Cool Demo. Can You Afford to Run It Every Day?
Most AI demos work. They summarize the document. They draft the memo. They answer the question in seconds. Then you try to run that same workflow every day, across a team, at scale.
That is when the real math comes into play.
The cost stack nobody talks about
Take a single "valuable" agent task. Something real, like drafting an investment memo, summarizing research, or updating a CRM after a call.
On the surface, it looks cheap. A few cents per run. Maybe less if you are using a smaller model.
But the actual cost stack looks more like this:
Model inference (often multiple calls, not one)
Retrieval and context assembly (vector search, embeddings, storage)
Orchestration logic (tool use, retries, fallbacks)
Latency overhead (which turns into real user time)
Monitoring and error handling
Engineering time to maintain and improve it
That "$0.03 task" quickly becomes $0.10 to $0.50 in real usage. Run it 50 times a day across a team, and you are at $5 to $25 per day. Per workflow. Per team.
Scale that across a firm, and you are no longer testing. You are operating.
The only question that matters
Is the output worth more than it costs to run?
The clean version of that ROI math:
Time saved per task x hourly loaded cost of that role
Error reduction or risk avoided
Speed to decision
Consistency across the team
If an analyst spends 60 minutes on something and the agent reduces it to 10, you have saved 50 minutes. At $150 per hour loaded cost, that is $125 of value per run. Compare that to $0.25 to run the agent, and the math is obvious.
But most workflows are not that clean.
Some only save 5 to 10 minutes. Some require human cleanup. Some fail just often enough to erode trust. And that is where most "working" AI falls apart in production. Not because it does not work. Because the economics do not hold up at scale.
What it looks like when you get it right
One of our clients receives a recurring legal document a couple of times a week. Marking it up manually took their team up to two hours each time.
We built an agent to do it.
The first version was not good enough. It cost $20 per document and hit about 75% accuracy. Not worth it. So we kept working: tightened the prompts, optimized the retrieval, and reduced unnecessary model calls. We reduced costs significantly and increased accuracy to above 95%.
What used to take two hours takes minutes. The economics justify it clearly, and the workflow is reliable enough that people actually trust it.
That is the path. Not "deploy AI everywhere." Start with one workflow. Measure the real cost. Measure the real time saved. Improve it until the economics justify daily use, or conclude they never will, and move on. Then do the next one.
What the teams getting this right are actually doing
They are not chasing a broad AI transformation. They are doing something much simpler:
Pick one workflow
Measure real usage cost
Measure real-time saved
Tighten the loop until it is reliable and cheap enough to run every day/week/month
AI is not blocked by capability anymore. It is gated by whether you can run it daily without second-guessing the cost, reliability, or output quality.
That is the difference between a demo and a system.