AI agents can now do real work inside the systems that run a business. They match invoices in the ERP, route cases in the CRM and answer employees from company knowledge. The models are capable, and every major platform ships agent features.

Getting from a promising demo to an agent that does that work every day is a different job. Gartner’s survey of more than 2,000 CIOs, published in October 2026, found that only 13% say their enterprise has realized significant or transformational value from AI. In our experience, the gap is rarely the model. It is the work around it: the process, the access, the testing and who keeps the agent running.

This plan covers that work in six steps.

In brief

  • Start with one workflow that has volume, clear rules and a result you can measure.
  • Define the process before you automate it. AI can’t learn a process that only lives in people’s heads.
  • Set access, approvals and logging before you build, not after the first incident.
  • Test before launch and keep testing after, because models and systems keep changing.
  • Decide who builds the agent and who runs it, and make sure your team can change it later.
  • Measure results and keep improving once the agent is live.

1. Pick one workflow with volume, rules and a clear result

The best first agent handles work that happens often, follows rules most of the time and has a result you can count. Good candidates include invoice matching, case routing, order exceptions and employee requests.

Avoid starting with the hardest or most visible problem. A narrow first agent that works builds trust, gives you real numbers and teaches your team how agents behave in your systems.

Write down what success looks like before anyone builds anything: hours saved, cycle time, error rate or share of items handled without a person touching them.

2. Define the process before you automate it

This is the step most often skipped, and the one that decides whether the rest works.

A common pitch is that AI will learn how your team has handled the work over time and then make the same decisions on its own. That only works if a consistent process exists to learn from. In many companies it doesn’t. The process lives in a few people’s heads, and it changes when they leave.

We saw this on an accounts payable automation project. The client’s AP team processed four different types of invoices, each handled differently. Nothing in their ERP identified the invoice type or the vendor type. The team simply knew each vendor and remembered how to process its invoices, and that knowledge passed from person to person when someone left.

No AI could have sorted that out on its own. So we started with the process. We defined how each invoice type should flow, set up the ERP to record invoice and vendor types, and built a proper intake so invoices arrived in a form that supports 2-way and 3-way matching. Only then did we add AI to handle the parts that need judgment.

Before you build an agent, make sure you can answer three questions:

  • What are the steps, and who decides at each one?
  • What data tells the agent which path to take, and is that data recorded in your systems today?
  • What are the exceptions, and where should they go?

If the answers live only in people’s heads, capture them first. The agent will be faster to build and far more reliable.

3. Set access, approvals and logging before you build

An agent works through the same systems your people use, so it needs the same care with access. Gartner advised infrastructure leaders in November 2026 to inventory their AI agents and “classify them all as nonhuman identities.”

In practice:

  • Give each agent its own identity and role in each system, never a shared admin account.
  • Grant the narrowest access that works, per system and per action.
  • Decide which actions need approval, such as payments or changes to customer records.
  • Log every action so you can see what the agent received, what it changed and what it returned.
  • Keep a way to stop the agent instantly, and test it before launch.

Platforms such as NetSuite, Salesforce and Workato provide controls for their own agents. Most companies run several of them, so someone has to design the roles and keep one view across all of them. Our upcoming guide to governing AI agents covers this in more depth, and our AI governance services map the controls to frameworks such as the NIST AI Risk Management Framework.

4. Test before launch and keep testing after

Agents change behavior when models, prompts, data or connected systems change. A test that passed in March can fail in June without anyone touching the agent.

Build a set of realistic test cases from your own process, including the exceptions from step 2. Run them before launch, then on a schedule after it, with a clear rule for what happens when a test fails. Our Ally Platform includes Test Lab for this reason: it runs simulated conversations against an agent before launch and keeps testing it after.

5. Decide who builds the agent and who runs it

Your agents can be built by your own team, your software vendor or an implementation partner. Whichever you choose, plan for the day after the build. Your company should own what is built, including code, prompts, configuration and documentation, and your team should understand how it works as it is built. Before the build ends, check that your team, or a partner you choose, can make a real change on its own.

Then decide who runs the agent after launch: who monitors it, reviews its access, handles model updates and fixes issues at night. In the 2026 GenAI partner survey by BCG and AWS, “insufficient ongoing support” was one of the top complaints once AI reached production. Agree on support hours and response times before go-live.

6. Measure results and keep improving

Track the measures you set in step 1 from the first day. Review them monthly with the process owners, look at the exceptions the agent sends to people and decide what to improve next.

Once the first agent is stable, the second one is much faster. The process work, the access model, the test approach and the support routine are already in place.

How Boston SoftDesign implements AI agents

Boston SoftDesign has built and supported system integrations since 1999, with NetSuite, Salesforce and Workato at the core. Our AI implementation services follow this plan: we start with the process, connect agents to the ERP and CRM systems where the work happens, and build governance into the first design.

Clients own what we build, including code, prompts, configuration and documentation. Knowledge transfer is part of every engagement, so your team can see how each agent works and why. After launch, our managed support, including 24/7 coverage, keeps agents and the integrations around them running.

Learn more about our AI consulting services, AI automation services and AI agents for business, or talk to our team. If you are still deciding where AI fits, our AI readiness audit guide is a good place to start.

Frequently asked questions

What is AI agent implementation? AI agent implementation is the work of putting an AI agent into production inside your business systems. It covers choosing the workflow, defining the process, connecting systems, setting access and approvals, testing, launching and supporting the agent after launch.

How long does it take to implement an AI agent? It depends on the workflow and on how well the process is defined. A narrow first agent on a well-defined process moves much faster than one where the process has to be designed first. Defining the process up front usually saves time overall.

Why do AI agent projects fail? Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. In our experience, a missing process, unclear success measures and no plan for running the agent after launch are common causes.

Can AI learn our process from past data? Only if a consistent process exists in that data. If the same kind of work was handled differently depending on who did it, AI will learn the inconsistency. Define the process and record the data the agent needs first, then add AI.

Who should run AI agents after they launch? Someone has to monitor the agents, review their access and fix issues when models or systems change. That can be your own team or a managed service partner, but it should be agreed before launch, with defined hours and response times.

Sources

  • Gartner, survey finds only 13% of CIOs say their enterprise has realized significant or transformational value from AI (October 2026): gartner.com
  • Gartner, weak AI governance will lead to business-critical service disruptions (November 2026): gartner.com
  • Gartner, over 40% of agentic AI projects will be canceled by end of 2027 (June 2025): gartner.com
  • BCG, How CTOs can choose the right GenAI partners at the right time (June 2026): bcg.com