An open letter to the C-suite: your Agents will get along as well as you do
Your front office has to model the coordination it expects from its agents.Here’s the argument in three sentences. A better model won’t save you, and neither will a bigger budget; what pays off is your functions owning their data and agreeing on what the shared numbers mean.That agreement is a…
Your front office has to model the coordination it expects from its agents.Here’s the argument in three sentences. A better model won’t save you, and neither will a bigger budget; what pays off is your functions owning their data and agreeing on what the shared numbers mean.That agreement is a behavior, not a purchase: you can’t reorg your way to it, and you can’t buy it. Get it, and your agents answer questions no single function could; skip it, and you get what most companies got last year.Which wasn’t much. The average company scrapped 46% of its AI proofs-of-concept before they ever reached production. And the payoff stays rare even past the pilot stage. McKinsey finds only 39% of organizations report any EBIT impact from AI at all, and most of those put it under 5%. That’s the yield on a technology every board has made a priority.Three-quarters of companies now have a Chief AI Officer, up from one in four last year, and 83% of CFOs plan to raise AI spending more than 15% within two years. A new title and a bigger budget are fair moves, but neither gets two functions aligned on the same goal. That comes from the top. If the leaders aren’t speaking the same language, neither will the agents.Do not underestimate the time spent on creating and maintaining processes, subprocesses, and workarounds involved in reconciling the numbers in your QBR. I’ve participated in these processes enough to know the teams are burying their actual opinion in a hidden risks and opportunities file, then figuring out how to give you the answer you want to hear. Is that how you want your teams spending their time?Say it concretely. Done right, one function owns each important number, publishes it in a form the others can use, and everyone else consumes that instead of standing up their own version. You license the model (three vendors will sell you a better one next quarter), but the data your functions produce, define, and stand behind is the asset that carries from one model to the next.NIST calls this the Measure function, and it’s what keeps trust from evaporating when the model changes. This isn’t a case against moving fast. It’s the only thing that lets you move fast twice.Start with the mess, not the modelNIST calls this the Map function: know which data lives where, who owns it, and where the definitions contradict. Companies skip it and build anyway. In a survey of 600 data leaders, 43% named data completeness, quality, and readiness among the biggest obstacles to moving GenAI from pilot to production.The mess is what the model inherits. In the forecast review, someone asks how many active customers we have. Three functions give three answers, each right by its own definition and useless next to the others. Every request for a new cut of a metric, a new report, another data product carries two hidden costs. One is operational.You maintain the change, retrain the people who rely on it, and fix the downstream reports it breaks. The other is the friction between the functions it forces to reconcile their definitions again. You’d rather those people focus on new ways to drive ROI for the business.As leaders, agree to test the three steps below and get a readout on the churn it creates:Name an owner for one high-value domain. Start with customer or regulated data, and keep the first pilot small. The owner answers for accuracy the way they answer for a P&L; the platform team still runs the pipelines. The hard half is cross-functional. Finance, sales, and ops sit at one table and agree this is now the source of truth.Settle the definition before the model touches it. One golden record per entity is a central goal of master data management. Agree “active customer” before the pilot runs, not after the board asks.Put the definition in the workflow, not a deck. The owner answers for the definition and tells downstream consumers when it changes; the rule travels with the data instead of sitting in a document nobody opens.Rented models, owned dataModels are rented, and the rent keeps falling. Epoch AI tracked what it costs to reach a given level of model performance and found the price dropping about 50-fold a year at the median, with the steepest declines arriving after 2024.The model is the fastest-depreciating line item you have; the data underneath it is the only part that appreciates. That’s where the ownership belongs, and ownership means a person rather than a system. Someone answers when the number is wrong.Ownership pays. McKinsey found a mining company that moved to domain-owned data products built use cases seven times faster than before, while raising data stability and reuse.The label for the pattern keeps changing under you; it was “data mesh” a few years ago, it’s “context” this year. Ignore the label. The durable unit underneath doesn’t change. A function owns the data and answers for it.Feed a strong model conflicting copies of the same data, and it returns a fast, confident answer that’s wrong. The questions worth the most can’t be answered inside one function. An agent takes a customer count from sales, revenue from finance, and account status from customer success, then combines them into one answer.It’s only as good as the weakest agreement between the functions it draws from. If sales and finance don’t count an “active customer” the same way, that answer absorbs the gap and reports it as fact.Three commitments, and the second is the hard one:Reuse before you rebuild. When another function already publishes the number, consume it instead of standing up your own copy. Reuse is what compounds one data product into leverage; a fresh copy is just another thing to maintain.Cede the definition. When another function’s version is authoritative, use theirs even when yours is more convenient. Ceding the definition is the whole trust.Rent the model, own the data. Swap models as the price drops; put the durable investment in the data layer underneath, the only part that appreciates.Govern: the function that spans the othersMap and Measure don’t hold themselves together. The NIST framework has a function for that too. Govern. It doesn’t sit beside the others because it cuts across all of them. Define the discipline once and every pilot after that inherits it; skip it and every pilot re-argues it from scratch. NIST drew the shape for risk. It works the same for value.That’s the line between the unity you architect and the unity you announce. An announcement resets every quarter.Watch where the market is spending to fix this. The purchase of the year is the control plane, the gateway that routes every model and enforces policy on every prompt. Useful, but it governs the most replaceable components you have. What doesn’t move is the owned data underneath, and that’s the part no gateway sells you. A control plane can’t make two