Fine-Tune Once, Charge Monthly: The Productized AI Service Model

The most common mistake AI consultants and agency owners make is treating fine-tuning as a project. A project has a start date, an end date, a deliverable, and a final invoice. The client is happy, you cash the cheque, and then you both move on.

This is leaving money on the table — and it is also leaving the client with a model that degrades over time as their business changes.

Fine-tuning is not a project. It is a capability that requires maintenance, improvement, and governance. The agencies that understand this have replaced project revenue with recurring subscription revenue — often at 4-6x the effective hourly rate of their project work.

Why Models Require Ongoing Care

A fine-tuned model captures the state of a client's business at the time of training. If you train on support tickets from Q4 2025, the model reflects Q4 2025 language, products, policies, and customer concerns. By Q4 2026, several things will have changed:

• Product catalog updates that the model does not know about
• Policy changes the model will give wrong answers on
• New edge cases that were not in the training data
• Quality drift as the task input distribution shifts

A model without maintenance becomes a liability within 6-12 months of training for most business use cases. This is not a flaw — it is the nature of any system that captures a snapshot of a changing business. It is also the natural business argument for a monthly retainer.

The Productized Fine-Tuning Service

Here is the service structure that converts project work to recurring revenue:

Tier 1: The Build (One-Time)

This is the initial engagement: data collection and cleaning, fine-tuning, evaluation, and deployment. Priced as a project, typically AU$3,000-12,000 depending on data complexity and deployment requirements.

Deliverables:

• Cleaned training dataset (the client owns this)
• Fine-tuned model deployed to Ollama or chosen inference platform
• Evaluation report with accuracy benchmarks
• API documentation
• 30-day support period

This is the engagement that generates the trust and the technical foundation for everything that follows.

Tier 2: Model Maintenance (Monthly Retainer)

After the 30-day support period, the client moves to a monthly retainer. The retainer covers:

• Monthly review of model output quality (sampling 50-100 production queries)
• Flagging quality drift or coverage gaps
• Minor prompt/system prompt updates
• Monitoring infrastructure (ensuring the inference server is running)
• Priority support for production issues

Pricing: AU$300-800/month depending on complexity and volume.

This tier is pure recurring revenue. Most months, the time investment is 1-2 hours — a monthly review session and any small adjustments. The value to the client is peace of mind and a point of accountability for their production AI system.

Tier 3: Retraining Cycles (Episodic)

Every 3-6 months (or triggered by specific events like a major product launch, policy change, or observed quality degradation), the model needs retraining on new data.

Deliverables:

• Collection and cleaning of new training data (new support tickets, updated documents, etc.)
• Retraining run incorporating old + new data
• Evaluation comparing new model to previous model
• Deployment of updated model
• Updated documentation

Pricing: AU$1,500-5,000 per retraining cycle depending on data volume and complexity.

This is the revenue event that repeats. A client with a well-functioning model will trigger retraining on a predictable schedule. You are not re-selling them; you are delivering the ongoing service they need to keep the model current.

The Annual Revenue Math

Single project approach:

• Initial build: AU$6,000
• Annual revenue from one client: AU$6,000

Productized approach:

• Initial build: AU$6,000
• Monthly retainer 12 months: AU$6,000 (AU$500/month)
• 2 retraining cycles: AU$6,000 (AU$3,000 each)
• Annual revenue from one client: AU$18,000

Same client, same technical work, 3x the revenue. With 10 clients, that is AU$60,000 in annual recurring revenue from retainers alone — before any new project work.

How to Have the Conversation

The transition from "project" to "ongoing service" requires a specific conversation. Most clients have not thought about model maintenance — you need to introduce the concept.

The frame that works: "The model we build together will be excellent on day one. But AI models are like a talented employee — they need to be updated as your business changes. What we are really building is a system for keeping your AI current, and that is an ongoing investment, not a one-time project."

This framing:

• Sets realistic expectations (the model is excellent but needs care)
• Positions maintenance as expected and normal, not a sign of failure
• Creates the category of "AI model maintenance" in the client's mind

The clients most likely to accept this framing are the ones who already have experience with software maintenance — SaaS subscriptions, managed IT services, accountants and lawyers on retainer. They understand ongoing service models. "Think of us like your AI system's IT team" is a frame that lands well with these clients.

Packaging the Monthly Retainer

The monthly retainer needs to feel like value, not like "we are charging you to do nothing most months." How to structure it:

Deliver a monthly report. Even a one-page PDF with: model health status, quality metrics for the month, upcoming retraining recommendation (yes/no), and any anomalies noticed. This makes the retainer tangible. Clients who receive a monthly report know they are getting something; clients who receive nothing quietly consider cancelling.

Include a fixed number of "minor adjustment" hours. 1-2 hours per month for things like updating the system prompt, adjusting the model's persona, or trialing a new configuration. This gives clients something concrete they can request each month.

Create a clear escalation process. When quality drops significantly, what happens? Who is notified? What is the SLA? Having this defined — in writing, in the service agreement — makes the retainer feel like real support, not just a recurring charge.

What This Requires of You

The productized model only works if you have a reliable, stable delivery infrastructure:

• The model needs to actually keep running month over month (Ollama as a systemd service, not a script you have to restart manually)
• You need tooling to run monthly quality checks without spending 4 hours per client (batch inference + automated scoring against your evaluation set)
• Your retraining process needs to be fast enough that a AU$2,000 retraining fee is profitable — if retraining takes you 20 hours manually, the economics do not work

This is where a fine-tuning platform like Ertas creates leverage. A retraining cycle that would take a day of manual Python work takes an afternoon with good tooling. The difference between a productized service with 20% margins and one with 60% margins is often the efficiency of the delivery infrastructure.

Getting Your First Retainer Client

Start with a client you have already delivered a successful project for. The ask is simple: "We built something that works well. We want to make sure it keeps working well. We offer a monthly maintenance service for AU$X that covers quarterly health checks and an annual retraining cycle. Would you like to keep it running well?"

If the client found the initial build valuable — and they will tell you if they did — the retainer conversation is an easy yes. They already trust you, they already have the model in production, and they understand the value.

The clients who say no to the retainer usually say no because they have not yet seen the model break or degrade. Give them 6 months and follow up.

---

Further Reading

• AI Agency Differentiation in 2026: Stop Reselling, Start Owning — Why proprietary models are the durable competitive advantage
• How to Price Fine-Tuning Services Profitably (Agency Rate Card) — Specific rate card and pricing methodology
• The Freelance AI Consultant's Stack in 2026 — Full toolkit for running a solo AI practice