Track A / 5 volumes

Working with agents

A public track for the people who will live with the workflow. By the end of Volume 1, a reader should be able to write one outcome, mark the gates, and name the approver. Later volumes cover data and privacy in plain language, the four kinds of work FusionWork is built for, the weekly rhythm of approvals and corrections, and a careful rollout.

Sample available Linked samples are ready to read. Other modules show planned outcomes and are marked “Coming soon”.

Volume 1

Agents, alongside your team

Sample lesson available · remaining lessons coming soon

Brief the work and inspect its controls. Interfaces overlap: a chatbot can call tools, a copilot can act, and an agent can perform one-off work. Judge what can execute and who controls it, not the label. Complete the minimum preflight above before any trial.

A1.01Read the lesson

What an agent is

Completed artifact: One-sentence workflow explanation for a chosen operations, sales, support or finance example. Outcome: You can explain, in one sentence, what an agent does in your team — and how that is different from a chatbot.

  • After the agent has run, the work has moved, not just the wording.
  • Outcome, steps, a human gate, and a record — that is the whole pattern.
A1.02Coming soon

Chatbots, copilots, and agents

Completed artifact: A comparison of who selects steps and what can execute in three interfaces. Outcome: You can tell which of the three you already have, and which one you still need.

  • A chatbot is a conversational interface; a copilot usually keeps a person directing the next step. Either may have tools.
  • An agent selects steps within delegated limits, for a recurring or one-off job.
  • Ask who picks the next step. Rules-based automation follows a path someone fixed in advance. With a copilot, the person picks. An agent picks within limits you set. One workflow can use all three.
  • Agents work from likelihood, not fixed rules, so the same request can come back worded differently. That is why the result needs a check and the log needs a record.
A1.03Coming soon

The handoff: outcome, steps, gates, record

Completed artifact: One-page four-part handoff. Outcome: You can sketch any recurring task as those four parts, on one page.

  • If a part is missing, the handoff is not ready.
  • This sketch still works when the underlying model or app changes.
A1.04Coming soon

Work that is ready for an agent

Completed artifact: Readiness grid with one missing prerequisite and its owner. Outcome: You can pick a task that is repetitive, bounded, and already living in your tools.

  • Look for work that follows a path most days, with exceptions you can name.
  • If nobody can describe the outcome in plain language, start there, not with the agent.
  • Run the four readiness checks: a clear boundary, data you can rely on, a result you can check, and controls in place. If one is missing, fix it first.
A1.05Coming soon

Work that should stay with people

Completed artifact: A marked list of decisions that stay with people. Outcome: You can name the judgements, relationships, and commitments an agent must not finish alone.

  • Sending, paying, and deleting are the obvious stops — not the only ones.
  • A person still owns the customer, the supplier, and the exception that feels off.
  • When a task needs judgement and a mistake would be costly, the agent prepares the evidence and a person decides.
A1.06Coming soon

The approval moment

Completed artifact: Approval card with exact content, named approver, backup and timeout. Outcome: You can design a gate a busy person can actually use.

  • The approver sees the exact action, not a vague summary after the fact.
  • A shared inbox is not a named person; a backup must be named too.
A1.07Coming soon

The action log

Completed artifact: A review of one intent/outcome pair and its retention limit. Outcome: You know what a useful log contains, and why it is part of the workflow.

  • What it saw, what it did, who approved it — in language a manager can read.
  • The log holds personal data, so it falls under the same PDPA duties as the work itself.
A1.08Coming soon

Your first handoff

Completed artifact: Completed handoff, preflight and fictional rehearsal result. Outcome: You can complete one four-part handoff and its access/privacy/stop preflight, then rehearse it on fictional data. A live trial requires the owner-approved preflight and approval route.

  • One job, one set of tools, one named approver.
  • Write the outcome as if you were briefing a careful new colleague.

For the A1 handoff, choose a missing delivery document, a CRM follow-up, a support escalation or invoice matching. The full A1 sample applies the same four parts to each and includes a Singapore/Bangkok bilingual review card with THB and SGD kept distinct.

Volume 2

Your tools, your data, your rules

Coming soon

How the agent sits inside the tools you already use, what it is allowed to see, and how personal data should be treated in Singapore and Thailand.

A2.01Coming soon

Work stays in the tools you already use

Outcome: You can explain why the workflow should not start by moving everyone to a new system.

  • The agent comes to your inbox, documents, chat, CRM, and finance tools.
  • A new place to "put the AI" is usually a new place for work to get stuck.
A2.02Coming soon

What the agent is allowed to see

Outcome: You can limit access to the mailboxes, folders, and records the job actually needs.

  • "Give it everything, just in case" is how quiet leaks begin.
  • If a folder is not required for the outcome, it stays out of reach.
A2.03Coming soon

Personal data under Singapore's PDPA

Outcome: You can spot personal data in a workflow and ask the purpose, access, and retention questions.

  • Names, emails, invoice contacts, and staff details are personal data, even in a shared mailbox.
  • An agent does not reduce the company's duty; it makes the path through the data more important to map.
A2.04Coming soon

When Thailand's PDPA also applies

Outcome: You can tell when a Singapore team also needs to think about Thailand's PDPA.

