Welcome to the strategy workshop

My dear enterprise AI friends. I think you all would recognize the exact model of windowless conference space that was named after some mountain that no person present in the meeting you’re at has ever climbed. There are twelve executives, four suited up consultants with big watches, two architects wearing typical German glasses, one data scientist with lots of notes and one visibly exhausted program manager who all gather at nine AM sharp to try and decide how AI will reshape their company for the next five years, and at least three of the twelve believe that Microsoft Copilot is itself a strategy. This is considered a strong starting position.

The program manager opens his plastic Windows laptop and the screen shows a PowerPoint slide titled “Our AI Journey” and it is featuring a winding road disappearing into a vage sunrise, with along the sides lots of sign posts pointing to the text besides it. And this apparently comforts everyone into believing that a destination exists somewhere past the quarterly budget. Then the hired consultant welcomes the room to a “highly interactive co-creation session,” and that – as you well know – is corporate lingo for “no one has actually bothered doing the thinking yet”, though there are now enough people present to distribute the blame attractively. He explains that the goal is not to jump to solutions (Copilot) immediately, a sentence that visibly wounds the vendor reps who just arrived with seventeen of them, all carrying laptops with chargers and fighting for a spot in the sockets.

Instead the group establishes ‘vocabulary’ first, then aligns on ‘ambition’ and identifies ‘opportunity spaces’, three verbs chosen specifically because “decide what to do” fits on one slide but “align, establish, identify” comfortably fills a whole morning and justifies the catering bill the CRE exec happily offered to pay.

The opening question arrives . . .

What does AI mean to you?

The CIO says ‘transformation’. The CFO says ‘efficiency’. The CHRO says something like ‘empowering people’. The CRO says ‘responsible adoption’. But no one in the room dares say that AI is a set of computational techniques that might improve specific decisions under specific conditions, because that would risk the workshop eventually discussing a specific decision.

And we won’t have that in a ‘strategy workshop’ now will we?

There is a reason that this question produces six different correct answers. “AI” is not one technology, it is a label draped over statistics, automation, robotics, chatbots and whatever the CEO saw his grandkids show him on a Sunday morning. It’s a term that is broad enough to mean everything and specific enough to commit an organization to nothing.

By eleven the walls are papered with Post-its. And since we’ve all been there, we can predict what’s on them. . .

Automate repetitive work
Improve customer experience
Personal productivity
Digital twin, despite none of them had specified what should be twinned or why it needs a sibling at all
Quantum AI, because regular AI has become insufficiently futuristic.

Self-driving enterprise underlined twice, with no one asking if the enterprise even has brakes at the moment. The facilitator annex cartoon artist then clusters everything into five themes called customer experience, employee productivity, operational efficiency, data and insights and new business models, the same five clusters produced by every technology workshop held since 2014, and their familiarity is read as proof of ‘alignment’ with global best practice because the consultant was involved.

Then of course, an architecture diagram appears with a circle for ‘AI Core’ and satellite circles for Data, Governance, People, Partners, arrows running both directions so it doesn’t imply sequence. Its power is diplomatic since every department gets a circle, so none of them have been excluded from the future – not even the executive of the business unit at the most far-right side of the org chart, which is usually the least important. Then somebody – usually the consultant – raises the question of “what is the operating model going to be?” and the facilitator immediately files it on the parking lot, a flip chart reserved for questions that would otherwise threaten the schedule, alongside ownership, funding and which business problem any of this solves.

Then something called ‘the lighthouse project’ gets to be chosen next, and it is never the highest-value project. It usually is the project that is demonstrable within twelve weeks without touching any of the core systems, changing ownership or reality. Practical projects like predictive maintenance dies for lack of enough sensor data. A supply chain agent dies because the supply chain system is mid-transformation until the next ice age. The organization lands, as it always does, on an internal knowledge chatbot, that is named before its architecture even exists, because a chatbot threatens no power structure and it demos beautifully on answers that are inserted into the knowledge base the night before.

Four pillars close the day.

Enable people.
Transform operations.
Accelerate decisions.

And of course, my favorite. . .

4. Innovate ‘responsibly’

Which is broad enough to cover everything the company already does and everything it might ever do, which makes them nearly impossible to challenge. The roadmap runs Explore, Scale, Transform, rising left to right with no dependencies drawn between them, because a dependency would imply that something must finish before something else begins.

And when asked to share one closing word, the room offers ‘energized’, ‘inspired’, ‘aligned’. The architect says ‘concerned’. The facilitator frowns, seeing the vibe being threatened and writes it down as ‘constructively challenged’ or something along those lines. Then everyone poses for a photo beneath the words innovation, AI and transformation, a fine record of the day the company decided to become AI-driven without deciding where to drive.

Yes, my smart friend, this blog is dedicated to something called ‘Enterprise AI strategy’, or more common ‘the AI-pilot petting zoo’. And it is a record of every program I’ve seen happening since November 22nd 2022 – a.k.a. the day that Machine Learning turned into a sexy thing.

More rants after the messages

Visit the ATLAS website. It’s free. I built it for myself. And now you can use it too.

Connect with me on Linkedin 🙏
Subscribe to TechTonic Shifts to get your daily dose of tech 📰
Please comment, like or clap the article. Whatever you fancy.

Step one, buy everything

Once you’ve got the four pillars in place, the organization does the only thing it knows how to do with genuine confidence, which is to spend money. Of course not because it has determined which capability it needs, that would require analysis and hard choices and above all the risk that some vendor does not get invited to the buffet – God forbid. No, it buys because purchasing creates movement and movement creates PowerPoint decks, and decks in turn create the sensation of transformation without the inconvenience of it.

Usually vendors arrive within days, because their digital cadaver dogs detected the phrase ‘AI-first’ in a press release and they explain that the organization is standing at an ‘inflection point’, which is a conceptual location that every technology sale occupies since no man has ever gotten board approval by announcing that the market is moving gradually and there is plenty of time to think. And somewhere around slide number five, the slide called ‘The cost of waiting’ appears. The cost of acting foolishly obviously doesn’t get its own slide.

The first purchase is always productivity licenses.

These are justified by a spreadsheet created by a Dunning-Kruger type promoted to interim AI-Officer who recounts a statement made by Julian Goldie and consorts, stating that that forty thousand employees who are collectively saving thirty minutes a week yield more than a million hours a year. That is, of course, a number that no man has the courage to inspect, since it assumes that every saved minute automatically turns into productive capacity instead of another meeting. Then a business controller in finance books the benefit as ‘non-cashable productivity’, which is financees for ‘don’t expect money to show up’. How I love business controllers’ humor.

And then comes the data platform.

This investment is justified by the observation that AI needs, of course, data. And they’re conveniently ignoring that the company already owns two data lakes, several data warehouses and at least one previous single source of truth. Clearly the missing ingredient was never discipline, but the correct architecture metaphor. So, they end up with a lakehouse that combines the suffocating depth of a lake with the mortgage of a house (a data architect would probably excommunicate me for saying this). A data mesh promises decentralized ownership to departments that spent twenty years refusing to own data quality. A knowledge graph gets added after someone returns from a conference, and when the team asks which relationships to model, they are shut down and told to begin with an enterprise-wide ontology, which is how a twelve-week initiative turns into a six-year philosophical inquiry into the meaning of the word customer.

Months later the executive committee asks for a consolidated view of what was actually bought.

It doesn’t exist of course, so the Dunning-Kruger officer produces a slide with fifty-two boxes and ninety-three arrows, including AI generated images about ‘Agentic Transformation’, each meaning something completely different but none of them notices. Then the sponsor – usually an older executive who has ‘seen it all before’ – asks the only honest question of the entire program.

“What can we do now that we could not do before?”.

The answer, after several rounds of ‘enterprise-grade foundation’ and ‘repeatable pathway to production’, turns out to be summarize documents. That’s a capability that took eighteen months and several data platforms to reach.

To avoid vendor lock-in the company then adopts a multi-vendor approach.

This means it had to buy overlapping capabilities from three suppliers and assigns architects to justify why, producing a model gateway, an API gateway, a data gateway, a security gateway and an AI gateway, each one protecting the enterprise from the risks introduced by the other four. A request now passes through identity checks, policy checks, filtering, loss prevention, routing, moderation, HITL assignment, and audit logging before it reaches the model, which answers below two nanoseconds, but the surrounding infrastructure and harness takes six weeks to approve the connection. This is described as agile.

Then vendors perform their sneaky little sales trick called ‘Innovation credits” that’s supposed to sweeten the early spend. Vendors like SAP, Microsoft and friends all present this as generosity, but in practice it’s functioning more like seeds scattered across the garden, and by the time the prototypes have sprouted and absorbed internal data and executive sponsorship, moving them elsewhere or burying them would require excavating the building.

No single purchase looks catastrophic on its own, a million for licenses seems manageable, several million for the platform development gets classified as foundational CAPEX, and together they form a large, permanently hungry creature living inside the operating budget, one that eventually starves the actual use cases competing for what crumbs are left.

The company ends up standing inside a magnificent technological cathedral built from subscriptions, slowly sucking it dry, and at the centre of it is one lonely knowledge chatbot, that is patiently answering two questions a day about a travel policy using a document that expired two years ago.

The vendor logo architecture

Some time after the dopamine releasing shopping spree, someone finally asks to see what the architecture is all about, a reasonable request, and it should produce a document that describes how components work together and what happens when something catches fire at three in the morning. But what usually arrives instead is an architect who is opening PowerPoint and arranging a few vendor logos with lots of thin blue arrows between them.

Microsoft is of course positioned near the centre, not through deliberate conclusion or some other thing that would make sense, but simply because the architect started from a Microsoft template he reused and never quite escaped it. Databricks floats somewhere above or below – depending on the preference – performing an almost spiritual purification, raw data enters from the left through three arrows labelled ingestion, transformation and feature engineering, and clean, governed, analytically obedient data exits on the right. The diagram does not show the twenty-seven versions of the customer table, the field called status_final_v2, or the business rule that lives only in the memory of a contractor who is currently on sick leave. It does have three arrows and a dream, that dream being that once enough data goes through the platform, intelligence emerges like green mold does when left to digest forgotten food, only difference is that is has stronger executive sponsorship.

SAP always is positioned underneath everything like it’s some sort of a geological layer, it is rarely drawn attractively but it is omnipresent, and every ambitious recommendation eventually collides with a valid cost centre, the correct G/L account, a transaction code and a configuration decision made in 2004 by someone now living peacefully beyond the reach of corporate email. Governance occupies a teensy weensy box in the upper right, connected to everything by dotted lines that imply influence without demonstrating control, and inside that box you can read the phrase HITL which stands for ‘human in the loop’, a universal solvent for architectural anxiety. By the way, no one would dare ask which human, which loop and with what information, or if that human gets seven seconds to review two thousand model-generated recommendations before the queue moves on.

The finished diagram is of course visually stunning and always structurally empty with seven layers, twelve domains, fourteen vendors and more arrows than an airport evacuation plan, and visual completeness is one of the most effective illusions in enterprise technology, because when every box has a label, then the mind assumes the system has been understood. But what it actually shows is where products have been bought, but not how intelligence works, nor which decisions the enterprise intends to improve. It gets approved, drawn on a white wall as some cartoonesque journey, and saved as AI_Target_Architecture_FINAL_v19_Approved_Updated.pptx, a filename that describes the architecture’s entire lifecycle more accurately than the diagram itself.

