AI Decision Support
Support means the person still decides, and can still say no.
What is decision support?
Decision support means putting relevant information and reasoning in front of the person making a decision, in a form they can question and override, as distinct from making the decision for them.
The distinction between supporting and making a decision sounds obvious and collapses quickly in practice. A system that produces a recommendation with a confidence figure, and that people are busy, becomes a system that makes decisions with an approval step attached.
That drift is predictable and it has a name in aviation and medicine: automation bias. People accept a system suggestion more readily than they would accept the same suggestion from a colleague, particularly when they are tired or the queue is long.
Designing against it means showing the reasoning rather than the conclusion, making disagreement easy, and recording how often anybody actually disagrees. A support system nobody ever overrides has stopped being one.
Why a low override rate worries us
Duara builds decision support in Kenya for organisations where a person must remain accountable for the outcome.
We measure override rates and treat a very low one as a warning. If nobody ever disagrees with the system, either it is never wrong or people have stopped thinking, and the second is far more likely than the first.
Who this is for
- Decisions need to stay with people
- Volume makes consistent judgement hard
- Reasoning must be explainable afterwards
- You worry staff will simply accept whatever appears
Industries we serve
- Financial services
- Insurance
- Healthcare administration
- Manufacturing planning
- Public sector
- Logistics operations
Problems this solves
These are the conditions where support helps and where it quietly turns into automation nobody chose.
- Support drifting into automation. A recommendation with a confidence figure and a busy team becomes a decision with an approval click attached.
- Automation bias. People accept a system suggestion more readily than the same suggestion from a colleague, particularly under pressure.
- Conclusions without reasoning. A recommendation nobody can interrogate cannot be meaningfully agreed or disagreed with.
- Inconsistent judgement at volume. The same case decided differently by different people, or by the same person at different times of day.
What we deliver
Showing reasoning rather than conclusions, making disagreement easy and expected, measuring override rates, and treating a low one as a design failure rather than a success.
Reasoning shown rather than a bare recommendation
Disagreement made easy and expected
Override rates measured and reported
A record of what was shown when a decision was made
Consistency support without removing judgement
Escalation where a case is outside the system range
Features
- Reasoning, not conclusions. What the system considered and why, so the person can genuinely agree or disagree rather than accept.
- Disagreement made easy. Overriding should take no more effort than accepting, since friction on one side produces a decision nobody made.
- Override rate treated as a health measure. A very low rate means people have stopped thinking, which is more likely than a system that is never wrong.
- A record of what was shown. What information was in front of the person at the time, which matters when a decision is later questioned.
- Out of range cases escalated. A case unlike anything the system has seen is flagged as such rather than given a confident answer.
Benefits
- Accountability stays with people. The person deciding can explain the decision, which is the requirement in most regulated and consequential work.
- More consistent decisions. Support reduces variation between people and across the day without removing the judgement itself.
- Decisions defensible afterwards. A record of the reasoning shown at the time answers the question that arises months later.
- Drift detected early. Measuring overrides shows when a support tool has quietly become an automation nobody approved.
Technologies we use
- Reasoning presentation. Showing what was considered and why rather than a conclusion with a score.
- Override capture. Recording disagreements and the reasons given, which is the most useful data the system produces.
- Confidence and range checking. Identifying cases outside what the system can reliably support and saying so.
- Decision record. Storing what was displayed at the moment a decision was taken.
- Consistency analysis. Measuring variation between decision makers without measuring individuals against each other.
Automation bias is well documented in aviation and medicine and it applies here exactly. The design question is not whether people will over trust the system but how much you slow that down.
How the process works, step by step
Decision support usually takes six to ten weeks, and the override measurement continues indefinitely rather than ending at delivery.
-
Step 1
Free decision review
We establish which decisions are being made, by whom, and what information they currently have.
-
Step 2
Accountability confirmed
Who remains responsible for the outcome is settled in writing before anything is built.
