AI Adoption Gap in CRE: Start with Trustworthy Basics

Your AI pilot looked great in the demo. It died on contact with real operations. Most enterprise transformation teams are tracking the wrong operational metrics entirely. They report on seat utilization, prompt volume, and active licenses. Then they wonder why the finance committee refuses to fund phase two. This is the structural failure pattern I call Adoption Theater. When you drop an expensive AI mandate into an execution environment with broken workflows and ambiguous decision rights, the technology does not fix the systemic friction. It scales it. The operating partners I talk with ask one question now: what will this do to improve the P&L? They do not care about your implementation roadmap or vanity telemetry. If you cannot point to a direct operational delta, you do not have a business case. You have an expensive software trial. Real transformation governance requires auditing for benefit outcomes like these: * Contractors eliminated from the back-office run rate within ninety days. * Customer support resolution times cut by hours, directly reducing headcount requirements. * Order processing error rates dropped below the historical baseline, preventing revenue leakage. * Chargeback volumes reduced through automated cross-functional validation loops. AI doesn't fix broken operations. It can often highlight the flaws. Clean data and mature governance get rewarded with clear efficiency gains. Systemic chaos simply gets amplified at a significantly higher weekly burn rate. If your AI strategy requires changing human behavior before it shows a return, you built a flawed strategy. The tool should make tangible business outcomes easier to accomplish. Start measuring how much legacy operational debt the tool systematically destroys. If your AI initiative cannot survive a basic P&L stress test, your operating design is the bottleneck, not the software.

Goldman Sachs just published its definitive report on the AI buildout. It names every gap except the one that matters most. The report sizes the buildout at 7.6 trillion dollars through 2031. Then, in the financing section, one line that does the real work: "the technology is arriving faster than the capital built to fund it." Read that as an operator, not an investor. Goldman calls this a capital problem. It is also an execution problem, priced in dollars. The report maps where money moves easily and where it stalls: → Digital infrastructure. Fundable today. Bonds, private credit, infrastructure debt. → Industrial AI software. Fundable through public markets and M&A. → Physical AI and robotics. Venture equity only. No one has built the structure to underwrite it yet. Same pattern at every layer. The ambition is fully formed. The mechanism to deploy it is missing. Here is the part a capital report frames from only one side: A financing gap and an execution gap are the same gap. Seen from opposite sides of the table. Capital cannot fund what it cannot underwrite. Enterprises cannot ship what they cannot govern, own, or measure. Same failure. Different vocabulary. This has a name on the enterprise side too. The AI Execution Gap. The distance between AI strategy and production accountability. Ownership, evaluation, and governance are the unglamorous structure that turns a plan into something that survives the real world. Goldman is pricing that gap in the capital markets. Your board is living it in stalled pilots. One question for your next budget meeting: When you approve AI spend, are you funding a strategy, or a structure that can actually deploy it? If the honest answer is the strategy, you just found your gap. 💾 Save this before your next AI budget conversation ♻️ Repost to the leader in your network who signs off on AI spend 🔔 Follow Gabriel Millien for daily insights on closing the AI Execution Gap

The demo cost twelve cents. The production bill arrived like a plot twist nobody asked for. A team shipped an AI assistant that summarized customer tickets. In testing, it was cheap, fast, and delightful. Everyone celebrated. Then real usage kicked in: longer conversations, retries after timeouts, users pasting entire email threads because "the AI is smart, it'll figure it out." Token usage tripled in a month, and finance started asking questions that engineering didn't have graphs for. Nobody had budgeted for the fact that AI systems don't fail loudly when costs spiral. They fail quietly, on an invoice, weeks after the decisions that caused it. The idea worth remembering: cost in AI systems is a design decision, not an accounting afterthought. Every retry, every oversized prompt, every "just send the whole document to be safe" choice compounds silently. Experienced teams treat token usage like latency or error rate: something you monitor in real time, not something you discover at the end of the month. The teams that get burned are the ones who only think about cost after finance starts asking why the AI budget looks like a rocket launch. The business consequence here is real. Uncontrolled token costs erode margins on AI features that were supposed to save money, not become a new line item nobody can explain. They also create a dangerous incentive: teams start throttling quality to control cost, quietly making the product worse without telling anyone, because nobody wants to be the one who explains the invoice. The trade-off is uncomfortable but simple. You can optimize for flexibility and let the model handle messy, oversized inputs gracefully, which costs more. Or you can enforce strict input limits and structured prompts, which is cheaper but less forgiving of real-world chaos. Every organization picks one of these by default, usually without realizing it, until the bill shows up. Who in your organization actually owns AI cost as an operational metric, not just a finance line item?

A useful cautionary tale below. Build-vs-lease assumes you can afford to develop systems. Most can't — so the real discipline is leasing defensibly. The under-priced risk isn't even the token cost — it's the sunk cost of the human interfaces SaaS builds over time. By the time a tool's embedded in how staff actually work, migrating means rebuilding trained workflows, not just re-plumbing an API. That's the lock-in that moderates the 'we'll switch if pricing moves' assumption....

As we approach the weekend⚡ , it's worth reflecting on the ongoing conversation about the cost of AI, particularly regarding token pricing. In Benedict Evans' recent essay, "Ways to Think About Token Pricing," he addresses a challenge that many companies are currently facing or will soon encounter. While AI is becoming easier to experiment with, the underlying pricing model can quickly become complicated. We often discuss the capabilities of AI, but Ben highlights a crucial issue: as businesses scale AI beyond pilot programs, token-based pricing may become problematic. Key takeaways from Ben's piece include: - We are currently experiencing a supply crunch, but this won't last forever. As AI infrastructure develops, token prices are expected to decrease. - Token pricing is influenced by various factors, including supply, demand, marginal cost, and ROI, all of which are currently unstable. - Over time, it's likely that model providers will have less pricing power. AI is expected to evolve into a commodity infrastructure, shifting the greater opportunity to the applications and experiences built on top of it. Why does this matter for businesses? - Unpredictable Costs: Token-based billing complicates budgeting. According to Gartner, 62% of IT leaders cite unpredictable monthly costs as their primary concern with AI pricing. - Inefficient Implementation: Without proper controls, teams may waste tokens on redundant or verbose prompts, leading to inflated costs without added value. Choosing the right partner is essential. see Bridgenext - Strategic Misalignment: Opaque token costs can obscure ROI, making it challenging to connect AI spending to measurable outcomes. So with all of that in mind, while token pricing may work for initial experimentation, it becomes increasingly difficult to manage as AI integrates into teams, workflows, and customer experiences. If AI infrastructure transitions to a commodity, the true value will lie in what companies develop on top of it, rather than merely the number of tokens consumed. Ultimately, AI must deliver measurable value without becoming an unmanageable cost center. The key to success for any company at any stage of AI investments is to do what you have always done in terms of rolling out programs. Define what success looks like. Document expected impact (both positive and negative), and commit to tracking and socializing the actual cost and impact.

The most significant AI story in M&A this year is not a model launch, as is part reason why I bang on about AI so much - to get businesses up and running with it to see both benefits and risks. The FT reports that Bain has been using AI to build working replicas of the software of takeover targets during due diligence. Hundreds of prototypes since 2023. What started with a specialist engineering team is now standard kit for ordinary diligence teams. And it has already impacted at least one deal. A PE investor told the FT that a Bain-built replica of an analytics platform contributed to their firm walking away from the bidding. Every software investor, founder and board should pause on that. For years, a lot of B2B software value was priced on the product itself. Feature depth, workflows, clever code. The new diligence question is blunt: if AI can rebuild a passable version of your product in days, what is the buyer actually paying for? The honest answer has to be something a replica cannot capture. Proprietary data. Distribution. Embedded workflow. Switching costs. Regulatory position. A customer base that genuinely depends on you. If the answer is “we are feature rich”, that is thin ice. A prototype built in during diligence by AI is not a business. It has no customers, no winning team, no trust, no support obligations, no years of workflow scars. You could clone the features of most enterprise platforms and still never break into their market. Which means the replication test only tells you half the story. It shows what can be copied. It cannot tell you whether customers would actually leave. That is why the two tests belong together. Run the replica to find out where the product value really sits. Then talk to customers to find out how deep the dependency runs. One without the other gives you a false read, in either direction. For acquirers, this should now be a standard part of commercial diligence. For software companies, run it on yourself before a buyer runs it on you. If your product can be copied, what cannot be? Questions? Don’t hesitate to ask.