Strategy by Gartner quadrant

Eventually someone has to pick a platform, and every principle about openness and business fit that was formulated in the strategy workshop leaves quickly through a side door, because platform selection is always political and mildly theological (just watch the debates between the architects). It is requiring faith in a vendor’s roadmap and a continued willingness to answer support tickets.

But fortunately an analyst firm has already arranged the market inside a comforting geometric shape.

Vendors get sorted into Leaders, Challengers, Visionaries and Niche Players, and this classification carries instant emotional weight. Because Leaders feel safe. Niche Players feel like they should apologize for attending, regardless of whether their narrow specialty happens to be exactly what the organization needs. The CIO asks the architecture team which Leader they recommend, and asks again when the answer starts with “it depends on the use case,” because that question can be resolved before lunch while the honest one cannot.

A consulting team is then hired from one of the ‘applied consulting’ firms, and a requirements matrix follows, usually with hundreds of criteria where each vendor gets scored, weighted and multiplied into totals with two decimal places.

The preferred vendor wins by 1.7 points and not a soul can explain what those 1.7 points actually mean, and the vendor happens to already be a Leader, which gets read as confirmation instead of a sign that the evaluation was built to reach the safest possible answer.

Analyst briefings add the latest buzzwords like sovereign (Europe), multimodal (because why not), agentic, and orchestration to the vocabulary, and the RFP that follows asks every vendor if their platform is scalable, secure, interoperable and human-overseen, and every vendor answers yes, because technically anything can integrate with SAP if the budget is sufficiently spiritual.

Then the demos start . . .

Every demonstration takes place inside a tiny artificial universe in which the data is clean, the user asks exactly the right question and no one has recently changed a policy without telling IT. The vendor then types a prompt that has survived forty rehearsals, the model produces the correct answer, and the room applauds as though intelligence itself has just been invented instead of a preselected scenario being executed by an exhausted sales engineer who spent the previous evening removing every possible hint of reality.

Then an executive volunteers a spontaneous question, which causes the sales team to experience a brief collective cardiac event. The vendor smiles, explains that the question is excellent, and redirects it towards a nearby use case that the system has actually been trained to survive. If the answer still fails, the presenter calls it an integration issue, a data-readiness issue or an opportunity for co-innovation, because no enterprise demo has ever failed when it can instead reveal a strategic ‘collaboration’ (read: you are now responsible for our roadmap).

All demonstrations run on clean scripted data that resembles the organization like a showroom kitchen is a representation of an actual kitchen during kids rush hour.

A smaller specialist vendor with an asterisk implying a ‘wildcard’, then happens to show an actually better fit for one critical use case, easier to integrate and purpose built, but then procurement asks what happens if it gets acquired, then what happens if it disappears off the map, with far more intensity than the identical question aimed at the platform giant already holding the company’s identity, data and office suite, hostage. The risk committee then labels the specialist unacceptable concentration risk and selects the vendor that already controls everything, who gets classified as ‘reducing risk’.

A year later the platform’s quadrant position drifts, a competitor moves further right, a proof of concept runs on the new option, and the organization now owns two platforms, which is now called ‘avoiding lock-in’. Meanwhile an engineer builds an open-source fix for the original problem in three weeks, without telling anyone. It is of course barred from production for not being part of the strategic landscape.

The Copilot delusion

The rollout begins with a message about smiling colleagues and floating interface elements, promising that software will liberate thousands of trapped minds from composing meeting invitations and automating tedious work. The business case starts with multiplication, forty thousand employees saving thirty minutes weekly recovers more than a million hours a year, a number that grows theoretically but not financially because it exists in presentations but not in a bank account.

Executives get early access so they can champion the tool visibly. A consultant turns three bullet points into six polished paragraphs containing the phrase “as we embark on this exciting journey together,” and the executive’s own assistant, who has done the identical job for years without inventing strategic priorities, watches in professional silence. The rollout email warns employees not to enter confidential, personal, financial or commercially sensitive data, which is the exact information that would make the assistant useful, leaving it free to help with tasks that are safe and largely beside the point*

Most people begin by using the thing to help them rewrite emails. One employee turns three bullets into a paragraph. The recipient turns the paragraph back into three bullets. The company has now paid twice to reconstruct the message it already had. A Microsoft supplied dashboard tracks active users, prompts, documents generated and potential ‘savings’, numbers that climb steadily because usage is mandatory since there is no alternative, and it is easy to count, and whether any decision improved, any customer got happier, or any process got shorter stays entirely uncounted.

Meeting summaries become the star feature, since the company already spends much of its life speaking into rectangles, and the assistant produces a record of decisions and actions even when the meeting contained neither, because the template expects negative entropy to have occurred during the meeting. Search runs into the same wall, an employee asking about travel policy gets a fluent answer stitched from five contradictory documents, one expired two years ago but ranks well because its formatting is nicer.

Whenever the tool fails to create value, the organization prescribes more training, regardless of employees distrust its results, misuse it, ignore it or use it constantly without achieving anything measurable.

The organization creates an AI champion network, whose members receive badges, newsletters and additional responsibilities, but no additional time in which to perform them. Leadership then begins treating adoption percentage as the primary measure of success because it is easy to compare across departments, even though it says almost nothing about value. A department in which hundreds of employees use AI to make emails sound friendlier may therefore appear highly progressive, while a small engineering team that uses AI less frequently but cuts defect rates in half may appear resistant because its usage numbers are lower.

A year in, the value review confirms that adoption is rising and satisfaction is high, and that direct financial impact remains hard to isolate. That is a phrase that describes the predictable result of buying a general assistant before deciding which work should change. The licenses get renewed regardless, since stopping reads as retreat and continuing is explained as scaling.

The tool turns out to be useful to a certain degree, but actually pales in comparison to what the users have at home, but that is barred from use. It drafts, it summarizes, and removes blank-page anxiety, and the delusion was never about its value, but that distribution alone equals transformation.

Real transformation would mean removing activities and killing outputs that you don’t need, and a license does none of that by itself. It has simply entered the organization and learned to take minutes.

*_According to Toms Hardware Microsoft updated Copilot Terms of Use, stating that “the AI large language model is designed for entertainment use only, and users should not use it for important advice”. Microsoft later on removed the ‘entertainment only’ phrasing after receiving backlash from the public (duh).

The pilot petting zoo is open for business

My dear friend, I think you have lived through enough of these programs to know the smell before you open the door. Fresh coffee, printed roadmaps, and the vibrant optimism of a room that has not yet been told the truth. Every enterprise AI program eventually builds a pilot portfolio, a carefully curated set of small experiments meant to demonstrate potential but it is structurally forbidden from becoming useful.

The first resident of the zoo is almost always an HR chatbot. It is answering holiday and travel-expense questions flawlessly in a demo, and none of the participants care to ask what happens when the employee works in Belgium, reports into Germany, sits on a Dutch legal entity, and has just been seconded to Singapore, because that one employee would send the bot into a request for political asylum. The second is usually something like an ‘invoice classifier’ and it is boasting a whopping ninety-six percent accuracy. It’s a number that survives right up until someone finds out that the test set was built with fifty invoices, forty-eight sharing the same template, the remaining two classified as wildlife, just for the sake of having two safe outliers.

Here I have to pause dripping acid, because this pattern deserves a real explanation, and I am still an educator and researcher next to someone eating his own dogfood. An accuracy score only means something relative to the distribution it was tested against, and a pilot that was built on curated data confirms that the model works when nothing interesting happens. This is the same mechanism behind our PASF research finding that only around 35 percent of enterprise processes we assessed actually qualify for full automation, what we call Zone I and II, while the rest run into governance walls, data quality walls, or a need for human judgement that a clean test set was built to hide. But let’s not spoil the fun any further.