-
Step 3
Reasoning designed
How the system will show its working is designed before what it will recommend.
-
Step 4
Override measured from day one
Disagreement is recorded from the start, since the rate matters more than the accuracy.
-
Step 5
Reviewed for drift
We check whether overrides have fallen toward zero, which indicates acceptance rather than agreement.
Quality assurance and how we work
The commitments attached to systems that influence what people decide.
- We treat a low override rate as a warning. If nobody disagrees, either the system is never wrong or people have stopped thinking, and the second is far more likely.
- Disagreeing is as easy as accepting. Friction on one side of that choice produces decisions nobody actually made.
- Reasoning is shown, not a score. A confidence figure invites acceptance. Showing what was considered invites judgement, which is the point.
- Out of range cases are declared. A case unlike anything the system has seen gets flagged rather than a confident answer it cannot support.
- We record what was displayed. When a decision is questioned months later, what the person could see at the time is the relevant fact.
Frequently asked questions about decision support
What is the difference from automation?
The person genuinely decides. The distinction sounds obvious and collapses quickly when a recommendation carries a confidence figure and the team is busy.
What is automation bias?
People accept a system suggestion more readily than the same suggestion from a colleague, particularly when tired or under queue pressure.
How do you design against it?
Show reasoning rather than conclusions, make disagreement as easy as acceptance, and measure how often anybody actually overrides.
Why measure override rates?
Because a very low rate means people have stopped thinking rather than that the system is never wrong. It is the clearest drift signal available.
What override rate is healthy?
It depends on the decision, and the useful thing is the trend. A rate falling toward zero over months is the warning.
Should the system show a confidence score?
We prefer reasoning. A score invites acceptance, while showing what was considered invites judgement, which is the entire purpose.
What happens with an unusual case?
It is declared out of range rather than given a confident answer the system cannot support.
Who is accountable for the decision?
The person making it, and we settle that in writing before building. Support does not transfer accountability and should not appear to.
Can it make routine decisions automatically?
That is a different system and a different conversation. Where you want automation we would say so rather than calling it support.
How soon would it be supporting decisions?
Six to ten weeks, with override measurement continuing indefinitely rather than ending at delivery.
Will it make decisions more consistent?
Usually, by reducing variation between people and across the day, without removing the judgement itself.
Can it be used in regulated work?
That is where it fits best, since accountability stays with a person who can explain the reasoning.
Does it record what was shown?
Yes, which is the fact that matters when a decision is questioned months later.
Will staff feel monitored?
Overrides are measured at aggregate level rather than to judge individuals, and we would say that clearly to the team.
What if the system is frequently wrong?
A high override rate is useful information and we would investigate rather than tune the system to agree with people more often.
Can it learn from overrides?
Cautiously. Learning to agree with people more often is not the same as being more correct, and that distinction matters.
Does it work for operational decisions?
Yes, and repetitive operational decisions at volume are where consistency support helps most.
Is this the same as analytics?
No. Analytics answers questions you ask. This puts information in front of a specific decision at the moment it is being made.
Who can see decision records?
Only your authorised staff, and we would keep them separate from performance management.
What does decision support cost?
Decision volume and complexity determine it, quoted after the decision review.
Glossary of terms
Decision support terms in plain language.
- Automation bias
- The tendency to accept a system suggestion more readily than the same suggestion from a person.
- Override
- A decision maker disagreeing with the system, which should be easy and is worth measuring.
- Out of range
- A case unlike those the system can reliably support, which should be declared rather than answered.
- Decision record
- What information was displayed at the moment a decision was taken.
- Support versus automation
- Whether the person genuinely decides or approves a decision already made.
- Consistency
- Reducing variation between decision makers without removing their judgement.
Related services in AI Solutions
Talk to us about decision support
Name a decision your team makes daily and ask how often two people would decide differently. That variation is what support addresses.
Start your project