How Perfect is YOUR Information?

I've been reading a lot of research lately. One of my favourite agents is Rosie Research - every morning I get up to a tailored list of relevant new research, case studies, use cases and AI transformation news, with the vital figures and "so whats" pulled out for my convenience. It's a wonderful way to start the day when you live deep in the world - as I do - of building meaningful, responsible, effective ai capability in regulated and professional spaces.

What Rosie delivers isn't head those breathless "AI will change everything" headlines - I'm as tired of those as you are. And if I see one more stock photo of a mechanical robot arm reaching out to touch a glowing human hand, or one more headline about how many jobs ai will replace, I may have to write a strongly-worded letter to someone.

What Rosie delivers to me that I'm lost each morning in is the harder stuff. The Gartner data. The McKinsey & Company surveys. The peer-reviewed studies from Stanford University and Massachusetts Institute of Technology that don't make the news because "carefully controlled experiment shows nuanced results" doesn't get clicks.

And something interesting is emerging through the research, the data, the case studies from across industries.

The headlines call AI technology. We all seem to believe that AI is a technology challenge. Governments fund technology departments to solve it. Organisations hire technology consultants to implement it. Many boards are still asking their CTOs to own it.

But this is the category error the vast majority of companies are making.

There is no technology problem here.

What we're facing is a human capability challenge - and it is mammoth.

The reason OpenAI can't solve it, why Anthropic can't solve it, and arguably why the technology departments of governments can't solve it, is because the technology problem...

...doesn't exist.

The technology works. It does an extraordinary amount of stuff. The power is immense.

The question now is: how do we put it to work?

Think about buying a new washing machine.

You don't call up the engineers at Bosch to ask them how to connect it to your water supply. It's nothing to do with them. Their job ended when they created the machine. No, you call the plumber - the person who specialises in putting the machine in place, connecting it to electricity, making sure it's in the right spot (not in the middle of your living room but tucked in the cupboard under the stairs), connected up to the right hoses in and out, and making sure it's safe to use.

That's their job.

And this misunderstanding - that tech companies are supposed to solve problems way beyond their remit - is everywhere. The machine works. If it doesn't have a safety grounding cable then yes, we go back to the manufacturer (got that, Grok?).

But beyond that? If people are putting the machine in the middle of their living room and wondering why the water's flowing all over the floor, the answer isn't to call Bosch. The answer is to connect the out hose to the waste pipe. Maybe you need a plumber for that. Or maybe you build some skills through YouTube and learn it yourself.

I think we're going to start making meaningful progress towards real results when we start thinking that way about AI.

The technology is ready. What's not ready is the human capability to use and leverage it effectively.

About 96% of current AI usage is at the most basic, preliminary, un-impactful level.

Yes those who use AI often believe themselves to be power-users (Dunning-Kruger at its best here: much of my role is gently helping those who believe putting a transcript into Chat or asking it to rephrase an email, or do a search better meant for Google, is not in fact deep usage.)

But we don't know what we don't know, right? So a lot of my time is spent teaching people to fish: we don't give them pre-made use cases and hey here, use these. Instead we say, here's how we can start thinking about the job you do, and start coming up with the highest impact use cases that really move the needle for you.

Having said that, there are a few AI use cases that it seems, no matter where we go, large clients or small, technology or consulting, education or law, that cause our participants walk out of the room with the determined look full of wonder of someone who finally understands what all the fuss is about. Keep reading.

The Concept That Keeps Coming Back To Me

If you studied economics, you'll remember perfect information. It's one of those foundational concepts that gets introduced and immediately contradicted.

Perfect information: the theoretical state where all market participants have complete access to all relevant information, all the time. Prices reflect reality. Decisions are optimal. Everyone knows what everyone else knows.

Beautiful model. Your professor then spent the next hour explaining why it's entirely fictional.

Real markets are riddled with information asymmetries. Buyers know different things than sellers. Insiders know different things than outsiders. Every economic calculation carries a heavy discount for imperfect information - the stuff people don't know, can't access, or can't find in time.

And as all of you know as well as me, the same principle applies inside organisations. Maybe more so.

Think about where your organisational intelligence actually lives when it comes to say, your clients. It's walking around in the heads of senior people. It's in the coffee chat someone had with a client last Tuesday. It's in the quick update at the end of a meeting about something else - the one that concerned everyone but never quite made it into a written note.

It's in emails that are technically searchable but practically unfindable, or the personal accounts of the lead. In transcripts that exist but nobody reviews. In CRM entries that are filtered through the human doing the logging - who may not remember fully, may file it late, may omit aspects they deemed irrelevant at the time, or may be in a rush to get home or to an event.

