Answer
What should happen when an agent and a person disagree?
Look at what each was working from. The disagreement is usually about inputs rather than judgement, and that is resolvable.
Compare what each was working from before deciding who is right. Most of these disagreements are about inputs rather than judgement: the agent read something the person had not, or lacked something the person knows.
The instinct is to treat this as a question of authority: whose judgement wins. That framing skips the step that resolves most cases, which is establishing what each party had. An agent that read forty records and a person who remembers the last six are not disagreeing about interpretation, and neither is a person who knows a customer's history that appears in no system and an agent that does not.
So the first move is comparison of inputs. What did the agent read, what did the person know, and does either contain something the other lacks. This is only possible if the agent's output carries its evidence, which is the practical argument for requiring references: a disagreement between two referenced positions resolves in minutes, and one between two unreferenced positions can only be settled by seniority.
The most common resolution is that the agent had data and the person had context. The records show one thing and the person knows why: a customer with an unusual arrangement, a period that was atypical, a field that stopped being maintained. In that case the agent is correct about what the data says and wrong about what it means, and the resolution is to write the context down so the next run has it rather than to override and move on.
The second most common is the reverse and it is worth taking seriously. The agent has read something the person has not, or has noticed a pattern across more cases than anyone reviewed. A person's confident recollection of how things usually go is a sample from memory, and memory samples badly. Where the agent's position is supported by records the person has not examined, dismissing it because it came from a machine is how a business ignores its own data.
The third case is that the agent is simply wrong, and the useful response is to find out how rather than to correct and continue. A wrong conclusion came from somewhere: a misread field, an assumption, a source that does not say what it was taken to say. Locating it takes a few minutes and tells you whether this is a one-off or something that recurs, which is a much more valuable output than a corrected answer.
Whatever the resolution, it should change something outside the exchange. The context written down, the check added, the assumption corrected in the brief. A disagreement resolved only in conversation will recur in exactly the same form, because nothing about the next run will be different, and the second occurrence is where people conclude that these systems do not work.
Before deciding which of you is right, find out whether you were looking at the same thing.
Siddharth Sharma, Context Theory
Related questions
Who should decide when it genuinely comes down to judgement?
The person, and specifically the person who will answer for the outcome. That is not a claim about whose judgement is better; it is that a decision needs somebody who bears its consequence, and the agent does not. Where nobody in the business bears the consequence, that is a finding about the process rather than about the disagreement.
Does the agent being confident mean anything?
No, and neither does its hedging. Confidence is produced regardless of the strength of the position, so it carries no information about which side of a disagreement is better supported. The evidence each side can point at is the only thing that separates them.
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 |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Most disagreements between an agent and a person are about differing inputs rather than differing judgement, so establishing what each had access to resolves them without either party being wrong. | Listing what the agent read and what the person knew for a specific disagreement and checking for content unique to each. |
| Response | A disagreement between two referenced positions resolves in minutes while one between two unreferenced positions can only be settled by seniority, which is the practical argument for requiring evidence in agent output. | Attempting to resolve a disagreement where neither side can point at a source. |
| Constraint | A person's confident recollection of typical behaviour is a sample from memory, which samples badly, so dismissing an agent position supported by records nobody has examined is how a business ignores its own data. | Comparing a recalled typical figure against the same figure computed from records. |
| Software | A resolution that changes nothing outside the exchange recurs identically, because nothing about the next run differs, and the second occurrence is where confidence in the system is lost. | Checking whether a previously resolved disagreement has recurred in the same form. |
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