PHASE 08 · SCALE

Scale & the 30 / 60 / 90 Plan

You have one evaluated, bounded prototype and a human-owned cadence. Now define acceptance gates for any additional workflow, shared infrastructure, and the conditions to stop.

Self-paced · write an acceptance-gated rollout Plan + commit Final phase

From one assistant to a team

One passing prototype does not authorize more workflows. Add only a bounded task whose owner, data boundary, evals, approvals, monitoring, incident response, and rollback are ready.

★ THE SCALING ORDER

Add the next assistant only when (a) the previous workflow passes its task-specific evaluation and required human review AND (b) you've identified another bounded, worthwhile task. Add one at a time so ownership, permissions, failure handling, and rollback remain clear.

There is no universal team size. Every added assistant creates maintenance, security, and coordination cost. Use the smallest set that improves a measured workflow; introduce orchestration only when the benefit exceeds that added complexity.

The 4 scaling traps

Building 10 assistants no one uses. Every assistant has maintenance cost — feedback logs, version updates, knowledge file refreshes. If you're not using one regularly, archive it. A team of 4 active assistants beats a team of 12 mostly-dormant ones.

Wiring up an automated chain before each component has been evaluated. Trace the full chain, test partial failure and retries, make writes idempotent, and keep high-impact actions behind explicit approval. A single aggregate success rate can hide a critical weak link.

The "every shiny new tool" trap. A new model or framework launches and you consider rebuilding everything. Define the task, baseline, minimum material improvement, migration risk and re-evaluation trigger before testing; no fixed multiplier or waiting period applies.

Ignoring the measured tradeoff. Under deadline pressure, teams may bypass the workflow—or force a weak workflow where manual work would be safer. Track why operators override it and improve, narrow, or retire it accordingly.

The "company of one + AI team" model

The endgame this roadmap aims at — for solos, freelancers, and small operators — is a model where you are the brain, the AI team is the hands, and your output looks like that of a small agency.

Imagine a freelance performance marketer evaluating this operating model. The following numbers are placeholders for planning, not a verified case study or expected outcome:

Before using a scenario like this as a target, replace every number with your baseline, include software and review cost, and test whether quality, safety, margin, and workload actually improve.

Templates library — what to keep, refine, share

By month 2 you'll have a small library of artefacts. Treat them like assets. They are.

📁SOPs/ — every SOP you've written. Each version-stamped. The most valuable folder you own.
📁system-prompts/ — generic, versioned prompt contracts with owners and eval links; no tenant or secret context.
📁knowledge/<tenant>/ — tenant-isolated sources, access, encryption, retention, deletion, export, and audit. Share only approved non-sensitive methods and taxonomies.
📁feedback-logs/ — your training history. The curriculum if you ever rebuild.
📁monthly-reports/ — your transformation record. Re-read every quarter.

On a cadence appropriate to the workflow, do a library audit: which SOPs are stale? Which configurations have not been re-evaluated after a model, tool, policy, or data change? Which knowledge files reference products or clients that no longer exist? Archive carefully and retain an audit trail.

An evidence-gated rollout plan

Advance by acceptance criteria rather than calendar milestones. Extend, narrow or stop a stage whenever evidence, risk review or operational capacity requires it.

STAGE 1 · FOUNDATION

Evaluate one bounded, low-risk workflow

STAGE 2 · CADENCE

Build the operating system + second assistant

STAGE 3 · CONTROLLED EXPANSION

Several bounded workflows with controlled handoffs

After the rollout evidence is accepted

Further expansion is not rinse-and-repeat. Each task, tenant, integration, and handoff creates new failure modes and requires its own threat model, evals, approvals, incident plan, and rollback.

Specialise harder

Take your current setup and niche it down: "AI Director for Meta + Shopify" beats "AI Director for marketing." The narrower your domain, the better your assistants and the higher your rates.

Productise the system

SOPs, prompts, evaluation cases, controls, and knowledge files can form part of a professional service. Scope and price from discovery, risk, integration, maintenance, and local market evidence—not a generic duration or revenue range.

Run multiple "books"

Reuse non-sensitive patterns, never client data or credentials. Keep each client's storage, retrieval index, secrets, configuration, logs, approvals, and deletion process isolated, and test for cross-tenant leakage before use.

The metric that matters most

No single number proves that the roadmap worked. Use a balanced view of task quality, critical failures, human-review burden, elapsed time, cost, accessibility, privacy, and operator experience.

★ THE METRIC

Compare the accepted workflow with its documented baseline.

Time spent directing versus executing can be useful, but interpret it alongside corrections, incidents, outcome quality, and total cost. More clients, higher rates, or less burnout do not follow automatically.

Your final task

Open your AI Director — My Roadmap.md file. Write the closing entry.

  1. Date: today.
  2. What I built (assistants, SOPs, knowledge files).
  3. What % of my week is now directing vs executing (honest estimate).
  4. The next assistant I will build, and by when.
  5. The skill I will practice until the next evidence review, with a date and acceptance criterion.

Save it as a versioned decision record. Revisit it at the scheduled evidence review and compare actual quality, risk, effort and cost with the baseline.

One last thing

The operating model of online work keeps changing, but urgency is not a substitute for evidence. Durable practice comes from bounded experiments, clear ownership, representative evaluation, documented limitations, and the willingness to stop a workflow that does not help.

The useful identity shift is from doing every step manually to designing and supervising a dependable process. Keep human accountability where consequences matter.

You now have a framework to test that approach. Use it carefully and revise it from evidence.

If you finished, comment "DONE"

Most people who download a 50-page guide read the first 5 pages. If you actually read all 8 phases — and especially if you did the exercises — let me know on Instagram @avamartech. I want to know who's serious. The people who reach back are usually the ones I end up working with directly.

Now go build your first SOP. ✌🏼