Even information that does get captured in your CRM is, by definition, imperfect. Subject to bias, recollection, interpretation.

And this information - or the lack of it - shapes everything. The preparation you bring to meetings. The context you have when someone calls unexpectedly. The institutional memory that either exists or doesn't when a long-standing client mentions something from three years ago.

What The Research Actually Shows

McKinsey's data suggests knowledge workers spend an average of 1.8 hours per day simply searching for information. Not creating value. Not serving clients. Not thinking strategically. But just... looking for things.

For a firm with 1,000 knowledge workers, that's somewhere in the range of £2.5-3.5 million annually. Not invested, not building capability, just pure waste. Money evaporating in the gap between knowing something exists and being able to find it.

Meanwhile, Gartner's latest numbers are sobering. Only 9% of organisations have reached what they call "AI maturity." 72% of CIOs report breaking even or losing money on their AI investments. And the trend perhaps most concerning: the failure rate is getting worse, not better. In 2021, 46% of AI prototypes failed to reach production. By 2025, it's 58%. Average project duration has increased from 7 months to 12 months. News says 95% of AI projects fail.

So the tools are better and better, more and more powerful, more connected, more usable... but outcomes are worsening.

If this were a technology problem, that pattern wouldn't make sense. We have dramatically better technology than we did in 2021. The models are more capable. The interfaces are more intuitive. The infrastructure is more robust.

What we have is a human capability problem wearing a digital disguise.

The Category Error That's Costing Millions

RAND Corporation's peer-reviewed research - based on interviews with 65 experienced data scientists - identified five root causes of AI project failure. Not one of them is "the technology didn't work." They're: misunderstanding the problem to be solved. Inadequate training data. Technology focus over problem focus. Inadequate infrastructure. Attempting problems too difficult for AI.

Notice what these have in common. They're human failures. Failures of understanding, of framing, of judgment, of organisation.

The Harvard/BCG study with 758 consultants is even more revealing. Within what they call the "technological frontier" - tasks AI handles competently - consultants produced 40% higher quality output and completed work 25% faster. That's the good news.

There's a complication though: outside that frontier, performance dropped by 19 percentage points. People who used AI inappropriately produced worse work than those who used no AI at all.

The technology amplifies. It amplifies good judgment and it amplifies poor judgment. It can accelerate effective work and it can accelerate ineffective work. The determining factor isn't the tool - it's whether people know how to use it.

Back to the washing machine: the machine works brilliantly. But if you put a wool jumper on a boil wash, that's not Bosch's problem. That's a capability gap. And no amount of machine improvement fixes it.

What This Actually Looks Like In Practice

Let's get concrete, because economic theory is interesting but your Wednesday afternoon is more immediate.

Across our clients - education, banking, professional services, manufacturing, asset management - certain patterns emerge when we ask the question: where can AI make the biggest, quickest positive impact (within our constraints of tools, compliance, security)?

Every organisation is different. But five structures come up again and again. High impact, low cost, implementable with existing tools.

Each of these can be configured in 3-5 minutes. No technical knowledge required. Using tools you likely already have access to. Security implications entirely manageable.

1. Meeting Intelligence → Organisational Memory

Your meetings are already being transcribed if you're using Teams or Zoom. Possibly by multiple systems simultaneously. But what happens to those transcripts?

In most organisations: nothing. They sit in a folder. Nobody reviews them. The information captured evaporates within days. Maybe our Dunning-Kruger power users possibly upload them to Claude or ChatGPT to write a quick follow up note.

Configure a project area with your company's templates, your preferred formats, your methodologies. Ensure your team account is configured with your brand guidelines, voice, purpose, etc. Set up the automation from notetaker to holding folder. Set up the project area to access the holding folder full of transcripts. Now it's the ultimate Client Knowledge Base - it extracts the key insights - client needs, decisions made, follow-up commitments - and can draft communications, create proposal outlines, or simply log the information in a way that's actually searchable. It can prompt you every morning with the current status of your client accounts. It can notify you when a distant colleague met with someone connected to your contact. Or when the news picked up that one of your clients has moved onto a new firm.

The intelligence that used to fester in corners now persists. The context that lived only in one person's head or deep in the crevices of their email and beyond their conscious memory, become accessible to anyone who needs it. Even if they didn't know they needed it.

Setup time: 30 minutes. Uses your existing £25/month ChatGPT or Claude licence. No technical knowledge required.