Build vs. Buy for AI in CS isn't a strategy debate. It's a decision about which cost you'd rather deal with: speed now or governance later. Buying first, for a common workflow, gets you into production faster. And production is where the real learning happens, which signals actually matter for your accounts, not which signals look good in a demo. Here's where it gets more complicated than "buy for speed, build for differentiation." Building also hands you a governance problem you didn't have before. Every build decision means you now own the audit trail, the drift in decision logic over time, and who's accountable when the model gets something wrong. Buying inherits a governance model someone else already stress-tested. Building means you're writing that model from scratch, in production usually before anyone's realized that's what they signed up for. So the real tradeoff isn't build vs. buy. It's speed vs. differentiation vs. who owns the governance debt. Buy first where the capability is genuinely common. Build first or don't buy at all where it's genuinely yours, and you're prepared to own what breaks. Skip that last part, and "build" quietly turns into a maintenance program nobody budgeted for.

Tokenmaxxing-Cost of Tokens-BudgetWipeout Vs values delivered: Tokenmaxxing is not alone the problem but cost of the token is. 1. Today, most of AI Tech biggies keep the cost of token low. They are not profitable and may take years and years to recover the cost 2. Most of the big AI Tech are going for IPOs this year or so. Means they have to answer to their shareholders. This means THEY MUST INCREASE Token cost. 3. Already, due to Tokenmaxxing, many companies face budget crunch as most of them are already eaten within 6 months. The Finance Departments are scared! They ask What is the Value and "Not Much" is the answer. Hence the perception games begin now among corporates. 4. Many companies have already weighing in AI Vs Human options Vs value delivered. 5. When these new IPOs roll out, expect significant raise in token cost. The companies finance department will ask tough questions. The board will scrutinize further for stakeholder values So, The Chandrasekhar limit (refer to my previous post), the 4 or 5 neutron stars are getting too closer due to AI Business Gravity... When they will go critical is something of a deep concern. My advise is to keep your core team in tact, while keep the critical AI based projects live. Shut down all other AI usage to optimize of cost. Wait for the C-Limit to pass by and make calls. So I suggest the following: 1. Do not use AI for recruitment for next 3 years. 2. Use AI for Finance, Core Mfg, and Pharma - only for areas where you want deep stakeholder value 3. Use AI in critical Healthcare programs and Retail sectors 4. For Tokenmaxxing people/programs, put in a audit or review mechanism to assess value delivered Refrain from training everyone in AI. Provide limited access to core projects to see value. Establishing Initial success is key here.

If the stock market is the benchmark, your AI strategy needs to beat it. If the current hurdle is 8–11%, then the company goal should be 12%+ growth from AI-enabled work. That is the bar. BCG’s real point is simple: token costs are not just an IT problem anymore. They are becoming a CFO problem. As AI moves from experiments into production, token bills get buried in software budgets while agents, workflows, and customer-facing products quietly drive usage higher. The company may feel “AI-native” because everyone is using the tools, but the finance team still cannot answer the most important question: Are these tokens creating return? BCG frames this as cost per successful outcome. Not tokens used. Not AI activity. Return on AI. That is the part every CFO should care about. Tokens that build reusable capability can look like investment. Tokens that run internal workflows are operating expense. Tokens inside customer-facing products are COGS, which means they need to be managed against gross margin. That distinction matters. A support agent that creates more escalations is activity, not return. A marketing agent that generates assets no one uses is activity, not return. A customer-facing AI feature with rising inference costs can look innovative while quietly weakening margins. This is where Owned Intelligence becomes practical. Owned Intelligence is not just “we own our AI.” It is the operating layer that lets a company see, route, govern, and measure AI work. It connects the model to the company’s data, workflows, permissions, context, business logic, and financial accountability. When you own that layer, you can start asking better questions: Which workflows are producing measurable return? Which AI features are protecting margin? Which tasks should use a lighter model? Which context should be cached and reused? Which agents should be stopped because they create activity, not outcomes? Which parts of AI spend belong in investment, opex, or COGS? That is the shift. The companies that win with AI will not be the ones with the most subscriptions. They will be the ones whose CFOs can prove which workflows got faster, which products protected margin, which teams created more output, and which revenue opportunities became possible because company knowledge finally connected to action. If your AI strategy cannot beat the market, it is not a strategy yet. It is an experiment.

Every AI deployment has two costs. The first is the implementation cost. This one gets modelled carefully, scrutinised by finance and presented to the board. The second is the ongoing operational cost - licensing at scale, model maintenance, data quality management, integration upkeep and human oversight. This second one is routinely underestimated. Sometimes it doesn't appear in the business case at all. An implementation that looks compelling on a one-year view can look very different on a five-year view. And the board that didn't ask the question at approval will certainly ask it at the first budget review. Build your business case to include both types of costs and present both to the board. The conversation might be harder in the short term but will be significantly easier in the long term.