An agent on your data, or a ChatGPT subscription

Where a paid subscription is genuinely enough, where it quietly fails and what an agent trained on your own documents adds.

If your team already pays for ChatGPT or Copilot, this is a fair question: what does a custom agent add that a subscription does not already do? Sometimes the answer is nothing. Here is how to tell which case you are in.

What a subscription gives you

A very capable general assistant that knows nothing about your company. It drafts, summarises, translates and writes code well. Every employee gets the same blank slate, and whatever context an answer needs has to be pasted in by hand, by that person, every time.

Where that quietly fails

  • Nobody pastes forty pages of HR policy before asking how many vacation days they have left. They guess, or they ask a colleague, or they ask HR — which is the cost you were trying to remove.
  • Answers differ from person to person, because the context each one pasted differed. Two employees get two versions of the same rule and both think theirs is official.
  • There is no trace. A month later nobody can tell what the assistant said, which document it leaned on or whether it was right.
  • Nothing enforces who may see what. The subscription has no idea that this person's salary band is not that person's business.
  • Personal data ends up pasted into a chat window by people acting in good faith, which is exactly the thing your data policy was written to prevent.

What an agent adds

  • Your documents are indexed once and searched on every question. Nobody pastes anything.
  • Each answer cites the section it came from, so it can be checked and trusted.
  • It knows who is asking. Personal balances, entitlements and dates come from the employee's own record, not from a generic policy quote.
  • Every conversation is logged, so HR, legal or you can read back what was actually said.
  • It sees only what your system handed it for the person asking, and it keeps no memory between conversations — where one shared subscription account sees everything anyone ever pasted into it.
  • Changes are regression-tested against a fixed set of questions before they go live, so quality moves in one direction.
  • Cost follows work done, not headcount: cents per dialog, whether five people ask or five hundred.

When a subscription is honestly enough

A handful of people. General-purpose work — drafting, brainstorming, code, translation. No rule that has to be applied identically every time, no audit requirement, no personal data in play. In that situation a custom agent is an expensive way to get something you already have.

When it stops being enough

The same question gets asked dozens of times a month and the answer must be identical every time. Someone must be able to point at the document behind an answer. The data involved cannot be pasted into a public chat. A person is spending hours a week re-typing, looking up or repeating themselves. Any two of those together, and the arithmetic usually favours an agent — the scenarios show that arithmetic per unit of work.

How to check on your own data

Do not take this on trust. A two-week pilot runs the agent on your documents with a small group of your people and gives you the numbers for your process: how many questions it closed without a human, where it fell back, what a dialog cost. Then the comparison is yours to make, not mine to claim. See how a pilot runs, or book a call and bring the questions your team actually asks.