2. Senior Expertise → Queryable Standards

If you're a senior professional, you have a way of doing things that's taken years to develop. Frameworks you've refined. Quality standards you recognise instinctively. Formatting preferences that matter more than anyone admits. An aversion to diamond shaped bullet points.

Most of this knowledge is trapped in your head. It transfers to junior staff through repeated feedback - if they're brave enough to ask, if you have time to explain, if they catch you at the right moment.

Put your methodology in a Custom GPT. Write out how you structure client reports. What you always check for. Your house style. Your standards for what "good" looks like.

Now your juniors can run their work through before they knock on your door. They get your-style feedback at 11pm when you're not available. They iterate before the draft reaches you. The quality of first submissions improves. Your time for detailed review reduces.

You're not being replaced. You're being scaled. What used to be one-to-one knowledge transfer becomes one-to-many.

Setup time: 30 minutes.

(Implications for consulting industry, anyone?)

3. Policy Documents → Accessible Answers

Somewhere in your organisation is a compliance manual that nobody reads until there's a problem. HR policies. Procurement procedures. Regulatory requirements. The stuff that should take five minutes to find and somehow takes forty-five.

Build a project area with your policy documents in the knowledge base. Call it policybot. Or Polly. Now anyone can ask "what's our policy on expenses over £500?" and get a straight answer with a source reference and a link to the correct, most recently updated form.

The information asymmetry that used to exist between tenured staff and newcomers - the stuff everyone was supposed to "just know" - becomes accessible to everyone equally.

Setup time: 3-5 minutes.

4. Calendar → Automatic Preparation

Configure a system that triggers from your calendar each morning. It identifies your external meetings, pulls relevant context - recent LinkedIn posts from who you're meeting, previous interactions your team has logged, company news, shared connections you might not have noticed.

Delivered before you leave the house. As a 3 minute audio if you prefer. As a two paragraph brief if you'd rather read.

Research shows top performers achieve 2.7x more conversions through personalised meeting preparation - but most people don't do it because who has time? This gives you the infrastructure to do it without the overhead. The cost-benefit analysis of taking the time to research in advance of every meeting every participant didn't make sense: no one has time for that except in the most important of cases. But when it costs nothing but 3 minutes of your time, why not?

AI shifts the cost-benefit maths.

You walk in knowing you both went to Durham. You mention the article they wrote last month. You start off on the right foot.

Setup time: 3-5 minutes.

5. Scattered Knowledge → Synthesised Intelligence

Every organisation generates enormous amounts of intellectual property that becomes unfindable three months later. Project learnings. Market analyses. Client insights. Strategic thinking buried in presentation decks and senior employees' accounts that nobody opens again.

Connect your well-configured GPT to your knowledge base. Ask: "What have we learned about private equity clients in the past year?" Instead of hunting through SharePoint or hoping you remember which folder something was in, you get structured insight drawn from your own institutional knowledge. Instantly.

This isn't querying a generic AI; it's nothing to do with ai-generated or human-generated. This is quickly and efficiently querying the accumulated intelligence, the IP, of your own organisation.

Setup time: 3-5 minutes.

The £25 Reality

If you have a £25 ChatGPT licence and a transcription tool, you have everything you need to save your team from wasting 2+ hours looking for stuff a day. To remove the need for manual CRM updates - which is considerable hours per week for most client-facing teams. To potentially remove the need for a separate CRM altogether.

I mean that literally. Why are you entering information into Salesforce just to query Salesforce? Why not query your actual communication flows - the emails, the transcripts, the meeting notes - and get real-time intelligence without the manual data entry in between? Worried about face to face meetings? Fireflies. Worried about security and encryption? Plaud.

When we built our own client intelligence system, we realised the CRM was the middle man. The friction. The thing forcing humans to re-enter information that already existed in more complete form elsewhere.

Farewell, Salesforce. Farewell, HubSpot. You're soon to be replaced by a 3 minute hook up.

Our team now gets proactive client updates on their own schedule. I prefer a short notification every morning. Louise Ballard (Moody) prefers a weekly roundup. We both like looking at Delphi, the beautiful real-time dashboard Eliza Moody made to visually display our client position and progress. All these are configured in seconds, all pulling from the same underlying knowledge, all delivering exactly what we need without manual effort so we can focus on the stuff we enjoy: delivering for our clients.

The Conversation I Keep Having

A client told me recently they'd come back to us in a year or so, after they'd done a mass data cleanup project.

I did try to keep a straight face.

