Applied AI and agents
Content generation, agents that do the work and MCP servers. The hard part is not making it work once in a demo: it is a cost you can quote, a result somebody can review, and nothing going out without a person seeing it.
- Who it is for
- Businesses with a repetitive, expensive manual task (writing, classifying, extracting data) who want to know up front what automating it will cost.
- Deliverable
- Flow in production · cost per operation measured and documented
What it includes
- The specific job, not "AI" in the abstract: writing, classifying, pulling data out of a document, or preparing a plan somebody approves afterwards.
- Cost measured per operation and capped per account, so the spend can be budgeted before it is switched on instead of showing up on next month's invoice.
- Human review by design: what the model produces arrives as a draft, and publishing it stays somebody's decision.
- Agents and MCP servers wired into your tools with narrow permissions: what it may read, what it may write, and what it must never do.
- Evaluation on your own cases: a set of real examples to compare against before and after touching a prompt or changing model.
Where it gets hard
A feature that costs pennies in the demo costs a salary at a thousand users. Consumption is measured per operation from day one, turned into a unit the customer understands (credits, for instance) and capped per account. Without that there is no price that holds.
The question is not whether it fails but what happens when it does: who sees it before the end customer, what is kept so it can be reproduced, and how you roll back. A flow that takes the first answer as good is a flow that publishes the mistake.
New models arrive every few months, cheaper or better, and the temptation is to change the name and deploy. Without a set of cases to compare against, nobody knows whether quality went up or down: only that the invoice changed.
The proof
The PlanVortex planner writes a week of posts from a topic, the customer's own photos or their shop's catalogue, and what it produces are drafts a person reviews. Consumption is charged in credits measured per operation, and its MCP server is published on npm.
- Check it at
- planvortex.com
What we have written about this
How we work
Half an hour of conversation, a closed proposal with scope, price and date, a deployed increment every two weeks in an environment you can log into, and handover. The four steps are on the home page.
Frequently asked questions
- Which model do you use?
- Whichever fits the task and the budget, and it can be changed: the code is not tied to one provider. What decides is not the brand, it is what each operation costs and how it does on your own test cases.
- Will my data be used to train a model?
- Not with the configuration we leave in place: we use the API terms, where the provider does not train on what you send. It is written into the contract and into your privacy policy, with the provider named, because it is a data sub-processor.
- Is this useful for my case or is it hype?
- It depends, and it is the first thing we look at. There are tasks where it does not pay off today: if reviewing the output costs more than doing the job by hand, we say so in the first conversation and there is no quote.