Answer
How should an agent handle a task it has never seen?
By establishing what it is looking at before acting, and treating an unrecognised case as a reason to stop rather than improvise.
Orient first: establish what this is, whether it resembles anything handled before, and whether it falls inside the defined scope. Where it does not, stop and report. An unfamiliar case is where improvisation is least appropriate and most likely.
This is the situation where the shape of the failure is most predictable. Faced with something outside what it has handled, a system will construct a reasonable approach and proceed, because producing a reasonable approach is what it does. The output looks like the output of a run that knew what it was doing, and nothing distinguishes them. That is why an unfamiliar case needs an explicit handling rule rather than a general disposition towards care.
The rule that works is to define the scope by the characteristics of the input rather than by the agent's confidence. This kind of record, these categories, these formats, these value ranges. Anything outside is out, and the determination is a check rather than a judgement. This is more reliable than asking whether it feels unusual, because unusualness is exactly the property a system with no history of the domain cannot assess.
Where a case is inside the scope but unfamiliar in some respect, orientation comes before action. What is this, what does it resemble, what is different about it, what would the usual approach do here and why might that be wrong. Producing that assessment costs one cheap turn, it is readable, and it frequently reveals that the case is ordinary and was merely presented differently.
The report on an out-of-scope case matters as much as the stop. What arrived, why it did not match, what the nearest match was, what it would have done. That last item is genuinely useful: it tells whoever reads it whether the boundary is drawn correctly, and a stream of near-miss cases where the proposed approach was right is the evidence for widening the scope. Without it, the boundary can only be adjusted by guessing.
Accumulating these is the mechanism by which the workflow improves. Out-of-scope cases cluster, and a month of them is a short list of things the design did not anticipate. Some become new rules, some become a widened definition, and some are correctly left to a person forever. All three are useful outcomes and none is available without the record.
There is a case worth distinguishing: a task that is unfamiliar to the agent and routine for the business. This usually means the workflow was scoped from a description rather than from observation, and the missing case was always there. It is not a novel situation, it is a specification gap, and treating it as the first is how a workflow acquires increasingly baroque handling for cases that should have been in the definition.
The novel case is exactly where an agent has least to go on and most freedom to invent, which is a poor combination.
Siddharth Sharma, Context Theory
Related questions
Should it try and then stop if it fails?
Only where trying is free and reversible, which for a read-only investigation it usually is. Where the attempt involves acting, the order should be reversed: assess whether this is in scope, then act. An attempt that has already changed something cannot be undone by noticing afterwards that it should not have been made.
How specific should the scope definition be?
Specific enough that the determination is a check rather than an assessment. If deciding whether a case is in scope requires judgement, the boundary will be applied inconsistently and the cases near it will be handled differently on different runs, which is worse than a boundary drawn slightly wrongly and applied the same way every time.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Firms that never responded to a web enquiry at all | 23% | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Faced with an unfamiliar case a system constructs a reasonable approach and proceeds, producing output indistinguishable from a run that knew what it was doing, which is why an explicit handling rule is required rather than a disposition towards care. | Submitting a case outside the intended scope and comparing its output against an ordinary run. |
| Software | Scope defined by input characteristics makes the in-or-out determination a check rather than a judgement, which is more reliable than asking whether a case feels unusual, since unusualness is what a system without domain history cannot assess. | Testing whether the scope boundary can be evaluated without any judgement about the case. |
| Response | An out-of-scope report that includes what the agent would have done tells the reader whether the boundary is drawn correctly, and a run of near misses with correct proposed approaches is the evidence for widening it. | Reviewing a month of out-of-scope reports for whether the proposed approach would have been right. |
| Constraint | A case unfamiliar to the agent but routine for the business indicates a specification gap from scoping by description rather than observation, and treating it as a novel situation produces increasingly elaborate handling for cases that belonged in the definition. | Asking whether the unhandled case is one the business encounters regularly. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one