Insights · November 2026 · 8 min

A buyer's guide to rolling out AI agents

How to choose the first function, define what success means, and write a brief a vendor can act on.

Short answer: To roll out AI agents, choose one function with high routine volume and measurable outcomes, define the outcome metrics that reward the agent, confirm data and system access, set governance and approval limits, run a measured pilot, and hand over in stages. A vendor brief should state the function, maximum budget, timeline and a clear description.

Most companies that try AI agents start with a tool and look for a place to use it. I would suggest the reverse. Start with a part of the business you understand well, define what good looks like, and only then decide what to build or buy. This guide follows that order. It is written for an executive who has to make the decision and live with the result.

Which business function should you automate first?

The first function sets the tone for everything after it, so it is worth choosing carefully. I look for five qualities:

Customer service often fits all five, which is why it is a common starting point. I describe it in detail in AI agents for customer service. Marketing and supply chain can also work well, though supply chain tends to need cleaner data across more systems.

I would avoid starting with a function that is politically sensitive, poorly measured or in the middle of another major change. The first deployment should be one where success is easy to see and agree on.

How do you define success for an AI agent?

This is the most important decision in the whole project, and the one most often rushed. An agent that learns from outcomes will optimize whatever it is rewarded on, so the reward has to reflect what the business actually wants.

A good reward has three properties. It measures an outcome rather than an activity, so resolved and stayed resolved, not tickets closed. It reaches as far toward business value as the data allows, so revenue from leads, not just lead volume. It includes the costs and risks the business cares about, such as refunds paid, policy breaches and complaints, so the agent cannot score well by cutting corners.

Write the reward down as a short list of metrics with targets and limits. For example, a support agent might be rewarded on confirmed resolution and satisfaction, penalized for reopens and policy breaches, and held to a ceiling on refund value per month. The reasoning behind outcome rewards is set out in reinforcement learning for business agents.

Is your data ready for AI agents?

An agent can only act on what it can see and reach. Before signing anything, check four things:

  1. Access. Can the agent read and write in the systems the function uses, such as the CRM, help desk, ERP or ad platforms, with permissions limited to what the job needs?
  2. History. Is there a record of past cases and their outcomes the agent can learn from and be tested against?
  3. Policy. Are the rules of the function written down, including limits on money and the cases that must go to a person?
  4. Outcome data. Will the company be able to see the results of the agent's actions, such as renewals, reorders or returns, and feed them back?

Gaps here are common and usually fixable. They should be found before the pilot, since they tend to set the timeline more than the agent itself.

What governance do AI agents need?

Governance means deciding in advance who owns the agent, what it may do and how anyone will know if it goes wrong. At minimum I would set:

Companies in regulated fields will also need legal and compliance review of data use, customer disclosure and record keeping. My background in clinical informatics makes me cautious here. The rules are easier to set before launch than after an incident.

How long does it take to roll out AI agents?

Timelines depend mostly on data readiness and on how quickly the company can make decisions about policy. As a rough illustration, a well-prepared company might move through these stages over a few months:

  1. Discovery. Map the function, agree the baseline cost and metrics, and write down the policy.
  2. Shadow mode. The agent processes real cases alongside the team without acting, and its choices are compared with what people did.
  3. Limited pilot. The agent acts on one channel or category with low approval limits and heavy review.
  4. Staged handover. Scope and limits widen as metrics hold, while the team shifts to oversight roles.
  5. Steady state. The agent runs the function, a smaller team oversees it, and the reward and policy are reviewed on a schedule.

I would be wary of anyone who promises a full handover in weeks, and equally wary of a pilot that runs for a year without a decision. Set a date at the start for deciding whether to proceed, and the metrics that decision will rest on.

Should you build or buy AI agents?

Building makes sense when the function is central to how the company competes, when there is an engineering team with experience in models, tools and evaluation, and when the company is prepared to maintain the system for years. Buying makes sense when the function is important but not unique, when speed matters, or when the company would rather pay for a result than run a new kind of software team.

There is a middle path. Some companies buy a platform and keep ownership of the policy, the reward and the data. Whichever route you choose, keep those three things in your own hands. They encode how your business works, and they should outlast any vendor.

What should you put in a brief to an AI agent vendor?

A short, clear brief saves weeks. I would include four items, and keep each one specific:

A good vendor will come back with questions about your data, your policy and your metrics. A vendor who comes back only with a demo has probably not read the brief. The reasons to hand over a whole function rather than one task are set out in automate a function, not a task.

Questions people ask

What is the best first function to automate with AI agents?

Choose one with high routine volume, measurable outcomes, writable policy, contained risk and accessible systems. Customer service often meets all five, though marketing and supply chain can also work.

How long does it take to deploy AI agents in a business?

It depends mostly on data readiness and policy decisions. A well-prepared company might move from discovery through shadow mode and pilot to staged handover over a few months.

Should I build or buy AI agents?

Build if the function is central to how you compete and you have an experienced engineering team to maintain it. Buy if speed matters or the function is not unique. In either case, keep ownership of your policy, reward metrics and data.

What should I include in a brief to an AI agent vendor?

State the function, the maximum annual budget, the timeline for pilot and handover, and a description of current work, systems, desired outcomes and hard limits.

How do you govern AI agents safely?

Name an accountable business owner, set written approval thresholds enforced by the system, log every action, review sampled work and escalations regularly, and keep a fast way to pause the agent.

Keep reading