The problem
A services business does not fail because the work is bad. It fails because the work does not scale, and the founder becomes the bottleneck on every path.
Response time decided who won the client. Inquiries arriving in the evening or during delivery work sat until someone was free. The client who replied first was frequently not us, and by then the conversation had happened elsewhere.
Qualification was inconsistent. Whether an inquiry got the right questions depended on who saw it and how busy they were. Poor-fit inquiries consumed the same time as good ones, sometimes more.
Production was a per-item craft. Every piece of copy and every visual started from nothing. Quality was reliable; throughput and cost were not, and both scaled linearly with headcount.
Reporting was reconstructed monthly. Pulling numbers together from several places by hand, late, with the result being a document nobody could act on because the moment had passed.
The legacy platform we inherited made all of this worse. Roughly 173 standalone page scripts and 30 scheduled scripts, with no router, no autoloader, hand-rolled cryptography instead of vetted libraries, raw database access and no queuing or caching layer anywhere. It ran, and it worked, and every operational task around it was manual — not because nobody had tried, but because a codebase with no seams offers nothing to automate against. That is the concrete reason the replacement was built around invariants, queues and explicit state machines.
Everything depended on the founder. Intake, judgment, quality control, escalation. That is a business that cannot take a vacation, and it is a hard ceiling on growth.
Previous workflow
Before: a services business run by hand
Every delay compounds. A slow first response makes a slower brief, which makes a later delivery.
- Human action
Inquiry arrives
- Problem: Sat until someone was free
- Problem: Evening and weekend inquiries waited overnight
- Problem: Faster competitors won the conversation
- Human action
Someone reads and qualifies it
- Problem: Question quality varied by person and workload
- Problem: Poor-fit inquiries consumed full attention
- Problem: No consistent record of what was asked
- Human action
Discovery call and manual brief
- Problem: Brief quality depended on note-taking
- Problem: Requirements often revisited later
- Problem: Rework discovered at delivery
- Human action
Copy written from scratch
- Problem: Cost and time scaled linearly with volume
- Problem: Tone consistency depended on the individual writer
- Human action
Visuals produced from scratch
- Problem: The most expensive step per unit of output
- Problem: Iteration cycles measured in days
- Human action
Delivery and client review
- Problem: Feedback arrived as unstructured messages
- Problem: Revisions restarted the production step
- Human action
Monthly reporting assembled by hand
- Problem: Numbers pulled from several places
- Problem: Reported too late to change anything
- Problem: No consistent measure across clients
- Human action
Goals and non-goals
Goals
- Every inquiry gets a qualified response in minutes, at any hour.
- Qualification asks the same questions regardless of who is available.
- Content and creative production start from a structured brief rather than a blank page.
- Reporting is assembled continuously rather than reconstructed monthly.
- The founder is not on the critical path of routine delivery.
- Every client-facing commitment is reviewed by a human before it is made.
Non-goals
- No autonomous client communication. AI drafts. A person sends anything that constitutes a commitment.
- No AI-authored final deliverables without review. Generation is a starting point, not an output.
- Not a replacement for strategy. Positioning and creative direction stay human.
- No client data in general-purpose tooling beyond what the engagement permits.
My role
As founder and operations lead I own this system completely — which also means the honest framing of it is mine to give.
- Discovery — mapped where time actually went across the business, which is how the response-time problem surfaced as the highest-leverage fix rather than the production cost everyone assumed.
- Workflow design — the intake, content production, creative production and reporting workflows documented in the workflow library.
- Boundary setting — decided which steps AI produces and which a human owns, and held that line when speed pressure argued against it.
- Tooling selection and integration — including the deliberate choice not to put client data into general-purpose tooling.
- Quality control design — the review step is the product; generation is the cheap part.
- Measurement — defined the operating measures the business is actually run against.
- Team enablement — trained staff to work with the system rather than around it, which was harder and more important than building it.
Product solution
Intake and qualification. An inbound inquiry is acknowledged and qualified immediately. The system extracts what the inquiry is about, asks the qualifying questions that a good account manager would ask, and produces a structured record. A person reviews before any commitment — timeline, price, scope — is made. The speed gain is in the first response and the structured brief, not in replacing the conversation.
Content production. Work starts from a structured brief carrying audience, objective, constraints and tone, rather than from a blank document. AI produces drafts against that brief. A human edits, fact-checks and approves. The brief is the leverage: the same brief produces consistent output regardless of who runs the step.
Visual creative production. Concepts and variants are generated against the same brief, then selected, refined and finished by a person. This is where the cost reduction is largest, because the expensive part of visual production was always the first ten discarded attempts.
Reporting. Campaign and engagement data is collected continuously and assembled into client-facing reporting rather than rebuilt monthly. The interpretation — what this means and what to do about it — is written by a human.
Delivery infrastructure. Campaign execution runs on MailerClub, which grew out of this business's own needs and became a platform in its own right.
Workflow
After: inquiry to delivered work
AI moved into the production steps. Human approval sits on every point where a commitment is made to a client.
- External service
Inquiry arrives
Any channel, any hour.
- AI / LLM
Immediate acknowledgment and qualification
Intent extracted, qualifying questions asked, structured record produced. Under two minutes to first response.
- Automated software
Routing and scoring
Deterministic rules on the structured record decide priority and owner. Not a model decision.
- Approval gate
Human review before commitment
Decision point: A person reviews the qualification and owns anything involving timeline, price or scope.
- Human action
Discovery conversation
The relationship step. Not automated, and not a candidate for automation.
- AI / LLM
Brief synthesis
Conversation and requirements structured into a brief carrying audience, objective, constraints and tone.
- Approval gate
Brief approved
Decision point: A person confirms the brief is right before production starts, because everything downstream inherits it.
- AI / LLM
Content drafting
Drafts produced against the approved brief.
- AI / LLM
Visual concept generation
Concepts and variants generated against the same brief.
- Human action
Edit, select and finish
Editing, fact-checking, selection and finishing. This step is the deliverable; generation was the input.
- Approval gate
Quality review
Decision point: Approved for delivery, or returned to production with specific notes.
- Automated software
Campaign execution
Delivery runs on the sending platform, with its own suppression and protection rules.
- Data store
Results collected continuously
Engagement and campaign data accumulated rather than reconstructed.
- AI / LLM
Reporting assembled
Draft narrative generated from collected data.
- Human action
Interpretation and recommendation
What it means and what to do next is written by a person and owned by them.
- Monitoring / feedback
Feedback into the brief templates
What performed well changes the next brief, which is where the compounding is.
- Human action
- Automated software
- AI / LLM
- Data store
- External service
- Approval gate
- Monitoring / feedback
Architecture
Operating system — sanitized view
An operating model rather than a single application: production tooling wrapped around human approval points.
Client-facing
Inquiry channels
External service
Inbound inquiries arriving through several routes.
Intake and qualification
AI / LLM
Immediate acknowledgment, intent extraction and structured qualification.
Human commitment gate
Approval gate
No timeline, price or scope leaves the business without a person owning it.
Production
Structured brief
Data store
Audience, objective, constraints and tone. The unit everything downstream inherits.
Content generation
AI / LLM
Drafting against the approved brief.
Visual generation
AI / LLM
Concept and variant generation against the same brief.
Edit and finish
Human
Editing, fact-checking, selection and finishing — the actual deliverable.
Quality review
Approval gate
Approval to deliver, or return with specific notes.
Delivery
Sending platform
Automated
Campaign execution with suppression enforcement and delivery protection. See the MailerClub case study.
Legacy platform
External service
The inherited monolithic system this operating model was built to work around and progressively replace.
Measurement
Continuous collection
Data store
Engagement and campaign data accumulated as it happens.
Report drafting
AI / LLM
Narrative draft generated from collected data.
Interpretation
Human
What it means and what to do next — written and owned by a person.
Brief feedback loop
Monitoring
Performance data updates the brief templates, which is where improvement compounds.
Diagram is sanitized for publication: it shows responsibilities and data flow, not hostnames, addresses or credentials.
Where AI is used — and where it is not
AI / LLM responsibilities
- Immediate acknowledgment and qualifying questions on an inbound inquiry.
- Extracting intent and requirements into a structured record.
- Synthesizing a conversation into a brief.
- Drafting content against an approved brief.
- Generating visual concepts and variants.
- Drafting the narrative section of a report from collected data.
Human and deterministic responsibilities
- Any commitment to a client — timeline, price, scope. Human, always.
- The discovery conversation and the relationship.
- Approving the brief before production starts.
- Final edit, fact-check, selection and finishing.
- The quality gate before delivery.
- Strategy, positioning and creative direction.
- Routing and priority scoring — deterministic rules on the structured record.
- Every number in a client report.
- Campaign execution, suppression and delivery protection.
Decisions and tradeoffs
Automate the first response, not the conversation
- Options considered
- Fully automated sales conversation · Automated response only after human triage · Automated qualified acknowledgment with human takeover
- Chosen
- Immediate qualified acknowledgment, human takeover before any commitment
- Reason
- Speed to first response is what wins the inquiry. Commitment is what carries the risk. Those are separable, and only one of them needs a person.
- Cost and risk accepted
- Clients occasionally notice the first message is assisted, and the handover has to be handled well or it reads as a bait and switch.
- Result
- First response under two minutes without ever letting a model make a promise on the business's behalf.
Invest in the brief rather than in prompts
- Options considered
- Prompt libraries per task · Fine-tuned models · A structured brief that every production step inherits
- Chosen
- The structured brief as the central artifact
- Reason
- Output quality tracked brief quality far more than prompt quality. A good brief produces good output from almost any competent tool; a bad brief cannot be rescued by prompting.
- Cost and risk accepted
- Briefing is a real step with real time cost, and people wanted to skip it under deadline pressure.
- Result
- Consistency became independent of who ran the production step, which is what made delegation possible.
Generate many visual concepts and discard most
- Options considered
- Manual production throughout · Generate and use directly · Generate broadly, then select and finish by hand
- Chosen
- Generate broadly, select and finish by hand
- Reason
- The expensive part of visual production was never the final piece. It was the ten discarded attempts before it. Generation makes those effectively free.
- Cost and risk accepted
- Selection fatigue is real, and unfinished generated work is recognizably generated.
- Result
- The largest single cost reduction in the business, with the finishing step protecting quality.
Build a sending platform rather than buying one
- Options considered
- Continue on the inherited legacy platform · Buy a commercial platform · Build
- Chosen
- Build, which became MailerClub
- Reason
- The legacy platform offered no seams to automate against, and commercial platforms did not give us control over deliverability behavior or the data.
- Cost and risk accepted
- By a wide margin the largest investment described anywhere in this portfolio.
- Result
- A platform that is now a product in its own right — but this was a bet, and a smaller business should probably have bought.
Keep the founder on quality, off production
- Options considered
- Founder stays across everything · Full delegation · Founder on the quality gate only
- Chosen
- Founder on the quality gate and on strategy, off routine production
- Reason
- Routine production was the ceiling. Quality judgment was the thing that could not be delegated immediately without losing the standard.
- Cost and risk accepted
- The quality gate is now the founder-shaped bottleneck. Better than the previous one, still a bottleneck.
- Result
- Routine delivery runs without me. The honest position is that the constraint moved rather than disappeared.
Challenges
Early AI content was fluent and empty. It read well and said nothing. The problem was not the tool; it was that we were prompting instead of briefing. Once the brief carried audience, objective, constraints and tone, the same tools produced usable drafts. That diagnosis took longer than it should have because the output looked fine.
People routed around the brief step under deadline pressure. Exactly when it mattered most, someone would skip briefing to save fifteen minutes and lose two hours in revisions. Making the brief a required input to the production step, not a suggested one, was the only thing that worked.
The intake handover reads badly if done clumsily. An inquiry answered instantly and then going quiet for a day is worse than a slower, consistent response. The handover had to be explicit and quick, or the speed advantage turned into a trust problem.
Unfinished generated visuals are recognizable. There is a distinctive look to work that went straight from generation to delivery. The finishing step is not optional polish; it is what makes the output ours.