Have you ever noticed that every pilot gets a cheerful human or mythological name? The HR bot is called Athena, the invoice model is called Atlas, but no one ever named one Cassandra, despite Cassandra being the one figure in the myth who was famously correct and consistently ignored, which would make her the most accurate branding decision available to the whole program.

And then, when someone says the word ‘production’, reality walks in carrying a clipboard with a checklist. Security asks about identities and access management, legal drones on about liability, finance asks when the benefits will finally show up on a P&L, and the pilot that spent its whole life in a warm sandbox eating hand-fed test cases, develops a sudden medical condition. The program calls the rebuild that follows ‘industrialization’ and it makes the cost multiply by roughly ten and twelve confident weeks balloon into eighteen humbled months. Then the pilot goes on hold and stays in the portfolio anyway, because removing it from the slide would drag down the reported success rate.

This is how the petting zoo grows, chatbots that not a soul uses and classifiers that have a lower trust score than Klarna, each is an animal with a slide, a sponsor wearing a Teflon jacket, and a profitable future that is exactly one budget cycle away . . . forever. However, eighty percent of pilots get reported successful, where successful actually means that the prototype worked during a scripted demo and that none of the participants got physically injured.

And pilots stay popular, especially in Europe, because they are emotionally safe, they generate excitement without threatening a single budget or role, and a production system can fail publicly and take your career with it, but a pilot can remain promising forever. It suggests movement, but it never has to materialize any results.

That is the entire business model of the pilot, and it is why the enterprise adores its pilots like we adore puppies before the vet bills arrive.

Governance arrives carrying a clipboard

Governance usually wakes up months after the first pilots launch, once somebody finally asks the question if systems are actually allowed to do the things the vendors demonstrated so enthusiastically to the board. It arrives holding an eighty-four-page Responsible AI Framework whose principles like fairness, transparency and accountability, are all sensible and all nearly useless to the engineer who has to decide if a model may auto-reject one specific invoice. The framework insists that humans remain accountable, without saying which human, when the model was selected by IT, trained by a vendor, configured by a consultant, funded by Finance, and operated by a business unit that took it in under the assumption that the AI Centre of Excellence owned it the entire time.

So then program builds a questionnaire in the form of an intake running on PowerAutomate with a dash of Copilot Studio and it has two hundred and nineteen mandatory fields, and question one asks if the solution uses artificial intelligence, at which point I have personally watched a project manager pause and consider answering no. The questionnaire takes longer to complete than the pilot took to build, and then an AI Review Board convenes once every four weeks, and one missing attachment can delay an initiative longer than several generations of the underlying model remain commercially relevant.

The phrase “human in the loop” gets used constantly in those meetings, and I have never once watched a single person identify the human or draw the loop on a whiteboard. And I think that this now deserves a proper definition instead of regurgitating Gartner, since it gets treated as a solved problem when it is closer to a design pattern with a hundred unstated variables. A genuine human-in-the-loop control needs a defined decision boundary, a realistic time budget per case for the human and visibility into the model’s confidence or trust score, and a review path that isn’t more expensive than the value generated by the agent.

But what people tend to forget is that, alongside HITL, there is also something called HOTL, or human-on-the-loop. Now, in this model, the system acts fully autonomous and a human supervises it from a dashboard and watches for exceptions and only intervenes when the shit hits the fan. This analogy is also why we call them ‘the poop-scoopers of AI’. The human is no longer reviewing each decision, but monitoring the behavior of the entire machine. That is considerably less comforting when you realize that the dashboard may be tracking twenty thousand actions per minute. In our research, we noticed that roughly a third of all agentic AI programs were cancelled because the cost of these interventions exceeded the value generated. If you want to know how and why, just read the previous blogs. I’ve got a few lined up below*

A genuine human-on-the-loop design needs explicit stop conditions and enough explainability and observability for the operator to understand what is happening before the system turns a small error into an enterprise-wide archaeological site that will dye the ROI from black to red. It also needs a realistic answer of how one person is supposed to supervise thousands of autonomous decisions without turning into an ornamental emergency brake. Most organizations skip those details, place a dashboard icon next to the words “human oversight” and declare the problem solved.

