Skip to content
PI Square
Book

Agentic workflows

Agents that carry a piece of work from request to done, inside the tools your people already use.

memorygoalonboard a supplieryoucheck documentsverify bankdraft recordcreate in erporchestratorcheckdatadraftactionreadresearchdocsreadrecordsreaderpwriteticketswritetools · mcpnew vendor30-day termsa person approvesaudit
  1. 1Goal
  2. 2Plan
  3. 3Agents read
  4. 4Ask back
  5. 5Draft
  6. 6A person approves
  7. 7Write and remember
Try it

An agentic workflow is a set of AI agents that each do one part of a process: check a request, ask for what is missing, decide where it goes and update the systems involved. PI Square builds them in your cloud, gives each agent only the tools its job needs, and puts a person in front of every decision that carries risk.

  1. A request arrives

    a form, an email or a chat message

  2. Your cloud

    Agents check and ask

    each with one job

  3. Approved tools only

    read, write, file, notify

  4. A person approves

    before anything changes

  5. Your systems, updated

    Jira, SAP, Google Workspace, Microsoft 365

What it's for

  • Retail

    Employee ideas, store requests or supplier queries collected, completed and routed without anyone sorting them by hand.

  • Financial services

    Applications checked for missing documents, and the customer asked for them before a person reviews the file.

  • Procurement

    Purchase requests matched to policy and budget, with exceptions sent to the right approver.

  • Shared services

    IT and HR requests answered, or passed on with the context already gathered.

How we keep it safe

  • Every agent has a written job and a short list of tools, and it cannot reach anything else.
  • Each agent's output is locked to a schema, so the next step gets exactly the fields it expects.
  • Duplicate and retry guards stop the same item being processed twice.
  • Hard limits on cost and rate, for each agent and each day.

Built with

  • Google ADK
  • Vertex AI Agent Engine
  • Model Context Protocol
  • Gemini
  • Claude
  • Python
  • Pub/Sub
  • Firestore

Where we've built this

  • A southern African retail group: five agents that take an employee's idea, ask them follow-up questions in Google Chat, score it, open it in Jira and draft the product brief. Open to about 500 employees.
  • A regulated consumer AI product: a gateway that decides which tools a model may call, and records every call.

How an engagement runs

  1. Understand

    We map one process step by step and mark where an agent can take work off someone's desk.

  2. Prove

    One workflow runs on live requests, with a person approving each outcome.

  3. Embed

    The workflow has a clear owner, review rules and an operating guide. Further steps move to agents only when the evidence supports it.

The full approach

Questions about agentic workflows

What is an agentic workflow?
A process run by AI agents that each handle one step: checking, asking, deciding or updating a system. The agents pass the work along the way a team would, and a person approves the steps that carry risk.
How is an AI agent different from a chatbot?
A chatbot replies to a person. An agent also acts on the reply: it can file a ticket, update a record or route a request. Because agents act, every agent PI Square builds has a fixed list of tools it may use, and a person approves anything with risk.
Can agents work with SAP, Jira or Microsoft 365?
Yes. Agents reach your systems through their own APIs, using an account with only the permissions the job needs. We have connected agents to Jira, Google Chat, Google Drive and SAP, and connect Microsoft 365 the same way.
How do you prevent AI agents from taking unauthorized actions?
We define permitted tools, least-privilege access and approval points around the risks of each workflow. Actions such as changing records or contacting customers can require human review. Retry controls, duplicate-action checks and spending limits are configured and tested as part of the agreed scope.
Which protocols and frameworks do you use to connect agents?
Options include the Model Context Protocol (MCP) for tool connections, Google ADK as an agent framework and Vertex AI Agent Engine as a managed runtime. We choose integrations around your existing systems and agree portability requirements, including any dependencies on a cloud provider.

The other five

Tell us what you're working on

Keep this general; leave out sensitive or confidential information.

Or email nikhil@pisquare.ai