No Licence Fee: Agent-First Software That Charges for the Work

A software licence made sense when the product waited. You paid for the right to open it. Whether anyone completed a case, matched a payment or cleared a queue was your problem, and the invoice arrived either way.

Agent-first software does not wait in the same way. Its job is to take a piece of work from intake to a checked result. Once the product itself does the task, the honest unit of price is the task, not the seat. The next generation will not carry a licence fee. It will charge for work done, and only for work whose outcome can be shown.

Why a licence is the wrong meter

A licence prices access. Access was a fair proxy when every extra user needed a copy of the software, a login and a place in the training room. It is a poor proxy for an agent. The agent does not occupy a chair. It either finishes the job or it does not.

That mismatch is already familiar. Firms buy a company-wide licence, run a short launch, and then discover that the tool is available while the work is unchanged. The invoice is real. The outcome is optional. Staff invent their own uses, some of them useful and some of them a quiet risk, and nobody can tie the spend to a finished case.

Usage fees do not fix this on their own. Tokens, calls and minutes are still a meter of effort inside the machine. A busy agent can be expensive and useless. A quiet one can be cheap and decisive. If you cannot point at the result, you are still buying activity.

What "the work" has to mean

Outcome pricing fails the moment the outcome is a slogan. "Improved productivity" cannot be invoiced. A finance director and an operations lead have to be able to look at one case and agree that it counts, or that it does not.

A usable outcome is small enough to point at:

• A membership renewal prepared, logged, and waiting only for the decision a person must still make.
• A payment reference matched to the open item, with the exception written down when it does not match. A person still releases the money.
• A case pack assembled from documents the organisation already holds, against a checklist, with the gaps marked rather than invented.
• A handover between two systems that used to live in a spreadsheet, completed to the rule that was written down before the agent started.

Each of those can be counted. Each can be opened afterwards and checked. Each has a version that is not done: the wrong member, the unmatched payment, the missing document, the handover that broke the rule. Unfinished work is not a discount. It is not billable as success.

This is the same boundary we described for agents in UK operations. A job description, a named owner, and a stop. Outcome pricing simply sends the invoice to the same place. If you cannot name the finished state, you cannot price it, and you should not automate it yet.

The commercial terms that keep it honest

Removing the licence does not make the software free. The model, the records, the audit trail and the people who design the job all cost money. The customer pays when the work is finished. The supplier still has to cover the cost of attempting it. Both sides need the rules written before the first case, not argued after the first surprising invoice.

• A definition of done, in one sentence, agreed by the person who owns the process and the person who will pay.
• Evidence on every charge. What went in, what came out, and who accepted it. A line on an invoice with no case behind it is a licence under another name.
• A clear treatment of failure. Reversed, rejected or stopped work is not billed as a completed outcome. Say so in the price, including any separate fee for the attempt if there is one.
• A cap. An agent that loops must not be able to write its own unlimited bill. The cap is part of the design, not a favour.
• A human acceptance where the risk sits. Paying for the outcome does not move a regulatory duty off the organisation that owns the customer, the member or the payment.

There is a second trap. A supplier can choose a metric that is easy to inflate: messages sent, files touched, steps taken. A customer can refuse to accept finished work because acceptance triggers the fee. Both behaviours show that the outcome was badly chosen. The unit should be something the customer wanted done anyway, checked the way they already check a person.

What this does not replace

Not every product should drop its subscription. Software that answers a lookup, holds a record, or keeps a service available is still selling access, and a subscription can be the honest price for that. The shift belongs to software whose job is to complete work.

Judgement that cannot be specified in advance is a different purchase. A board still needs advice it can question. An awkward case still needs a person. Those should stay visible in the operating model, not be forced into a per-item fee that pretends every hour has a countable output. Governance, skills and the records underneath decide whether an outcome fee is safe. A price list will not.

Nor does outcome pricing excuse a vague pilot. If the queue has no definition of done, another commercial model will only make the argument louder. Find the work first. Price it second.

How to test it on one queue

Pick one process you already measure. Write the outcome in a sentence both operations and finance will accept, and write the cases that do not count. Run the agent on that queue for a month beside whatever you pay today. Compare cost per completed case, not cost per seat, and read the exceptions. Widen the price only when the exception rate is boring and the evidence for each charge can be opened by someone who was not in the room.

If you cannot name the case, start with AI Discovery. If the team is already building the agent, the commercial rule belongs in the same design as the technical one: that is the work of Agentic AI technical lead and audit. The invoice should describe the same job the agent was allowed to finish.