Latest
Deal Heidi cuts headcount by 18% and co‑founders leave after $475 million raiseDeal What buyers look for when they acquire an AI companyGlobal Agent or hire: how to decide whether a job should grow through people or systemsGlobal How to put an AI customer service agent to workGlobal How to put an AI agent on accounts payableGlobal How to use an AI sales assistant without annoying your customersGlobal What an owner can hand to AI agents first
United States edition
Friday, 6 November 2026
Infercline.
Advertise
Operators Building with Agents

The first 90 days of an AI agent rollout

Most agent projects that fail do so in the first three months, for reasons that have little to do with the technology. A month-by-month plan for getting it right.

The Infercline Desk · Friday, 6 November 2026 Sign in to save
black asphalt road during night time
Photo by Billy Freeman on Unsplash

The first three months of an AI agent rollout decide whether it becomes part of how your business works or a forgotten trial. Gartner predicted in June 2025 that over 40 per cent of agentic AI projects would be cancelled by the end of 2027, pointing to escalating costs, unclear business value and inadequate risk controls. All three can be addressed in the first 90 days. This plan breaks the period into three months, each with a clear goal.

Before day one, settle three things. First, the job: one specific, frequent task, such as answering order status questions or entering supplier invoices, not a broad ambition. Second, the owner: one named person accountable for the agent's results, with the authority to change or stop it. Third, the baseline: how many cases the job involves each week, how long each takes, the error rate and the cost. Without a baseline you will not be able to show value at day 90.

Month one is for setting up safely. Map the current process step by step with the people who do it, including the exceptions. Choose the tool, starting with any agent your existing software already offers. Give the agent the least access its job needs: read rather than write where possible, no ability to pay, delete or send to customers without approval. Decide in writing which actions need a person's approval. Check the data protection rules in each country whose customer data the agent will see. Then test the agent on past cases, with the people who do the job today grading its work.

Anthropic's guidance on building agents is useful here: start with the simplest approach that works and only add complexity when it clearly improves results. If a fixed workflow with one AI step does the job, use that. You can give the agent more freedom later, once it has earned it.

Month two is for a controlled launch. Switch the agent on for a small share of real work: one channel, one team, one region or a fixed share of cases. Keep a person approving its actions at first. Read a sample of its work every day. Log every mistake, what caused it and how it was fixed. Expect to change the agent's instructions, your help pages or your process several times. This is normal, and it is the most valuable learning of the whole rollout.

Bring the team in early. People whose work the agent touches will have questions about their jobs. Be honest about what the agent will and will not do, and involve them in grading its work. They know the exceptions better than anyone, and their trust will decide whether the agent is used or quietly worked around.

Month three is for measuring and deciding. Compare the agent's results against the baseline: volume handled, time per case, error rate, cost per case including checking time, and customer or staff feedback. Where the agent has proved reliable on a type of case, consider letting it act without approval and keep approval for the rest. Then make a clear decision: expand, adjust or stop. Stopping a trial that is not working is a success of the process, not a failure.

Frameworks can help structure this work without slowing it. The US National Institute of Standards and Technology published its AI Risk Management Framework in January 2023. ISO/IEC 42001, published in December 2023, sets requirements for an AI management system. Singapore's Model AI Governance Framework for Agentic AI, launched in January 2026, recommends bounding an agent's risks upfront, keeping humans meaningfully accountable and applying technical controls across its life. Borrow the parts that fit the size of your business.

McKinsey's 2026 State of AI survey found that about two in ten respondents had scaled AI agents across their organisation. The ones that get there usually treat the first agent as a way to build the habits, controls and confidence for the second and third.

Why it matters to your business

A disciplined first 90 days turns an agent from an experiment into a measured part of operations. It limits the damage if something goes wrong, gives you evidence to justify further spending and brings your team with you. Skipping the baseline, the access limits or the measurement is how promising trials end up among the cancelled projects.

What to do on Monday

  1. Write a one-page brief naming the job, the owner, the success measures and the date of the day-90 decision.
  2. Record this week's baseline figures for the job: volume, time per case, error rate and cost.
  3. List what the agent will be able to access and which actions will need a person's approval, and have the owner sign it off.
  4. Book a weekly 30-minute review for the next 12 weeks with the owner and the people who do the job today.

More on this topic: AI agent guides

Sources