Don't Hire Agents to Run the Old Envelope

Most teams do not start an AI project by asking what the customer still needs. They start by listing the desks they already have.

Someone writes summaries because two systems cannot read each other. Someone copies a quote from one screen into another. Someone “checks with finance” because the promise lives in a thread instead of a record. Those steps have owners and meeting time, so they feel like work. When agents arrive, the instinct is to hire digital labor for the same desks.

That is how an ambitious AI program becomes a more expensive version of the old envelope.

The customer did not ask for a better envelope. They asked for an authorized, accurate quote, a durable record of what was sold, and a clear path when the request is not ordinary. If two tools cannot share that record, the person sitting between them is not a job. It is a defect with a salary. An agent in that chair is only a faster messenger.

AgentC Foundry’s stance is blunt: do not rent agents to operate inherited handoffs until you have named which work should still exist.

Start at the outcome, not the current procedure

Take a blank sheet. Name five to eight value streams that actually pay the bills: win the work, deliver it, keep the customer, collect the money, and the exceptions that keep you honest. Write each outcome in customer language.

Then look at the current standard operating procedure. Almost every “required” step will try to reappear. Some of those steps are real: a licensed review, a payment hold, a safety check. Many are translation layers. They exist because a form cannot be trusted, a file cannot be found, or two tools were never designed to share a record.

If you start from the current desk map, agents will be assigned to the translation layers. They will look busy. They will consume tokens. They will produce unread reports about unread reports. The bill will rise while the customer still waits.

If you start from the outcome, some desks disappear. The remaining work is either ordinary and repeatable, or it is an exception that needs judgment. That split is the whole design.

Ordinary work gets a thick harness. Exceptions get a thin one.

Once the leftover work is real, model choice is a cost-to-serve decision, not a leaderboard.

Ordinary requests — the 99 percent you can actually name — belong on a cheaper model wrapped in a thick harness: instructions, tools, checks, allowed paths, and a visible “done” test. You are not trying to impress anyone with intelligence. You are trying to make a known job cheap, inspectable, and boring.

Exceptions — high consequence, unclear, or outside the named 99 percent — belong on a stronger model with a thin harness. You paid for judgment. Do not clip it with a twenty-step chain that pretends the edge case is ordinary. Route it, log it, and put a human on the consequence.

This is the opposite of how many teams spend. They put the most expensive model on every production copy-paste, then starve the rare decision that actually needed it. Or they put a weak model on the exception with no structure, then conclude that AI does not work.

The token dashboard will not save you. Spend that shortens a real customer wait is different from spend that produces six agents writing status no one reads. Both can look like adoption on a chart. Only one is cost-to-serve.

What to delete before you hire

A practical test for a small operating company:

1. Name the customer outcome in one sentence. If you cannot, you are not ready to rent an agent.
2. Circle every step that exists only because systems cannot share a record. Those are not agent jobs. They are integration, filing, or permission work.
3. Keep only the steps that change a real state: a quote that can be honored, a job that can be scheduled, a payment that can be collected, an exception that can be decided.
4. Split leftover work into ordinary and exception. If you cannot name the ordinary 99 percent, you do not have a routing door. You have a hope.
5. Assign harness thickness after the split. Cheaper model plus a thick path for the ordinary work. Stronger model plus a thin path for the exception. Never the reverse as a default.
6. Refuse unread output. If no one picks up the artifact, it was not service. It was theater.

This is not a halt rule. A loop that cannot stop is a different failure. This is a prior question: should the loop exist at all?

It is also not model shopping. You can unbundle models all year and still pay rent on desks that should have been deleted.

What this looks like in a real shop

Take quoting. The inherited envelope is familiar: a salesperson drafts, operations retypes, finance looks at it, someone emails a PDF, and a follow-up meeting exists because nobody trusts the folder.

An agent hired into that envelope will draft, retype, and email faster. The customer still does not have one authorized quote with a record.

The redesigned job is narrower: one quote object, one authority, one exception door. Ordinary configurations complete on a cheap, tightly checked path. Odd configurations go to a human with stronger assistance, not through five inboxes.

The same pattern shows up in intake, scheduling, collections, and content operations. Anywhere people summarize because tools do not talk, an agent will happily summarize forever. Anywhere the business can store the actual object — the quote, the ticket, the approved file — the summary desk can go. Deleting that desk does not look like innovation. It looks like a missing meeting and a shorter wait.

Ambition is not the token bill

Ambition is useful. Cheaper tokens are useful. Neither is a strategy if you are paying digital labor to carry paper between desks that should not exist.

The cost-to-serve question is older than agents. Agents only make it louder. When the unit of labor gets cheaper, inherited process looks affordable again. That is the trap. Affordable waste is still waste, and it compounds because agents do not get bored of doing the wrong job.

AgentC Foundry helps businesses with operations, AI, and agentic-AI challenges build real harnesses. The harness is not a pile of prompts. It is the owned path around the work that should still exist: what “done” means, which model is allowed, how thick the checks are, and where a human owns the exception.

If you want agents, start with a red pen. Delete the envelope. Then hire.