leaders agree on one number; only the two leaders can. That agreement starts at your table.Start with two leaders, then scaleThe pilot is really one relationship. Start with the two functions that argue the most and matter the most, sales and finance on the customer. Certify the north star metric that governs the rest, then the KPIs that roll into it, then the supplementary metrics beneath those. Each one you settle narrows the arguments under it.Follow that tree far enough and it stops being two functions. The metrics underneath pull in ops, marketing, and customer success. That’s the Map exercise, done by walking the territory instead of inventorying it first.Scaling doesn’t mean running that pilot again, once per domain. The first domain sets the pattern: who owns the number, how its definition changes, how the functions that use it get told. Write that down once and every domain after inherits it.McKinsey’s data-mesh work calls this federated governance, though the mechanism is plainer than the name. The leaders who certified the first metric write down what they agreed to, and the next two start from that instead of from zero.What compounds is the relationships, not the software. Each reconciled number is one producer and its consumers agreeing once and reusing that agreement from then on.The second domain is easier than the first, because the standard already exists and the trust is already paid for. Do this across the functions that matter, and you land somewhere specific. The numbers a cross-functional question needs are already reconciled, so the agent can answer it.This is also how the model you rent gets better without swapping it. The real upside isn’t a cleaner answer to an old question. It’s the questions you couldn’t ask before, when the metrics they needed lived in three systems under three definitions.Ask one:“Which customer segment should we invest in next year?”Churn is where this question breaks, and it’s the classic case. Marketing counts unsubscribes, customer success counts product inactivity, finance counts lost revenue. Three honest numbers, one word. On unreconciled data, the agent joins them anyway; its churn figures count unsubscribes, and its margins come from a finance model that buckets “segment” differently, so its numbers describe different groups of customers than its conversion rates do:“Segment B converts at 4.2%, churns at 6% a year, and carries a 34% margin. Segment D converts at 2.1%, churns at 9%, and carries an 18% margin. B leads on all three. Concentrate next year’s spend on Segment B.”On reconciled data, same question, same model:“Segment B converts at 4.2%, loses 11% of its revenue a year, and carries a 12% margin. Segment D converts at 2.1%, loses 5%, and carries a 41% margin. Conversion favors B; the economics favor D. Concentrate next year’s spend on Segment D.”The first answer isn’t a malfunction. Every number in it is real somewhere in your systems; they just belong to different definitions wearing the same labels. The second exists only because marketing, finance, and customer success settled what “segment” and “churn” mean before the agent ever ran.But what about…Five objections come up every time, and they’re fair.“Isn’t this circular? Domain-owned data needs the cross-functional cooperation it’s supposed to produce.” Partly, yes. Data products don’t manufacture cooperation from nothing; leadership still has to decide the functions will cooperate. What they do is make that cooperation compound instead of evaporating between offsites. The alignment is the price of entry; the architecture is what keeps you from paying it again next year.“This is just master data management with a new name.” The lineage is real. What changed is where accountability sits. MDM was a central team’s backlog the business waited on; this puts the ownership, and the blame, inside the function that generates the data. Same plumbing, different org.“Federated governance sounds like a committee, and committees produce documents, not decisions.” That’s the most documented failure in this space; Thoughtworks’ review of mesh implementations names committee reliance the most common governance mistake.The committee fails because it’s asked to manufacture agreement between leaders who haven’t agreed on anything yet. Sequence is the fix. Two leaders reconcile one real number first; the standard gets written down from that working case, and the group’s only job from then on is recording agreements, not producing them. A committee can’t substitute for the behavior. Nothing can; that’s the point of this letter.“Our security posture won’t allow this. Least privilege exists for a reason.” It does, and sharing runs under controllership, not around it. But least privilege was scoped for data that sat in one system, and if every request now routes through a ticket queue, the control is the bottleneck.Decide access when the product is published, not when someone asks. That moves the work onto the owner once, instead of onto a queue forever.“We’ve seen data-product efforts stall. Why would ours be different?” Most do stall; Gartner expects 80% of data-and-analytics governance initiatives to fail by 2027. Gartner’s stated reason is the absence of a crisis urgent enough to force the change; ours is adjacent. They stall when the sharing between functions never starts. Someone bought a platform and stood up a governance team, and the sharing never began. It works only to the degree the functions actually share.None of this is easy. Easy things don’t create advantages.Try this in your next leadership meetingStrip away the frameworks, and this letter asks one thing: model the behavior you want from your teams and agents. Run the company on numbers you hold in common, and do it visibly, because the next tier of leaders sets its norms by watching yours. That was worth doing before AI. AI just set the price of not doing it.Ask each leader at the table to write down the number your functions argue over most. Compare the answers. If they match, you’re ahead of most companies in this letter’s statistics. If they don’t, you’ve just watched the failure happen in one minute, with your own numbers.The comparison tests the room. What you do next tests you. Name the number, name the owner, before the meeting ends. That’s the pilot.Cohesion cascades from the top down. Your enterprise AI will work across functions exactly as well as you do.Sources: S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning (2025); McKinsey, The State of AI in 2025 (Nov 2025); Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks” (Mar 2025); IBM Institute for Business Value, “The rise of the chief AI officer” (Apr 2026); Bain & Company, CFO AI investment survey (Apr 2026); NIST, AI Risk Management Framework 1.0*; Informatica,* CDO Insights (Jan 2025); McKinsey, “Demystifying data mesh”; Thoughtworks, “The state of data mesh in 2026”; Gartner, D&A governance prediction (Feb 2024); IBM, “What is master data management?”.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!An open letter to the C-suite: your Agents will get along as well as you do was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI