Groundwork
All articles

AI in product management

AI agents and MCP for product managers: connecting AI to your product tools

Updated

Most product managers use AI by copying things in and out of a chat window: paste the feedback, paste the spec, copy the answer back. AI agents connected to your tools remove that step. They can read your tickets, research notes and analytics directly, and act on them. The Model Context Protocol (MCP) is the common way those connections are made. This article explains what changes, where it helps, and what to set up so an agent does not cause more work than it saves.

What MCP is, in one paragraph

MCP is an open standard, introduced by Anthropic in 2024, for connecting AI assistants to other software. A tool such as an issue tracker, a whiteboard or a research repository offers an MCP server. An AI assistant that supports MCP, such as Claude, ChatGPT or an AI coding tool, can then use it to look things up and take actions in that tool. Build the connection once and any compatible assistant can use it.

What an agent adds

A chat assistant answers a question. An agent works towards a goal over several steps: it looks something up, decides what to do next, uses another tool and checks the result. With tools connected, a product manager can hand over whole tasks instead of single prompts.

  • Feedback triage. Read this week’s support tickets, group them by problem, link each to existing insights and flag anything new.
  • Evidence lookup. “What do we know about how nurses prescribe infusions?” answered from your actual research, with links, instead of from the model’s general knowledge.
  • Spec to tickets. Turn an approved requirement into tracker tickets with acceptance criteria and a link back to the evidence.
  • Release notes and updates. Draft release notes from the tickets that shipped, and a stakeholder update from what changed.
  • Keeping things in sync. Update a roadmap item when its experiment finishes, or flag requirements whose evidence has changed.

Before you let an agent write

Reading is low-risk. Writing is where agents create mess: duplicate tickets, overwritten notes, a requirement approved by the wrong person. Settle these before you switch on write access:

  1. Least privilege. Give the agent only the tools and permissions the task needs. Triage needs to read tickets and create insights, not delete projects.
  2. People approve the decisions. Let the agent draft and propose, but keep approvals, sign-offs and anything sent to customers with a named person.
  3. Search before create. Instruct the agent to look for existing items before adding new ones, and to link to them instead of duplicating.
  4. Everything is attributable. You should be able to see what the agent changed, when and on whose behalf, and undo it.
  5. Watch for instructions hidden in content. An agent that reads customer emails or web pages can be steered by text written to manipulate it, known as prompt injection. Treat what it reads as data, and keep irreversible actions behind human review.

A first agent workflow to try

Goal:   Turn last week's interview notes into insights.

1. Read the notes in /research/2026-11.
2. For each interview, propose up to 5 quotes, verbatim, with the
   persona and source. Leave out names and contact details.
3. Search existing insights first. Link instead of duplicating.
4. Show me the proposed insights. Create nothing until I approve.
Read, propose and wait for approval. Widen the agent's freedom once you trust its output.

Where agents still fall short

  • They are confident when wrong. An agent will report success on a task it half finished. Check the result, not the summary.
  • They lack unwritten context. The reason a decision was made often lives in someone’s head. Agents work best where the evidence is written down and linked.
  • They multiply bad structure. If your backlog is a mess, an agent will produce more of it, faster. Tidy the structure first.

Agents are most useful when your product knowledge is structured and linked: insights tied to hypotheses, hypotheses to experiments, experiments to requirements. That structure is what lets an agent answer “why are we building this?” with evidence. See requirements traceability.

Sources

Keep your evidence and decisions linked

Groundwork takes your team from customer insights to evidence-based requirements. Free to start, with your whole team.

Get started