The bottleneck moved rather than disappeared. Getting the founder off routine production was a real change. Putting the founder on the quality gate created a new constraint in a better place. I have not solved this, and claiming otherwise would be dishonest.
The metrics on this page are self-reported. They come from my own delivery records, not from an audited system, and they are labeled that way on every surface of this site. I would rather state the provenance than present them as something they are not.
Testing and validation
| Layer | What it covers | Evidence |
|---|---|---|
| Acceptance criteria | Response time, brief completeness and quality-gate criteria defined before rollout. | Written operating standards for each workflow. |
| Pilot | Each workflow run in parallel with the manual process before replacing it. | Parallel running period per workflow. |
| Output quality review | Generated drafts reviewed against the brief before the workflow was trusted. | Review notes fed back into brief templates. |
| Intake accuracy | Qualification records checked against what the human conversation established. | Discrepancies drove qualification question changes. |
| Client feedback | Delivery quality and responsiveness tracked through the engagement. | Feedback captured per engagement. |
| Operational monitoring | Response time, brief completion rate, revision cycles per deliverable. | Tracked continuously as operating measures. |
| Rollback | Every automated step retained a manual path for the period after cutover. | Manual fallback documented per workflow. |
Results
| Measure | Before | After | How it is measured |
|---|---|---|---|
| Time to first qualified response | Hours, or overnight for inquiries outside working hours. | Under two minutes, at any hour. | Measured from inquiry receipt to first qualified response. Self-reported. |
| Project turnaround | Baseline manual process. | 40% faster end to end. | Inquiry to delivery, after intake, briefing and production handoffs moved onto the system. Self-reported. |
| Recurring manual steps | Intake, briefing, production and reporting performed by hand. | 70% of recurring manual steps eliminated. | Counted across intake, production and reporting workflows. Self-reported. |
| Creative production cost | Fully manual production, cost scaling with volume. | 45% lower. | Visual and written creative production after AI-assisted generation with human finishing. Self-reported. |
| Clients served | — | 170+ cumulative. | Across campaign, creative and automation engagements. Self-reported. |
| Founder dependency | On the critical path of intake, production and quality. | On the quality gate and strategy only. | Routine delivery runs without founder involvement. The constraint moved rather than disappeared. |
All figures are self-reported from delivery records rather than an audited system. They are stated with that provenance everywhere they appear on this site.
Evidence
Intake and qualification record
A structured qualification produced from an inbound inquiry, before human takeover.
Placeholder — evidence pending sanitization review before publication. Brief template
The structured brief that every production step inherits.
Placeholder — evidence pending sanitization review before publication. Creative variant set
Generated concepts alongside the selected and finished deliverable.
Placeholder — evidence pending sanitization review before publication. Client reporting view
Continuously collected engagement data with the human-written interpretation.
Placeholder — evidence pending sanitization review before publication.
Lessons
What worked. Separating speed from commitment. Automating the first response while keeping every promise human captured almost all the competitive benefit with almost none of the risk. This is the pattern I have reused most.
What failed. Treating AI output quality as a prompting problem. Months went into prompt refinement that a structured brief solved in two weeks. The input artifact mattered more than the instruction.
What changed. The brief became the central artifact of the business. It is what makes delivery consistent regardless of who runs a step, and it is what allowed delegation at all.
What I would do differently. I would have measured the manual baseline properly before automating. I have honest figures for the improvement because I kept records, but I would rather have had a rigorous before-state than a reconstructed one — and the reason those numbers are labeled self-reported on this site is that I did not.
What this demonstrates about AI-product leadership. This is the project where AI does the most visible work, and it is also the one with the strictest human boundary. Every commitment to a client is human. Every number in a report is computed. What AI actually bought was the removal of the blank page and the overnight delay — which turned out to be most of the value, without any of the risk that comes from letting a model speak for the business.
Next milestone
- Delegate the quality gate. The founder-shaped constraint moved rather than disappeared. Making the quality standard explicit enough to hand over is the next real change.
- Close the performance loop automatically. Brief templates are updated by hand from what performed well; that feedback should be systematic.
- Retire the remaining legacy platform dependencies, so operational work no longer routes around a system with no automation seams.