In a year, none of the gains we'd just discussed would be available to them. A whole year of 2+ hours per day per person, wasted. A whole year of manual CRM updates. A whole year of information evaporating after meetings.

What I tried to help them understand is that the gains from AI are immediate. Today. Now. This week. This afternoon!

There are ways - even in a regulated space, without touching your internal knowledge base, without going near sensitive data - to gain hours on your day with no security implications whatsoever. We call them Tier 1 applications, and it's where we start with any client.

The data cleanup can happen in parallel. But waiting for perfect data before starting is like refusing to use a bicycle until the roads are all repaved. You'll be waiting a long time, and you'll miss a lot of journeys. And let's face it, it just probably isn't even happening.

What The Research Shows About Who Benefits

The Harvard/BCG study found something that changed how I think about implementation.

Below-average performers improved by 43% when using AI effectively. Top performers? Just 17%.

AI compresses the skill distribution. It disproportionately benefits people who were struggling - giving them access to patterns and knowledge they couldn't previously access. The learning curve compresses. Stanford/MIT data shows new workers using AI-based tools reached the performance level of workers with six months more tenure in just two months.

But here's the catch. This only works if people know what they're doing.

Gartner's forgetting curve data is brutal: after one hour without application, you lose 50% of training. After six hours, 70%. After days, 90%.

This is why most AI rollouts fail. They deploy tools. Maybe they run a training session of limited relevance. Then people go back to their desks and... continue doing what they were doing before. The training evaporates. The tools get used sporadically, inconsistently, often incorrectly.

Usage might increase over time via those curious souls who poke about and chat about it at the water cooler; but meaningful value doesn't materialise.

Is Usage What You're Measuring? Stop!

This needs to be said clearly: usage is not a success metric.

AI usage could be terrible. AI usage could be transformative. Measuring usage tells you precisely nothing about which one you're getting.

Gartner documented what they call the "work swap" pattern. Unproductive users: write email with AI → recipient uses AI to summarise → write document with AI → recipient uses AI to summarise. Volume increases. Value doesn't. This is officially called "enshitification" - adding to the quantity of work rather than reducing it.

That's usage. It shows up in your dashboards. It looks like adoption. But it's actively making things worse.

If you've encouraged AI adoption without proper training - if people are using these tools without understanding how to verify outputs, how to configure their inputs, when AI is appropriate and when it introduces risk - you've opened significant exposure. Security risks. Reputational risks. Regulatory risks.

Deloitte's recent experiences are instructive, and they're not alone. There are documented cases across regulated industries of organisations that pushed adoption without building the capability to use these tools safely and effectively.

That's bad usage. And bad usage is genuinely dangerous. The fear of rolling out AI is justified!

Good usage is different. Good usage means treating AI as an amplifier of human capability. Steve Jobs called the computer "a bicycle for the mind" - something that doesn't pick your destination, but allows you to move further, faster, more efficiently, for less cost. Something that makes new places accessible that you couldn't reach before.

That's what AI can be. It's more like an F1 car than a bicycle. And like an F1 car, owning one doesn't make you Jensen Button. It requires training. A different mindset to win an F1 race than you might have driving your polo to the shops. It requires a governance structure. a track. Tiering that makes clear what's appropriate at what level of risk. A skilled team. Organisational context that helps people understand when AI adds value and when it introduces problems. When to go fast and when to slow down and check.

The question isn't "are your people using AI?"

The real question is "are they using it well?"

The All-Hands Reality

We run all-hands workshops with client teams. Two hours. Everyone in the room - or on the call - with their laptops open.

By the end of those two hours, every participant walks away with 3-5 immediate use cases configured and working. High impact. Using tools they already have access to. Saving a minimum of 4-6 hours per week. Often considerably more.

But I've started warning leaders: don't ask us to run an all-hands workshop if you haven't thought through the repercussions.

This sounds counterintuitive, but there is a reason and it is this:

When I can sit down with a team and, in about four minutes, configure a project area that does what previously required significant manual effort - when the room watches something that took hours happen in seconds - the reaction is not always pure enthusiasm.

Sometimes it's anxiety. Sometimes it's "what does this mean for my role?" Sometimes it's the four people whose jobs just got dramatically easier looking around and thinking "well... what now, boss?"

Many leaders come to us thinking a bit of AI here and there would be nice. Efficiency gains. Time savings. Nothing dramatic. Surely it won't be the hundreds-of-percent gains like the hyped-up headlines mention.

What they're often not ready for is the scale. The compressed timescales. The fact that we're not talking about 10% productivity improvements - we're talking about the kind of restructuring that requires a plan.