The acceptable accuracy threshold is arbitrarily set at ninety-five percent because it sounds serious and barely anyone would dare to run the math on what the remaining five percent means across two million live transactions. Then the AI architect adds monitoring for drift, and drift deserves a real definition too, it occurs when the patterns that a model learned during its training no longer match what is showing up in live production data, and then, when an indicator turns red the threshold gets adjusted (rather than investigated), and the dashboard returns to green.

Badabing badaboom. Check-in-the-box!

And in the end of your agentic program, you will see that governance has been growing until it turns into the largest workstream in the program, and that it has surrounded it with enough excel sheets to make its danger fully auditable. Governance is not the villain here, to be fair, the organization truly needs controls, accountability and boundaries, but the comedy begins when governance arrives as paperwork instead of engineering. A PDF cannot interrupt an agent mid-workflow when it selects the wrong tool and a written principle cannot halt a harmful transaction when it is the middle of executing. The AI system is a dynamic beast and it is wired into live business processes but the control framework that has to govern it lives exactly where it was born, in SharePoint.

*_Like here:

Summoning the ghost of ROI through séance

Every program eventually reaches a point where someone in Finance asks what all of this is actually worth. It’s a reasonable question, after 18 months of toiling, and this is exactly why the room instinctively tries to avoid answering it.

The business case usually rests on four categories:

productivity gains
strategic benefits
risk reduction
and future opportunities

And only the first legally requires actual numbers.

Fifteen minutes saved for writing an email that could have been a tweet, than gets summarized by the recipient by the same AI, is now multiplied across every employee for every remaining year of projected civilization, and not a soul dares to ask if the employee actually leaves the office fifteen minutes earlier, because enterprise value refuses to operate through anything that vulgar.

The program reports twelve million euros in productivity value. Finance then asks if anybody had seen the actual operating costs drop by twelve million, and then the room goes quiet. The controller sighs, takes a deep breath and notes ‘benefits non-cashable”. He knows the money probably exists, but in a higher financial dimension where a spreadsheet is legally recognized as a bank account.

Strategic value covers innovation, competitive positioning and future readiness, none of it measurable directly, which is exactly what makes it ideal for an executive presentation. The future has never once shown up at a steering committee to dispute a claim made in its name. Platform costs, by contrast, get treated with real discipline, because licence fees and cloud consumption are stubbornly real unlike strategic value, so the program spreads them across several years and places every uncertain benefit in the earliest quarter available.

The return improves further whenever a team counts the thirty minutes saved drafting a report but unknowingly ignoring the forty minutes a user spent checking if the model hallucinated a market trend. Employee surveys report seventy-eight percent improved productivity and forty-two percent greater creativity, but no one ever asks why creativity was ever part of the accounts payable role in the first place. And when Finance finally asks for a benefits realization plan, every claimed benefit gets assigned to a named owner who is surprised to learn they own it, and a senior manager gets picked largely for being out of the room when the decision was made.

None of this means that AI creates no value. It often creates real value, when it changes an actual decision or prevents an actual failure from happening – yeah, the stuff we did with Machine Learning back in the day. But the trouble starts when you measure enthusiasm instead of outcomes and hypothetical hours instead of real money. A genuine return requires cutting some activity and spend less or earn more, or reduce a real, real risk. Everything else might still matter to somebody, but it is not a return.

The program announces the investment has exceeded expectations regardless, licenses get renewed, and the return sits somewhere between the adoption dashboard and the strategic roadmap, floating peacefully above the enterprise like a financial spirit you can feel deeply and never once actually reconcile.

Signing off,

Marco

For my daytime job, I’m a researcher and factory builder at Eigenvector Research, a research lab operating at the frontier of unsexy AI. In the evenings, I write about why building this stuff is a lot harder than the vendors would have you believe.