YoursSherpa
Contact us

SystemsSheet 03 of 08

Workflow automation and AI agents for your manual processes

We start by mapping how the work is done today, step by step. Steps with clear rules become ordinary automation. Steps that need reading, judgment or drafting go to an AI agent, and a person approves anything that matters. Each agent is customised to your process, your tools and your approval rules.

Starts with
Process mapping and gap analysis
Agents
Built for your process and tools
Runs in
Your cloud, Databricks, Microsoft 365, or private servers

Agent orchestration with human approval

Switch the environment to see the services it runs on. Play the walkthrough, or pick a step.

Deterministic codeAI stepHuman reviewYour existing systemAI assist, approved by a person
    Agent orchestration with human approval

    Work arrives, code routes it, an orchestrator delegates to specialist agents with scoped tools, and a person approves anything that matters.

    Step Functions and EventBridge for triggers, LangGraph on ECS for orchestration, Claude on Bedrock for the agents.

    See it run

    One invoice exception, end to end. Code handles the rules, agents handle reading and drafting, and a person approves before anything is sent.

    Agent run traceOrchestrator, three specialists, one approvalIllustrative example

    Run 4812: invoice exceptionCompleted

    Email received trigger0.1 s
    Classify request invoice exception1.1 s
    Extract invoice fields 14 fields2.4 s
    Schema check passed0.1 s
    Look up PO-77120 in ERP tool0.6 s
    Tolerance rule 6.2% over PO0.1 s
    Research vendor history 3 tool calls4.7 s
    Draft reply and exception note2.9 s
    Approval in Teams AP lead28.8 s
    Update ERP, notify vendor code0.8 s
    0 s12 swaiting, compressed41.6 s

    Step 8: Draft reply

    Input the drafting agent received, built from the research step:

    {
  "vendor": "Northwind Medical Supply",
  "po": "PO-77120",
  "variance_pct": 6.2,
  "history": "2 prior variances, both approved",
  "tone": "firm, polite"
}
    Tools used3erp.get_po, erp.list_invoices, email.search
    Approved byAP leadin Microsoft Teams, 28.8 s after request

    Every step, input, output and approval is kept in the run log.

    Component view

    The same system as an exploded 3D drawing. Each plate is one step, and each block is one component in one of four materials.

    Drag to rotate. Select a layer to inspect it.

    Parts list

    Numbered bottom to top

    1. Agents work inside the tools your team already uses. We connect through APIs or MCP servers, and use a controlled browser session only where no API exists.

      • Email and inbox Your existing system
      • CRM and ERP Your existing system
      • Spreadsheets Your existing system
      • Ticketing Your existing system

      Typical toolingMicrosoft 365, Google Workspace, Salesforce, HubSpot, SAP, Jira, ServiceNow

    2. Work starts from an event: a new email, a form, a file, a status change or a schedule. This part is plain code.

      • Schedules Deterministic code
      • Webhooks and events Deterministic code
      • Inbox watcher Deterministic code

      Typical toolingn8n, Power Automate, Airflow, EventBridge, Azure Logic Apps

    3. The agent reads unstructured input such as emails, PDFs and forms, and turns it into structured fields. Code checks the output against a schema before anything moves on.

      • Document reader AI step
      • Classifier AI step
      • Schema check Deterministic code

      Typical toolingLLMs with structured output, OCR, JSON Schema, Pydantic

    4. If a rule can be written down, it is written as code. Routing, thresholds and approval limits stay deterministic and testable.

      • Business rules Deterministic code
      • Router Deterministic code
      • Exceptions queue Deterministic code

      Typical toolingPython, SQL, decision tables, rules engines

    5. Agents handle what rules cannot: researching a case across systems, drafting a reply or report, reconciling records that do not match. They act only through tools we define, and each tool has its own permissions.

      • Research agent AI step
      • Drafting agent AI step
      • Scoped tool calls Deterministic code

      Typical toolingClaude, GPT, Gemini or open-weight models; LangGraph, OpenAI Agents SDK, MCP

    6. A person approves anything that sends, pays, deletes or commits. Approval happens where they already work, such as Teams, Slack or email, and every decision is logged.

      • Approver Human review
      • Audit log Deterministic code

      Typical toolingMicrosoft Teams, Slack, email approvals, audit tables

    7. Approved actions are written back to your systems by code. Each run is recorded, so you can see what the agent did, why, and who approved it.

      • System updates Deterministic code
      • Notifications Deterministic code
      • Run history Deterministic code

      Typical toolingSystem APIs, RPA where no API exists, run-history dashboards

    Deterministic codeAI stepHuman reviewYour existing system
    Sheet03 of 08
    Layers7
    SystemWorkflow automation and AI agents
    Drawn byA. Adhikari
    IssuedSeptember 2026
    ScaleNot to scale

    How we find where agents fit

    For each step in your process we ask the same three questions. The answers decide what gets built.

    Can the rule be written down, and does it stay the same?

    Deterministic automation. Cheaper, faster and testable.

    Does the step need reading unstructured text, judgment or drafting?

    AI agent, working only through tools we define.

    Is the action costly or hard to undo?

    A person approves before it happens.

    We also look for the gaps between systems: data typed from one tool into another, reports assembled by hand, and checks that depend on one person remembering to do them. These are usually the first things we automate.

    Agents we customise

    Each agent is built for one job in your process, with its own tools, permissions and reviewer.

    Intake and triage

    Reads incoming requests, pulls out the details, and routes each one to the right queue.

    Document extraction

    Turns invoices, forms, contracts and claims into validated, structured records.

    Research

    Gathers context on a customer, case or supplier from your systems and summarises it.

    Drafting

    Writes replies, reports and summaries in your format, for a person to review.

    Data quality

    Explains failed checks, suggests fixes, and opens tickets with the evidence attached.

    Reporting

    Builds recurring reports from your data and points out what changed since last time.

    Where it runs

    The drawing stays the same. The services change with the environment you already run, and the feasibility analysis picks the fit.

    Environment
    What we use
    When it fits

    Your cloud: AWS, Azure or Google Cloud

    What we useContainers or serverless functions, managed models behind private endpoints
    When it fitsMost common. Data and logs stay in your account.

    Databricks

    Built before
    What we useDatabricks Apps, Mosaic AI Agent Framework, Model Serving
    When it fitsProcesses that run on data already in Databricks.

    Microsoft 365

    What we usePower Automate, Copilot Studio, Azure OpenAI
    When it fitsProcesses centred on Outlook, Teams and SharePoint.

    Workflow tools

    What we usen8n, Make or Zapier with LLM steps
    When it fitsSmaller processes and quick pilots.

    Private or on-premises

    What we useOpen-weight models with vLLM, self-hosted n8n
    When it fitsProcesses that cannot use cloud AI.

    Feasibility first

    Before anything is built, we check whether this system is worth building for you, and where it should run.

    • How the process runs today, observed with the people who do it.
    • Volume, variation, and the cost of an error at each step.
    • Which systems have APIs, and which need another route.
    • Which steps can be deterministic, which need an agent, and which must stay with a person.
    • How success will be measured against the current process.
    • Security, data access and approval requirements.
    What you receiveA process map with every step marked as code, agent or person, a pilot scope, and a go or no-go recommendation.

    Built before

    Work delivered by Ashish Adhikari, who leads engineering at YoursSherpa.

    • Validation frameworks and automated quality checks for healthcare pipelines handling 500K+ claims a day.Cedar Gate Technologies
    • Central AI agent hub: one MCP endpoint that gives every AI tool on a team the same approved playbooks.Agent infrastructure
    • Assessment of where machine learning fits into application workflows.Clevero

    Start with a feasibility call

    Tell us about one process or data problem. In the first call we will say which parts we would automate with code, which need an agent, and which we would leave alone.

    Send a short note through the contact form and we will set up the call.

    Contact us See Ashish Adhikari's profile

    What helps us prepare

    • The process or system you have in mind, and who works on it today.
    • Where the data lives: cloud, platform and main tools.
    • Security or hosting rules we need to work within.