Less routine work with AI and checks that fit the task

The workflow is delivered and set up for your company. An AI agent is a program that uses AI to carry out a defined task. In the workflow, it also has clear requirements for what to deliver and when to stop.

The price is checked before the quote is sent

An invented example: a kitchen company regularly prepares quotes from a price list. An agent gathers the information and creates a draft, but includes an old price. Another agent compares the quote with the current price list and sends the incorrect line back for correction.

After the correction, the quote is checked again. The employee receives the draft with a short summary of the correction and any open questions. The quote waits for the employee's approval before it can be sent. If the current price list is missing, work stops and the employee gets a note saying what is missing.

The benefit lies in the preparation. The employee does not have to retrieve every detail and transfer it into the quote. Prices are compared before the final review, catching the wrong price in this example before the customer sees it. Professional judgement and approval stay with the company.

Checks that match the consequences

Each task gets a level based on how far a mistake could reach. A small change can need thorough checking if it reaches customers. An internal experiment may need less, even when it involves substantial work.

• Low: The agent's own drafts and experiments, which can be undone and are not yet used by others.
• Mid: Work that colleagues see or use and that can be undone.
• High: Work that is visible outside the company or hard to undo.

One agent does the work. A coordinator, also an agent, checks the result and prepares the summary. From mid level upwards, a separate checking agent first reviews the material itself. It must check the sources and result, even when the agent doing the work reports that everything is in order.

At high stakes, a person always approves before the result is used. At low and mid stakes, the person responsible in your company decides whether the task must wait for approval. This choice can be set once for a particular kind of recurring task and then applies each time. No agent can make that choice or approve the work.

Even at low stakes, the person responsible receives a short summary of the result, what was checked and anything missing. The employee does not need to follow every attempt along the way. Anything needing a decision is highlighted.

Agree access and spending limits upfront

Connections and access are set up together with the person responsible in your company. An agent works only in the files and systems it is given access to. Before setup, it is clarified whether your systems can provide the access needed. The task brief also states what the agent must not change, such as customer data it may only read.

Each task gets a clear objective, a defined scope and requirements for evidence. You also set a limit on money spent on AI, along with limits on time and attempts. When a limit is reached, work stops and goes to the person responsible. The summary shows what was spent on AI. Setup and running costs also belong in your assessment of the total cost.

A delivery you can check

What must be delivered depends on the task. The workflow distinguishes five types so you can see what a finished result needs to include.

• Answer: An answer with a source that can be checked.
• Explore: Different possible solutions and a recommendation. You make choices about taste and direction.
• Build: A change that works and an automatic check that can flag it if it later causes errors. In the invented kitchen firm example, the check could compare the prices in the quote with the price list.
• Check: An assessment with evidence for what was examined.
• Maintain: An update with a short note explaining what changed.

A larger task can combine several types in turn, for example Explore, then Build and finally Check. The coordinator compares the delivery with the source or the result itself. A report that a price has been corrected must be verifiable in the quote itself.

When work cannot be completed

If a source is missing, a check fails or the next step requires an action outside the task brief, the agent must not guess its way forward. If the problem cannot be solved within the agreed limits, work stops. The person responsible receives a note explaining the problem, what was attempted and what is needed to continue.

The summary distinguishes between completed and checked work, a departure from the brief that a person has accepted, giving a reason, and work waiting for help. An agent must not change the requirements to call a task finished. If an error is discovered later, the person responsible must be able to trace the sources and changes and decide on the correction.

Start with a task you already repeat

Choose a routine task whose result stays inside the company. The workflow can be set up using copies of your material so you can assess the result before use. Try both an ordinary task and one with missing information. This lets you see whether the work actually stops in the right place.

Assess whether staff have less to gather, check and correct when the whole task is taken into account. That way you can see whether the agents take over the routine work and whether staff get more room for the judgements that need their experience.