← Insights

Insight · Workload selection

Which workloads should a Digital Worker run first?

A five-test scoring framework for shortlisting the work a governed AI system should take on first — and the workloads to defer until the platform has earned trust.

July 2026 · 8-minute read · Free — no email wall

The first workload decision carries the program

Most AI programs in health, government and regulated enterprise don't fail on model quality. They fail on workload selection: the first thing automated is either too ambitious (open-ended judgment, irreversible consequences, no evidence trail) or too trivial to matter (a demo nobody defends at budget time). Either way, the program loses the one thing a first deployment must buy — institutional trust.

A Digital Worker is a multi-agent AI system with a defined scope of work: it reads, classifies, drafts, reconciles or assembles within set boundaries, escalates outside them, logs everything it does, and answers to a named human. That definition is also a selection filter. If a workload can't be given boundaries, an escalation path and a named accountable human, it isn't a first workload — whatever the demo looked like.

The framework below is the one the practice uses in the first week of a Digital Workforce Blueprint. It is published in full because the thinking shouldn't be the secret — the discipline of applying it is the work.

Five tests of a suitable first workload

Test 1 — Rules-first decidability

Can most of the workload be decided by rules a subject-matter expert can write down — with the model doing the reading, extraction and assembly rather than the judging? Workloads with high rules density (eligibility checks, evidence assembly, reconciliation, structured triage) let every output trace to the rule that fired, the evidence it read and the calculation it made. Workloads that are mostly judgment (clinical diagnosis, discretionary approvals, anything a tribunal might later unpick) put the model where the accountability is — a place it should not be first.

Test 2 — An evidence trail exists

Does the workload run on documents, records and feeds the worker can cite? A Digital Worker's credibility is its ability to show its working: this finding, from this document, under this rule. If the real inputs are hallway conversations, tacit knowledge and phone calls, there is nothing to cite and nothing to audit. Digitised, structured, citable inputs are a precondition — not a nice-to-have.

Test 3 — A clean escalation boundary

Can you state, in one sentence each, what the worker handles and what it hands to a human? Good boundaries are crisp: confidence below threshold, escalate; any clinical implication, escalate; anything outside these five document types, escalate. If drawing the boundary takes a workshop and still leaks, the workload will generate boundary disputes instead of throughput — and every dispute erodes trust in the whole workforce.

Test 4 — Failure is visible and reversible

When the worker gets something wrong — and it will — is the error caught by a human checkpoint before it acts on the world, and can it be undone? Drafts reviewed before sending, flags reviewed before action, reconciliation exceptions queued for a person: reversible. Payments issued, notices served, records deleted, care decisions made: not reversible, and not first. Autonomy is graduated to risk; the first workload should sit at the shallow end deliberately.

Test 5 — Volume that justifies the platform

Is there enough repetitive volume that removing it changes someone's week? The first Digital Worker carries the fixed cost of the platform, rails and governance it deploys onto. A workload that occupies hours of skilled-staff time every day repays that foundation and produces a defensible number at review time. A workload that occurs monthly does not — however elegant the automation.

Scoring the shortlist

Score each candidate workload 1–5 against the five tests and total them. The scale matters less than the discipline: scoring forces the conversation about escalation boundaries and reversibility before anything is built. A worked example from a hypothetical care provider:

Scores are illustrative — the point is the conversation each cell forces.

In this example, evidence assembly goes first: highest score, zero external blast radius, and it produces an artefact (a cited evidence pack) that demonstrates the audit-trail discipline to the very governance forums that will approve worker number two. Intake triage follows on the same platform. The complaint letters wait — not forever, but until the boundaries and the governance have been proven on safer ground.

Workloads to defer — whatever the enthusiasm

• Open-ended judgment dressed as admin. If the "processing" step is actually a discretionary decision, the workload fails Test 1 no matter how much paperwork surrounds it.
• Irreversible external actions. Anything that pays, notifies, denies, deletes or treats without a human gate. These come later, with autonomy graduated to the evidence.
• Thin or contested evidence. If humans doing the job today argue about what the source of truth is, a Digital Worker will simply automate the argument.
• Politically loaded workloads. A first deployment under hostile scrutiny needs everything above plus luck. Earn the track record first.
• Someone's whole job, framed as a workload. A Digital Worker takes a scope of work, not a position on the org chart. Workloads that are really restructures fail on escalation design — and on trust.

From shortlist to scope of work

A scored shortlist is not yet buildable. The top candidates get written up like position descriptions: the boundaries, the inputs and the systems they arrive from, the escalation paths, the named human accountable, and what "done well" measures as. That scope-of-work document is the contract between the operation and the worker — it is what the evaluation harness tests against, what the review board approves, and what the audit trail is read against in twelve months.