Ertas for AI Automation Agencies

AI automation agencies can replace per-client API pass-through with fine-tuned models they own. Train a custom model for each client in Ertas Studio, export it, and run it on hardware you or your client control, so per-token API bills give way to a predictable hosting cost and inference stays on infrastructure you choose.

The Challenge

AI automation agencies building chatbots, voice agents, and workflow automation for clients face a brutal margin problem. Every client engagement means another OpenAI or Anthropic API key, another line item of variable per-token spend that scales with usage, not with value delivered. An agency managing 10 to 20 or more clients on GPT-4 or Claude can easily burn through AU$3,000 to 5,000 per month in pure API pass-through costs, and those costs are completely unpredictable. A single client's spike in usage can wipe out the margin on an entire account. Tools like Make.com, n8n, Voiceflow, and Stammer.ai make it easy to wire up AI-powered workflows, but they all funnel inference through the same commercial APIs, leaving agencies with zero control over their largest variable cost.

Beyond cost, the differentiation problem is even more existential. When every agency is reselling the same GPT-4 or Claude API behind a slightly different prompt template, there is no moat. Clients eventually realise they can cut out the middleman and call the API themselves. Meanwhile, client data (customer conversations, proprietary business context, sensitive operational details) flows through third-party infrastructure with every API call. Enterprise clients increasingly push back on this, demanding to know where their data is processed and stored. Agencies that cannot answer "your data never leaves our infrastructure" are losing deals to competitors who can.

The Solution

Ertas turns an agency from an API reseller into a provider of custom AI. Instead of paying for a separate API subscription per client, the agency trains a model for each client in Ertas Studio on that client's own data: tone of voice, product catalogue, FAQ answers and past conversations. Studio runs LoRA fine-tuning of open models such as Qwen, Llama, Gemma and Mistral on Ertas's cloud GPUs through a visual canvas, so nobody on the team writes training scripts or rents GPUs. Several runs can train in parallel and be compared side by side before one goes live.

Each finished model is exported from Hub as GGUF for llama.cpp, Ollama or LM Studio, or as full safetensors for vLLM and other server runtimes. The agency runs it on its own server, or hands it to a client who wants it on their own premises. From then on the model runs with no Ertas dependency and no per-token bill, so serving a client costs the hardware and hosting the agency already budgets for, and the conversations the model handles stay on that infrastructure. Training data is uploaded to Ertas to train the model, so strip personal details from conversation logs before upload.

A model trained on a client's real conversations tends to handle that client's domain better than a general model steered by a long prompt, and the agency can check this on its own held-out examples before switching. When quality needs to move, the agency retrains in Studio, compares the new run with the current one, and swaps the exported file.

Key Features

Studio

Per-Client Fine-Tuning

Upload a client's conversation logs, product data or knowledge base as JSONL, or import a dataset from Hugging Face, then configure a LoRA fine-tuning run on the visual canvas. Studio trains on Ertas's cloud GPUs and produces a model that carries that client's domain and tone, with no training scripts and no GPU infrastructure to manage.

Studio

Base Model Selection

Pick a base model from Studio's catalogue of open models, including Qwen 2.5, Llama 3.2 and 3.1, Gemma, Mistral 7B and Phi-4 Mini. Train small and larger candidates in parallel on the same client data and compare the results before deciding which size each engagement tier needs.

Studio

Client Review in Data Craft

Build or clean each client's dataset in Data Craft, Studio's guided dataset builder. Add and edit USER and ASSISTANT turns, version the examples, and track which ones the client has approved, so the model learns answers the client has signed off.

Hub

A Separate Model Per Client

Hub lists every model the agency has trained. Download each one as GGUF for Ollama, llama.cpp or LM Studio, or as safetensors for vLLM, and serve it from the agency's server or the client's. Each client's model is trained only on that client's data, and running it needs no connection to Ertas.

Example Workflow

An AI automation agency in Melbourne manages chatbot and voice agent deployments for 15 small-to-medium business clients across real estate, dental and trades. All inference currently runs through GPT-4 via Make.com and Voiceflow integrations, at AU$4,200 per month in API fees, and three high-volume clients account for AU$1,800 of that. The agency starts with its highest-spend client, a real estate agency whose chatbot handles 12,000 conversations a month about listings, inspection bookings and pre-qualification questions. The team exports six months of conversation logs (45,000 message pairs), removes names and phone numbers, and uploads the result to Ertas Studio as JSONL. In Data Craft they fix weak answers and mark the examples the client has approved. They train three LoRA runs in parallel on Qwen 2.5 3B, Qwen 2.5 7B and Llama 3.1 8B, and compare each against a held-out set of real questions and the answers the current GPT-4 setup gives. The smallest run that holds the client's quality bar is exported from Hub as GGUF and served with Ollama on a Mac Mini in the agency's office. The agency sends part of the client's traffic to the new model first, checks the answers, then moves the rest across. Each further client gets its own model the same way, and the agency's AI spend shifts from a per-token bill that grows with every conversation to an Ertas subscription plus hardware it owns.

Compliance & Security

When the exported model runs on the agency's server or the client's premises, the conversations it handles stay on that infrastructure and are never sent to a third-party AI provider. That gives agencies a concrete answer to the data-handling questions in enterprise procurement questionnaires. Training works differently: Studio trains on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Remove or mask personal details in conversation logs before upload, keep only the examples the model needs, and delete datasets from Studio when an engagement ends. Obligations under the Australian Privacy Act, GDPR or a client's contract remain for the agency and the client to assess.