  • It can apply when the work includes personal data of people in Thailand — staff, customers, or suppliers.
  • Flag it early, keep the data minimised, and do not treat this module as legal advice.
A2.05Coming soon

The knowledge the agent should use

Outcome: You can point the agent at the tone, policies, and documents it should draft from.

  • A knowledge base, a price list, or last month's good replies will beat a generic draft.
  • If the document is out of date, the agent will be confidently wrong.
A2.06Coming soon

Messy cases and exceptions

Outcome: You can describe what the agent should do when the file is incomplete or the request looks unusual.

  • The path for a missing delivery note should be as clear as the happy path.
  • Unusual cases get escalated; they do not get forced through.
A2.07Coming soon

Feedback that improves the next run

Outcome: You can correct a draft in a way that changes the next one, not only today's.

  • A short, specific correction is more useful than a general complaint.
  • If the same correction is made three times, the brief or the source document needs to change.
A2.08Coming soon

Who owns the workflow

Outcome: You can name a business owner, an approver, and (where you have one) a technical counterpart.

  • Agents do not own work; people do.
  • When something fails, it should be obvious who is called.

Volume 3

Four kinds of work

Coming soon

FusionWork is built for the work that slows small and mid-sized teams down: operations, sales, support, and finance. This volume helps you pick one workflow and brief it well.

A3.01Coming soon

Operations: forms, records, and chasing

Outcome: You can describe an operations handoff that keeps records in sync and follows up on what is missing.

  • Forms and documents are a good first job when the path is already known.
  • Chasing a missing field is work; inventing a policy is not.
A3.02Coming soon

Sales: research, follow-ups, and a clean CRM

Outcome: You can brief an agent to prepare the ground for a salesperson, without speaking for them.

  • Research and a draft follow-up can be prepared; the send still needs a person.
  • A CRM that is only updated "when someone has time" is a workflow, not a personality trait.
A3.03Coming soon

Support: triage, drafts, and escalation

Outcome: You can set a first-response workflow that routes, drafts, and stops before it speaks for the company.

  • Fast triage is useful; a wrong answer sent at speed is not.
  • The tricky cases should reach a person with the context attached.
  • A good split: the agent handles the routine questions and drafts the rest; people handle exceptions, check quality, and own the relationship.
A3.04Coming soon

Finance: matching, anomalies, and the month-end pack

Outcome: You can brief matching and exception-flagging without handing over the authority to pay.

  • Matching invoices to purchase orders is repetitive and checkable.
  • Releasing a payment, or explaining an anomaly to a director, stays with finance.
  • From order to cash, the agent can watch due dates and draft the chasers. A disputed invoice or a payment-plan negotiation stays with a person.
A3.05Coming soon

How to pick the first workflow

Outcome: You can choose one job that is frequent, annoying, and easy to judge.

  • Prefer a weekly burden with a clear "done" over a rare, political process.
  • If you cannot tell a good run from a bad one, pick a different job.
  • Place each step on the certainty × consequence grid. A first workflow should sit mostly in "run it automatically" and "bounded autonomy", with the costly steps gated.

Practice note: Two short, bounded pilots beat one broad programme. Pick a job you can already describe, and measure time to first useful result — how quickly a real user sees a useful result — not completeness.

A3.06Coming soon

Writing a brief the agent can follow

Outcome: You can write an outcome that names the tools, the frequency, and what good looks like.

  • Write it as instructions to a careful colleague, not as a slogan.
  • Include the exceptions you already handle by habit.
A3.07Coming soon

What "good" looks like in a week

Outcome: You can set a small number of checks for a quiet trial week.

  • Time returned to the team, errors caught, and approvals made on time are enough to start.
  • Vanity counts — "messages processed" — are not a measure of trust.
  • Measure three levels: did the agent do the task well, did the whole workflow get faster or cleaner, and did the business notice? Track one guardrail too, such as complaints or errors that reach a customer.

Practice note: Agree the measure before the trial starts. If the underlying records are messy, the agent will miss the target no matter how quickly you build.

A3.08Coming soon

A short list, not a catalogue

Outcome: You can keep a short, honest list of workflows instead of a wish list.

  • One working handoff beats six half-described ideas.
  • Add the next job only when the first one is boring in a good way.

Volume 4

The weekly rhythm

Coming soon

Once a workflow is live, the work is in the approvals, the exceptions, and the corrections. This volume is about running that loop without slowing the team down.

A4.01Coming soon

Approval rules that don't slow the team down

Outcome: You can set gates on the steps that commit the company, and leave the rest to run.

  • If everything needs approval, you have rebuilt the old inbox.
  • If nothing needs approval, you have given up the control you wanted.
  • Watch for rubber stamps. If every request is approved in a few seconds, the gate has stopped working. Look at a few approvals each week, and check what was approved, not only that someone clicked.
A4.02Coming soon

Named people, backups, and time off

Outcome: You can name a primary approver and a backup who can cover a Thursday public holiday.

  • Approvals should follow a person, not a generic "finance team" address.
  • Plan for leave before the agent is stuck at 5pm on the eve of a long weekend.
A4.03Coming soon

When the agent is stuck

Outcome: You know what "blocked" should look like, and who is notified.

  • A blocked agent should say what it is waiting on, not fail quietly.
  • The owner decides whether to supply the missing file, change the brief, or take the step by hand.
A4.04Coming soon

Reading the log without reading every line

Outcome: You can review a week's work by exceptions, gates, and odd patterns.

  • Start with failed matches, overdue approvals, and anything sent or paid.
  • A log nobody reads is only a comfort until something goes wrong.
A4.05Coming soon

Correcting an agent — and making it stick

Outcome: You can turn a correction into a change in the brief, the source documents, or the gate.

  • Fixing one draft by hand teaches you, not the workflow, unless you record why.
  • Repeated corrections are a design smell.
A4.06Coming soon

Adding a second workflow

Outcome: You can decide when the first handoff is stable enough to add another.

  • Stability means the gates are routine and the exceptions are familiar.
  • Reuse the same pattern; do not invent a new one for every team.
  • A simple ladder: the agent assists one step; then it shares the work with people; then you redesign the workflow around it; then it improves within limits, with a person reviewing every change. Climb one rung at a time.
A4.07Coming soon

Working as a team with agents in the loop

Outcome: You can describe how people and agents share a queue without stepping on each other.

  • Everyone should be able to see what is already being handled.
  • The team stays accountable for the customer or the supplier, even when the draft was prepared for them.
A4.08Coming soon

A simple weekly review

Outcome: You can run a fifteen-minute review that a busy lead will actually attend.

  • What ran, what waited, what was corrected, what should change.
  • Write one decision, not a report.

Practice note: A standing weekly write-up that reconciles the week's work into one account is more useful than a chronicle of meetings. Organise it by goals and results. Keep a short reflection on one problem and a decision, then check the following week whether it closed. An agent can help draft from the record; the person who owns the week still edits and sends it.

Volume 5

Trust, rollout, and what lasts

Coming soon

How to start carefully, ask good questions of any vendor (including us), and keep the pattern when the tools move on.

A5.01Coming soon

Questions worth asking any vendor

Outcome: You can ask where the agent acts, who it acts as, what it can do without approval, and where the log lives.

  • "It uses AI" is not an answer.
  • You are looking for gates, least privilege, and a record you can export.
  • Also ask: who answers for the result, how errors are found, and how you roll back or pause in an emergency.
  • Use IMDA’s Model AI Governance Framework for Agentic AI as a governance reference alongside the regional briefing above. Exact edition, URL, sections and attributed recommendations are pending verification. Academy controls are teaching policy, not certification or a legal-compliance claim.
A5.02Coming soon

Data that should not travel

Outcome: You can name classes of data that should stay in the tools you already trust, or not be used at all.

  • National identity numbers, bank credentials, and unrelated personal files are not "context."
  • If the workflow does not need it, the agent should not see it.
A5.03Coming soon

Access you grant — and how you take it back

Outcome: You can grant access for one workflow and revoke it without a treasure hunt.

  • Write down what was connected, in whose name, and how to disconnect it.
  • A departed staff member should not leave an agent behind with their keys.
A5.04Coming soon

A quiet test week

Outcome: You can run a trial in which sending, paying, and deleting stay off or stay gated, on purpose.

  • Start with drafts and matching; turn on the last mile when the drafts are boringly good.
  • Tell the people whose work is in the trial what will happen, and how to raise a problem.

Practice note: Run a quiet internal trial first, on your own work. Human review of the output is mandatory. Package what worked as a reusable briefing, then take two small external pilots rather than one large one. Access to data often takes longer than the build; plan for that.

A5.05Coming soon

Talking to your team about agents

Outcome: You can explain the handoff to colleagues without hype or threat.

  • The point is time back on the work only a person can do.
  • People should know what the agent will touch, and how to correct it.

Practice note: Adoption starts with boring, repeating work, not with a flashy demo. Treat anything the agent writes as something that can leak, and agree how knowledge is shared case by case.

A5.06Coming soon

Tools will change; the handoff should not

Outcome: You can judge a new model or a new app with the same four-part test.

  • Ask where the outcome, the steps, the gates, and the log live now.
  • A shiny assistant that cannot stop or cannot record is still the wrong shape.
A5.07Coming soon

Keeping the workflow list honest

Outcome: You can retire a workflow that is not earning its keep.

  • Not every idea deserves to stay live.
  • An unused agent with leftover access is a risk, not an asset.
  • When you weigh a workflow, count the costs teams usually forget: handling exceptions, people reviewing output, changing how the team works, and the risk of a mistake.

Practice note: When you account for agent work, it helps to separate four kinds of cost — the opportunity, the project, the person, and the artifact — so spend does not hide in a single lump.

A5.08Coming soon

What to do when something goes wrong

Outcome: You have a first-response path: stop the step, read the log, escalate to the customer incident/privacy lead, and verify the repair.

  • Speed matters; so does not guessing.
  • Privacy incidents follow your PDPA process — the agent does not get a special exception.