at 27% CAGR
The work
A 26,000-word strategic analysis: market sizing, competitive mapping across all ten proposed
agent categories, and business-model evaluation. Well-funded incumbents (raising in the
hundreds of millions) each own a narrow slice.
The finding
Two of the ten categories — team conflict resolution and workplace
wellness compliance — have virtually no dedicated AI solutions. The recommendation was
to build those two first and ignore the eight where funded competitors already sit, with an
app-marketplace launch as the distribution channel the incumbents had not tapped.
Pattern: when a client arrives with a list of ten things to
build, the highest-value deliverable is usually the competitive analysis that removes eight of them.
Marketing automation SaaS — channel partnership
Partner-ecosystem evaluation: positioning inside a platform's partner network rather than competing with it
Platform with $100M ARR and 8,000+ brand clients
An evaluation of referral and implementation partnership with a major messaging platform. The
differentiating angle identified: most of that platform's partners do configuration and
migrations. The gap is custom AI workflow layers — document extraction and operational reasoning
built on top of an existing platform rather than replacing it.
AI governance
AI agent governance vendor: partnership fit assessment
Strategic evaluation
A capability and strategic-fit assessment against an AI agent governance vendor. No client
automation engagement — included because it is part of the record.