Brink Work with us

Is AI making your decisions better?

Improving decision making in your organisation isn’t just a matter of introducing AI; you have to design new ways for people and AI to work together.

Scroll

When we speak to the funders, governments, and companies we work with, the same concern keeps coming up:

AI is being adopted across their organisations, but they’ve got no way of telling whether their decisions are getting better, or just getting faster.

In a 2026 EY survey, half of UK CEOs said AI tools were one of the biggest contributors to their productivity gains, but only 14% could see clearly what AI was costing and returning. EY-Parthenon CEO Outlook Survey, September 2026

Systems designed for a much slower world

So many current systems, the ones created to decide what gets funded and what gets built next, were designed for a much slower world, one with less to keep track of and more time to keep track of it in.

Today teams using these systems find themselves firefighting, reacting to things they should have seen coming, while their databases and evidence libraries grow more bloated and the time they have to weigh critical decisions keeps shrinking.

Under that kind of pressure the temptation is to throw automation at the problem: find the decisions that are logjammed and insert an AI model into the process.

But all that happens then is new information is delivered faster, outpacing the system that was designed to filter it and turn it into useful knowledge.

So the decision-making system gets left behind when it should be improving.

There are two lessons here:

  1. A faster decision isn’t automatically a better one; and
  2. Humans and AI can only make better decisions if the system around them is designed to continuously improve their interactions.

The evidence

When nearly 1,000 high school maths students practised with AI, the ones whose tutor gave hints instead of answers improved more than two and a half times as much as the ones using a standard ChatGPT-style assistant that simply answered their questions.1

Compared with students using only their textbooks and notes, the hints group did 127% better in practice, and the answers group only 48% better.

Finally, all three groups were tested again with nobody given access to AI. Those who had been given hints scored about the same as those who’d never used AI, while the students who had been handed the answers scored 17% lower.

The design of the interaction decides how much AI adds, and whether it costs people their own skills along the way.

A 2026 framework for human-AI decision-making teams reached the same conclusion: teams only do better than either working alone when the collaboration is deliberately designed.2

AI tutor that gives hints

Teachers helped set it up to guide students towards the answer, so they did the thinking themselves.

Standard ChatGPT-style assistant

It answered whatever it was asked, so students leaned on it, and fell behind once it was gone.

Dashed line: students who used only textbooks and notes

Hints instead of answers gave students the biggest boost, and took nothing away when the AI was gone.

Bastani et al., Generative AI without guardrails can harm learning, PNAS, 2025. Nearly 1,000 students. Scores compared with students who used only textbooks and notes; the final test was taken without AI.

The most pressing choices we have to make about AI right now are choices about how people and AI make decisions together.

A person can hand over a decision completely to the AI, or instead they can work alongside the AI to reach a mutual understanding.

At the same time, the way that decision gets made can stay fixed, or it can improve each time it is made.

Put these choices together and we get four ways of making decisions with AI.

Type 1 of 4

Meat Proxies

In this quadrant, the decision is outsourced completely to AI, and there’s little review or oversight, which means the decision-making process never improves.

Meat proxy is the name the internet has already created for the kind of person who forwards AI output without reading or checking it.

The pull towards it is strongest when the AI is usually right, because the better a tool performs the less reason people see to question it.

In one study, 27 radiologists read mammograms with suggestions from what they were told was an AI system, which was deliberately wrong on 12 of the 40 cases.3 The authors call what happened automation bias: once a trusted system has made a suggestion, people give it more weight than their own reading of the scan.

When we put our faith in a fallible machine, even the most experienced can see their decision-making capabilities suffer dramatically.

Least experienced readers

80%to20%

Correct when the AI was right, then correct when the AI was wrong.

AI right

AI wrong

Most experienced readers

82%to46%

Over 15 years’ experience on average, and still pulled off course when the AI was wrong.

AI right

AI wrong

Even radiologists with years of experience followed the AI when it was wrong.

Squares show the share of cases each group assessed correctly, out of 100. Dratsch et al., Radiology, 2023.

Type 2 of 4

Hoarders

Hoarders work through decisions alongside their AI, and the two of them get better at it, but the knowledge and the lessons stay with them.

When information is contained like this, one cog in the machine might move faster, but the rest of the system struggles to benefit and, in the worst cases, the system becomes unbalanced and starts to perform worse.

In software, where output is easy to measure, a study of about 10,000 developers in 1,255 teams tracked what happened when individuals adopted AI coding tools.4

When individual parts get faster but the systems around them stay the same, there’s no discernible gain to the overall business.

What changed when developers used AI

Each developer’s output, set against the results of the company as a whole

Developers using AI got much more done, but their companies saw no difference.

Faros AI, The AI Productivity Paradox, 2025. Telemetry from about 10,000 developers in 1,255 teams. Faros sells engineering analytics software.

Type 3 of 4

Offloaders

Here the decision is handed to an automated system, which runs its own loop and improves from its outcomes while the people around it watch.

It’s a closed system: the machine learns, but what it learns isn’t passed back to the people using it, so they have little idea what’s going on under the hood and no choice but to trust it completely.

Handing the thinking over in this way is known as cognitive offloading, and leaders are already seeing its effects: in a BCG survey, senior executives reported it eroding the critical skills their people need, and cutting back the mentoring and knowledge sharing that used to pass those skills on.5

When the system learns and the people don’t, the organisation is left trusting a machine that nobody fully understands, and there’s nothing to stop the machine ‘learning’ in the wrong direction.

What 70 senior executives had seen in their own organisations

Each square is one executive

Half

had seen critical skills eroding

A third

had seen less mentoring and knowledge sharing

As more of the thinking moves to AI, senior leaders are watching their people’s skills erode.

BCG Institute, When Everyone Uses AI, Companies Risk Critical Skills, 2026. A survey of 70 senior executives plus a dozen interviews, reporting what leaders have observed.

Type 4 of 4

Evolutionists

Evolutionists work through decisions alongside an AI, and that back and forth informs the way those decisions are made, so the process evolves and improves each time.

In this dynamic the person gets better at working with the AI, while the AI adjusts to support the person better, so the method they share keeps improving.

In one study, physical therapists assessed patients’ rehabilitation exercises with an AI system that showed them which movements its assessment was based on. The therapists told the AI which of those movements mattered for each patient, and the AI adjusted its assessment in response.6

Both sides got better: the AI learned from the therapists’ feedback, and the therapists made more consistent assessments with the AI’s help.

Accuracy of the assessments

Therapists alone, the AI alone, and the two working together

How each side improved

0.84to0.91

The AI’s own score, after learning from the therapists’ feedback

0.66to0.71

How often therapists agreed with each other, using a standard system and then this one

When the therapists and the AI learned from each other, together they beat both.

Lee, M. H. et al., A Human-AI Collaborative Approach for Clinical Decision Making on Rehabilitation Assessment, CHI 2021. Scores run from 0 to 1 (F1, a standard accuracy measure). The accuracy percentages come from a 2024 review of human-AI collaboration studies (source 7).

From our work

At Brink, an increasing amount of our work sits in that top right quadrant.

The people who know the work lead the work, while the AI fits around them.

EdTech Hub AI Observatory
Pakistan, Rwanda and Tanzania

Putting teachers in the lead on the AI tools built for them

In three case studies, each run over four months, teachers were put in charge of shaping the AI tools meant for their classrooms. In Pakistan, the team spent a month inside teachers’ working lives before any design began. In Rwanda, every product sprint started with teacher feedback, and in Tanzania teachers steered the tools towards problems the designers hadn’t planned for.

Teachers went from testers to product owners, and the tools that worked best were the ones that fitted how teachers already work.

Read the report

1–2 minutesto plan a lesson in Pakistan, down from about an hour

Frontier Tech Hub
Kigoma, Tanzania

Designing a legal advice app around the people women already trust

Dada Wakili is an AI-powered legal assistant app in Kiswahili and English. The team first assumed women would use it directly on WhatsApp, but workshops with more than 40 community paralegals showed that many women rely on feature phones or shared devices and can’t always afford data, so a chatbot built for direct use would never reach many of them.

The tool was redesigned for the paralegals women already turn to. The same conversations changed its content: the law was reorganised around how people describe their problems, such as “My husband has died and his relatives have taken everything. Where shall I go?”, rather than around legislation. Paralegals now get reliable legal information, checklists and next steps to share with their clients, and the next phase tests it in real casework.

Read the case study

PlannedWomen using a WhatsApp chatbot on their own phones

BuiltA tool for the paralegals women already go to for help

Brink Foundation
Gender and Learning Evidence Fund

Keeping funding decisions in human hands

When we reviewed applications to our open call, we decided where AI belonged before we started. People read, scored and discussed 100% of applications, and reviewers didn’t use AI to summarise them, because deciding who gets funded needs nuance and context that AI tools couldn’t match.

AI had two other jobs. It double-checked our eligibility decisions: Claude disagreed with us 27 times, and on review human judgement proved far stronger. And a dashboard we built with it showed the shape of the applicant pool, so we could see how different funding choices would change the portfolio. In our own words, AI was “most valuable to improve the digestibility of information, and least reliable wherever it was asked to judge.”

100%of applications read and scored by people

  • PeopleJudging, scoring, deciding
  • AIChecking eligibility, mapping the applicant pool

Your most consequential decisions should be made from an evolutionist standpoint.

But getting to that top right square depends mostly on how people and AI are set up to work together, and on a system that improves and evolves that dynamic each time it runs.

Anyone can buy an AI model off the shelf and point it at their nearest bottleneck; most of your competitors already have.

The real gains and impact come from designing how your people and that model work together, and that means dealing with the messy human reality around your decisions.

Want to put your people in the lead?

Brink brings together technologists, designers, strategists and behavioural scientists to help organisations design how their people and AI work together, so the decisions that matter improve each time they’re made. If you’d like to explore what that could look like in your organisation, we’d love to hear from you.

Get in touch coffee@hellobrink.co