How We Replaced 12 Hours of Manual Work With an AI Agent

The client had a problem that did not look like a problem on the surface. Three people on their ops team were spending about four hours each, every single week, doing the same sequence of tasks. Pull new leads from a form aggregator. Cross-reference them against a scoring rubric built in a spreadsheet. Write personalized outreach drafts. Log everything into HubSpot. Repeat.

Twelve hours a week. Every week. Nobody flagged it because it had always been done that way.

We built an AI agent to replace that entire workflow. Here is exactly how it works, what it cost, and what the real numbers look like seven months in.

The Workflow, Step by Step

The agent runs on a scheduled trigger every morning at 7am.

Step 1: Lead ingestion. It pulls all new form submissions from the previous 24 hours via a webhook connected to their lead capture tool. Each lead comes in as a structured JSON object with name, company, role, company size, source, and any self-reported context.

Step 2: Enrichment. The agent calls a lightweight enrichment API to append company revenue range, industry vertical, and LinkedIn employee count. This takes roughly 1.2 seconds per lead.

Step 3: Scoring. Here is where the model does real work. We fine-tuned a smaller open-source model on 800 historical leads the client had already manually scored. The model reads each enriched lead and outputs a score from 1 to 10 with a two-sentence rationale. It does not use GPT-4 or Claude for this step. A fine-tuned smaller model, running inference for fractions of a cent per lead, outperformed the frontier model on accuracy by 11 percentage points on their specific rubric. The $100k/month API bill that other companies are running into? Not a problem here.

Step 4: Draft generation. For leads scoring 7 or above, the agent generates a personalized outreach email. It pulls from a retrieval layer of approved messaging, case studies, and product context built specifically for this client. The output is a draft, not a sent email. A human reviews and approves before anything goes out. That human-in-the-loop step is not a compromise. It is the right call for their sales motion.

Step 5: CRM sync. Everything gets written to HubSpot automatically. Score, rationale, draft email, enrichment data. The sales rep opens HubSpot in the morning and sees a clean queue with context, scores, and ready-to-review drafts.

What It Cost to Build

Total build cost: $2,800. That covered workflow architecture, the fine-tuning process on their historical data, API integrations, the retrieval layer setup, and two rounds of testing with real lead batches.

The fine-tuning was the most time-intensive part, not the most expensive. Cleaning and formatting 800 historical leads into a usable training set took about six hours of work. The actual fine-tuning run cost less than $40.

Monthly operating cost: approximately $60. That covers enrichment API calls, model inference, and hosting for the orchestration layer. They process between 200 and 400 leads per month.

What It Saves

At a blended hourly rate of $50 across the three team members doing this work, twelve hours per week is $600/week. That is $2,400 per month, every month.

The build paid for itself in 35 days.

Beyond the dollar figure, two of the three team members shifted their time to actual sales conversations. The quality of outreach improved because the drafts are more consistent and better researched than what was being written manually under time pressure.

Why Most Businesses Are Not Doing This Yet

It is not that the technology is hard to access. It is three things.

First, nobody has mapped the workflow precisely enough to automate it. Most business owners know they have repetitive work. They could not tell you every step, every input, every output, and every decision point. You cannot automate what you have not documented.

Second, there is real fear around AI reliability in production. That fear is legitimate. Hallucinations in a customer-facing workflow can cause serious damage. The answer is not to avoid AI agents. It is to build the right fallback logic. Every high-stakes output in our client's system gets human review. The agent handles the volume work. The human handles the judgment call.

Third, the conversation around AI has been dominated by demos and hype for three years. Founders are tired of being sold transformation. They want to see a real workflow, real costs, and real results before they commit budget. That skepticism is healthy. It just means the bar for proof is higher now.

The Practical Takeaway

If you want to find your own version of this, do one thing today. Open your calendar and tag every recurring task your team does that follows a predictable input-output pattern. Not creative work. Not judgment calls. Predictable, repeatable steps with a defined start and a defined end.

For most businesses, that audit surfaces two or three candidates inside an hour. One of them is probably worth building. The build cost is lower than most founders expect. The monthly cost is often under $100. The payback period is usually under 60 days.

The hard part is not the AI. It is doing the documentation work first.