Once that project area we configured for a client was set up, once it was producing the reports that previously lay with 4 FTEs, in 2.3 seconds, to a higher quality, the genie doesn't go back in the bottle.

The organisations that are getting this AI stuff right aren't the ones with the best tools, the biggest budgets, or the hundreds of shiny complex automations. They're the ones whose leaders have thought through what happens after the efficiency gains land, and what they're really trying to achieve (higher quality client service? a reputation of being at the cutting edge? time savings? cost savings? a one-man unicorn?).

On Valuations And Multiples

Working in this space, there's often the question of whether the AI companies are really worth these valuations. Whether the hype is justified. Whether we're in a bubble.

My honest take on this: when I can configure a system in four minutes that does the work of four full-time positions - and does it faster, more consistently, around the clock, while everyone's asleep - the £25/month I'm paying is absurdly underpriced.

The real question isn't whether the technology companies are worth their multiples. The real question is whether the organisations using these tools have any idea what they're actually sitting on. And at the moment as someone who exists in this realm day in day out, I'd say most don't.

They've bought licences. They've maybe run some training. Their usage dashboards look healthy. But they haven't done the work to understand what becomes possible when you stop treating AI as a tool to be adopted and start treating it as capability infrastructure to be built.

The Regulated Spaces Question

We work across sectors, but a significant portion of our clients operate in regulated environments. Banking. Financial services. Healthcare-adjacent. Public sector.

The governance requirements are real. The security considerations are real. The compliance frameworks exist for good reasons.

But here's what I've learned: regulation is not a barrier to AI value. It's a framework for capturing AI value safely.

Tier 1 applications of AI, the ones we normally start with, exist precisely because they're designed for regulated environments. No sensitive data exposure. No additional security approvals required. Using tools that have already been procurement-approved. Working within existing compliance frameworks rather than around them.

The organisations that say "we can't do AI because we're regulated" are often the ones that haven't done the analysis. They've assumed that AI means OpenAI having access to their client data. It doesn't have to mean that at all. And the most impactful use cases most likely don't involve that level of access.

Good governance isn't a blocker to AI delivering value. AI governance is necessary; it's the structure that makes safe deployment possible. The organisations with mature governance frameworks are often better positioned to capture AI value than the ones operating without guardrails - because they can move with confidence rather than constant anxiety about what might go wrong.

What Perfect Information Actually Looks Like

Let me bring this back to where I started... with that perfect information.

Perfect information in markets remains theoretical. There will always be asymmetries. Buyers will always know different things than sellers. Insiders will always know more than outsiders.

But inside organisations? With the right configuration?

We can get closer than we've ever been.

The information that used to evaporate can persist. The knowledge trapped in senior leads can become queryable. The context that lived only in one person's memory can become organisational memory. The friction of searching alone - that 1.8 hours per day per person - can approach zero.

The discount we've always applied to every organisational calculation - the discount for imperfect information, for things people don't know, can't access, can't find in time - can shrink dramatically.

And if your competitor makes that investment and you don't... if their people walk into meetings with full context while yours are still scrambling to remember what was discussed last time... if their institutional knowledge compounds while yours evaporates with every departure.

That competitive gap will compound too.

The Position I Keep Coming Back To

AI isn't overhyped. If anything, it's underhyped - but for the wrong reasons.

The headlines focus on capabilities. What AI can do. How impressive the technology is. What's coming next.

The reality is that the capabilities have been there for two years. What hasn't been there - for most organisations - is the human infrastructure to use them.

The training. The governance. The tiering. The understanding of what works, what doesn't, and how to tell the difference. The organisational context that turns a tool into capability.

That's the work. Not procurement. Not implementation. Not "digital transformation" as a project with an end date.

Building the human systems that make technology valuable. Systematically. Continuously.

The washing machine works. The question is whether you've got a plumber or whether you're still calling Bosch wondering why the water's on your floor.

The organisations that figure this out will look back at this moment the way electrified factories looked back at the ones still running on steam. Not because the technology was secret - everyone had access to it - but because some organisations understood what it meant and others didn't.

We're in that window right now. It is a bit scary; but it might be exciting too.

Mackenzie Howe is Co-Founder and CEO of Atheni.ai, working with organisations to build AI capability as human capital infrastructure rather than technology projects. MSc Human Capital Economics (King's College London), MIT certification in Applied Generative AI.

Research Referenced:

Productivity & Information Friction:

AI Failure Rates & Root Causes:

Human Performance & Learning Curves:

Adoption & Governance:

Case Studies: