Don't pay for an AI agent until you know what done means

AI agents start to sound cheap when the price moves from "per user" to "per completed task". It is tempting. A customer question resolved for a small fee. A sales lead recommended only when it seems worth following up. Less subscription fatigue, more "pay when it works".

But there is a catch small teams can miss: AI cannot know what "done" means in your business if you have never written it down.

That is today's practical signal. Not that every company needs a new CRM agent. Not that every small business should connect AI to everything on Monday. The useful signal is simpler: before an agent books, sorts, recommends, writes, or follows up, write a definition of completed work.

Today's signal: AI is being priced as completed work

HubSpot announced that Breeze Customer Agent and Breeze Prospecting Agent move to outcome-based pricing from April 14, 2026. According to HubSpot, Customer Agent costs $0.50 per resolved conversation. Prospecting Agent costs $1 per lead recommended for outreach. Both are paid through HubSpot Credits.

The important part for Hammer readers is not the exact credit model. It is that vendors are starting to sell AI as completed work, not only as text generation or more seats.

Source: HubSpot: Customer Agent & Prospecting Agent Move to Outcome-Based Pricing

HubSpot describes Breeze as AI built into the CRM, with access to customer data, conversations, and deal history. It highlights agents for customer service, prospecting, data enrichment, meeting preparation, follow-ups, and AI-search visibility.

Source: HubSpot: Breeze AI Tools for Marketing, Sales & Service

Salesforce points in the same direction from a different angle. In its Summer '26 release, Salesforce writes about multi-agent orchestration, Slack-first workflows, real-time data activation, and AI-powered customer engagement. The release is larger and more enterprise-focused, but the pattern is the same: AI is moving closer to real business workflows.

Source: Salesforce: Summer '26 Release

For a Nordic small business with only a few people, this is both promising and a little uncomfortable. Promising because recurring work can get help without hiring for every small process. Uncomfortable because unclear routines become more expensive when AI starts acting inside them.

An AI agent needs more than a prompt

An AI agent is an AI system that can follow instructions, use tools, and try to complete a task across several steps. It can be as simple as reading new form responses, suggesting the right next step, and drafting a reply. It can also be more advanced: checking a CRM, sending a message, creating a task, and logging what happened.

It sounds like magic until you see where it usually breaks.

A human often knows when something is "done" without being able to explain it. An administrator sees that a booking change still lacks payment status. A teacher notices that a parent email is technically correct but will land badly. A shop owner knows that a product question should not be answered before stock has been checked.

The AI agent does not see that quiet experience. It sees instructions, sources, tools, and the rules you actually gave it.

That is why the first step should not be "connect AI to the CRM". The first step is to choose one small task and write a plain definition of done.

Write the done rule before you calculate the cost

Outcome-based pricing feels safer than a large subscription. But it asks a question many teams have not answered: what counts as a successful outcome?

For a customer service agent, "resolved conversation" may mean that the customer received an answer and did not write back. But for your business, a conversation may only be resolved if:

• The customer received the right answer.
• The answer used the latest price list or policy.
• The agent did not promise a discount, delivery date, or custom solution without approval.
• The case was logged, so the next person can see what happened.
• The agent escalated when the question involved contracts, health, children, money, or dissatisfaction.

For a prospecting agent, "recommended lead" sounds clear. But you may want a lead to count as done only if the agent can show:

• Why the contact fits your target customer.
• Which source supports the judgment.
• What the next step should be.
• Who on the team reviews before anything is sent.
• What weekly cost or volume limit applies.

This is not bureaucracy. It is how you avoid AI producing activity that looks good in a dashboard but creates more work for the people.

35 minutes: create a definition of done for one agent task

Do not choose your biggest process. Choose a boring, recurring task where you already know what a good result looks like.

Good starting points:

• A booking change that often arrives by email.
• A reminder about missing information before invoicing.
• A course registration that needs the right group, date, and next step.
• A quote request where budget, location, and timing must be known first.
• An internal weekly summary from meeting notes and tasks.

Then do this:

1. Write the task in one sentence. Example: "When a customer wants to move a booking, the agent gathers the right information and drafts a reviewable reply."
2. Write what counts as done. Example: "Done means the new date option, any fee, source, and responsible person are included. Nothing is sent without human approval."
3. Write the stop points. Example: "Stop if the customer is angry, payment is missing, the booking involves a custom solution, or the information conflicts."
4. Write which sources the agent may use. Example: booking terms, price list, calendar, CRM notes, and previous customer dialogue. Start with read access where possible.
5. Write the log you want to see. Example: source, proposed action, uncertainty, cost per unit, and who approved it.
6. Set a small budget limit. Example: maximum 20 cases in the first week, or a small fixed amount before you review the results.

That is enough for a first pilot. Not perfect. But clear enough to see whether the agent helps or just moves work around.

Copy this prompt: define done for the first agent

Use this prompt in any AI chat before building anything in HubSpot, Salesforce, Make, Zapier, Notion, Airtable, or your own system.

You are my AI operations partner. Help me define one small task an AI agent can help with without us losing control.

Business: [short description]
Task we want to test: [one recurring admin, customer, sales, school, or document task]
Tools/sources available: [CRM, email, calendar, documents, price list, form, policy]
Human owner: [role/person]

Do this:
1. Write the task in one sentence.
2. Write a clear definition of "done".
3. List which sources the agent may read.
4. List which actions the agent may suggest but not perform by itself.
5. Write stop points where the agent must escalate to a human.
6. Suggest a simple log: source, decision, uncertainty, cost, and approver.
7. Suggest a first test with maximum 20 cases or one week of runs.
8. Write three test cases: one simple, one messy, and one where the agent should stop.

Be concrete. If something is unclear, suggest a reasonable default and mark it as an assumption.

This is a better starting point than asking "which AI agent should we buy?". Tool choice gets easier once you know which task should become done.

Three small workflows worth testing

Booking change to reviewed reply

The agent reads the customer's email, checks the calendar and terms, suggests two new times, and drafts a reply. The human approves before anything is sent. Done means the customer gets a correct option, not that the agent has "answered quickly".

Missing invoice information to clear reminder

The agent sees that information is missing, retrieves the customer name, project, and deadline, and writes a short reminder. It may not change the invoice or threaten fees. Done means the right person gets the right question with the right context.

Course registration to next step

The agent reads a registration, matches the requested level with course options, and drafts the next step. It stops if the student is a minor, if special needs are mentioned, or if payment questions are unclear. Done means a human can make a quick decision.

Integrate safely without making AI useless

Safe AI integration is not about keeping AI away from everything important. Then it mostly becomes an expensive writing tool. The point is to give the agent the right access, the right limits, and a clear receipt for what it did.

Practical patterns go a long way:

• Start with read access before write access.
• Use environment variables or a secret manager for API keys.
• Give scoped permissions, not full access to the whole tool.
• Require approval before messages are sent, records are changed, or money is affected.
• Redact sensitive fields from AI logs when they are not needed.
• Keep an audit log: input, source, proposed action, human decision, and timestamp.

Then the agent can do real work. It can read, prepare, sort, and suggest. The human does not have to start from zero, but keeps the decision where it matters.

Hammer's angle: build the routine before the agent

For many small teams, the first win is not a full-scale AI agent. It is writing the routine clearly enough that a person, a temporary colleague, and an AI agent can follow the same logic.

That is Mindset Forge when you decide what the task is worth and where the boundaries are. It is Tool Forge when you connect the right systems with the right permissions. It is Skill Forge when the team learns to read logs, improve instructions, and adjust stop points without needing a consultant for every small change.

If you take one thing from today's signal, take this:

Do not pay for "AI that works" until you have written what completed work means. Otherwise, you are buying activity. Not outcomes.

FAQ

What is outcome-based AI pricing?

Outcome-based AI pricing means the customer pays when a defined task is completed, such as a resolved customer conversation or a recommended lead, instead of paying only per user or per generated response.

What is an AI agent in a small business?

An AI agent is an AI workflow that can use instructions, data, and tools to try to complete a task. In a small business, that can mean sorting incoming questions, preparing replies, updating internal lists, or suggesting the next step.

What should you write before connecting AI to a CRM, email, or calendar?

Write the task, the definition of done, allowed sources, stop points, permissions, budget limit, and the log you want to review. That makes the integration safer and makes it easier to see whether the AI agent is actually saving time.