View Recording: AI Budgeting for 2027: Where to Invest, What to Delay, and How to Prove ROI

Brian Haydin 0:05 Well, hello everybody. I think talking to my customers, it’s getting to be budget season. So this is a pretty good, timely conversation. For those of you who don’t know me, for those of you that don’t know me, my name is Brian Hayden. I’m a solution architect here at Concurrency. And what I want to do today is treat you all like you’re part of an investment committee, like an AI investment committee. I’m going to give you 3 different AI funding requests that you might see in 2027. You know, and then what I want to do is I want to bring the evidence that a good committee would start to look at one piece at a time. And then we’re going to watch what that does with the decision. So by the end of this, hopefully you’ll have an understanding of what kind of budgets are going to get approved and what you should be asking for. So before I get to that, just a quick word on who Concurrency is. We’re A Microsoft Solutions partner, a consulting firm, and we help organizations turn technology into measurable business outcomes. We take a people-first AI approach, and that means that we focus on helping people work smarter, make better decisions, and create better customer experiences. It also means that people stay accountable. AI brings speed, scale, and insight, but human judgment, creativity, and expertise are what turns those things into actual business results. And whether the work is with Microsoft Copilot or custom AI agents, cloud modernization, data or analytics, and maybe even process automation. The goal should be pretty much the same. Growth, better decisions, lasting competitive advantage, and more efficient operations. So if you’d like to keep the conversation going after this session, once you follow us on LinkedIn, there’s the QR code. You can find upcoming events, insights, and even the recordings of webinars like this. So the premise of this session fits into six words. The pilot was cheap and production isn’t. And that isn’t a complaint about AI. It’s just what a pilot really is. A pilot has the privilege of skipping things. It runs for a small group, it runs on a small workload, and it runs for the people who. built it and that are standing right next to it and managing it. Nobody asks it for things like an uptime commitment. Nobody asks who reviews the output when it’s wrong, or who gets the call at 2 o’clock in the morning when something breaks. Now production that has to carry all of that. And every one of those things is something that people, you know, budget. So here’s a mistake that I’m seeing most of the time, you know, and this year. The pilot that goes pretty well, and then in 2027, the number that gets built is, you know, just by taking that bill and scaling it up. And it’s, you know, like a multiplier, and that’s the pilot map. And today what I want to do is walk you through real math and something that you can actually defend. So… Okay. I want to like lay out this fictitional company. It’s A composite of all of you that I’ve talked to over the course of the year. And this company has 1200 employees. They’re A Midwest manufacturer. They have 1000 knowledge workers that are on Microsoft 365 and 300 of those users today are have copilot seats allocated. So in this fictitional company, we’ve got three requests that have been made to the committee that you guys are going to review. Request A comes from the CIO. It says renew copilot, expand it to all 1,000 knowledge workers, and add Agent 365 for everyone so that the agents that people are building are getting governed. That costs $540,000. And the case is pretty simple for this one. The people who have it don’t want to give it back. All right, let’s look at request B. That comes from the head of customer operations. They’ve piloted an agent this year that reads incoming order documents and keys them actually into the ERP. 100 documents a month and it worked. They want to run all 10,000 a month through it now, and that’s going to cost $25,000. This one, if you notice, is the smallest ask that’s on the table, but it is the only one that has a working pilot behind it. Now let’s look at request C. This comes from the VP of sales. an autonomous quoting agent that pulls pricing, inventory, and customer history, and it turns around and sends the quote out itself. Their request is $450,000, and it had the best demo that anybody in the building had seen, you know, this year. So add it all up. just a little bit over $1,000,000 and all three of these have a sponsor and a real reason for us to do it. So hold on to that. I want to ask you the same question at the end, and I’m going to try to start this poll that I set up so we can all. I don’t do this very often on the polls, but I want to start getting used to it because they are kind of fun. I do want to mention that we’re going to come back to this same poll at the end. What I want to do is… Um… is we’re going to add a fourth one on to this one, but we’ll talk about that towards the end of it. But so we had three requests, copilot, order intake, autonomous coding agent. What does everybody think? All, we got a few responses. Order that order document intake seem to be the best. All right, so. Oddly, it’s almost as if I planned this. Let’s start with the request that you probably should have, and you all did kind of vote for the request B. So let’s go through the math. Here’s how the sponsor got to $25,000. The pilot ran 100 documents a month, and the cloud bill that they measured was $280. Now, most of that was just basic search service, and the model itself only ran, it ran like $13. So production is 100 times the volume. We’re going to scale that up, right, from 100 documents a month to $10,000 a month. So they scaled the model cost by 100, and they kept pretty much everything else the same. and gets $19,000 a year. Then they did like the responsible thing that you should do and they rounded it up to 25,000. So I want to be fair to this fictitional sponsor. Every number on this slide actually adds up. The math is correct. They padded it. The problem isn’t what they what they put on this slide or what they put into it. When this workflow actually has to go to production, the business starts asking it for things the pilot never really had to deliver. And I want you to hear each of these things as a requirement because it’s really what it is. First, it had to work at 2 o’clock in the morning during the month end. So in the pilot, if it went down, somebody just restarted it after they got back on their lunch break. But in production, that means that the redundant capacity, that there needs to be redundant capacity and that there’s people that have to like service those kind of phone calls. Second, the order has to land on the ERP. In the pilot, a coordinator was copying the output over by hand. And in production, we have to do an integration and somebody has to repair it when the ERP changes. you know, so there’s going to be continuous maintenance. Now, the third thing is that nobody ships the wrong part. The agent’s going to get things wrong and people need to remember that. It’s not perfect. So a person is going to review a share of what this agent’s going to produce. Now, remember that a little bit because it’s going to matter more than anything else that we’ve that we’re going to talk about on this page. And then the fourth thing is that you can show an auditor what it did. That’s the logging, the data controls, and retention. Something I’m going to be talking about a little bit later today is, you know, agent ops, not on this session, but later tonight. Now, none of these are hidden fees. Each one is something that business would insist on if you ask them. But each one of them actually costs money and none of them is included in the $25,000 that we showed. So all right, natural question is, how do I figure that out? What’s the multiplier? There’s got to be a multiplier, right? Well, no, there isn’t a multiplier. multiplier. I looked, my teams looked, and there’s no defensible number that you can just copy and paste and say that this is what the pilot costs and this is what it’s going to cost to put it in production. What does exist, though, are three different kinds of numbers, and each of these tell you a little bit something different. So Gartner estimates that an agent uses 5 to 30 times the tokens of a chat bot for the same type of task. And the reason behind that is because agents plan, they call tools, and they check their work, and then they retry things that didn’t work the first time. So Gartner doesn’t, they didn’t publish the task sample behind that range. So I want to just treat that as like a warning and not an actual measurement. But, you know, it comes to the tokens, that’s the token number, but it’s not the total cost number. Now, McKenzie, they published a cost model this summer that, you know, for an agent that runs in a bank. And in that model, Human review is 70 to 75% of the variable cost. Tokens are about 1/4 of the overall cost. Now, that’s one illustrative workflow, and it’s a model, but not like the actual measurement. But let’s look at one that’s like a little bit bigger, too. A search service, remember we used it, we built this on top of a search service. and that had one replica that cost $245 a month. To get the kind of uptime commitment that we want for the reads and writes, you have to triple that. You need to have three instances. That’s three times the cost on one component, simply because we want to have failover and redundancy. So those are three numbers and none of those were actually in that multiplier, the fictitious multiplier that we wanted. So instead of just a number or a multiplier like this, I want to give you an equation that you can actually use to measure these things. All right, so here it is. Take everything that it costs to run the workflow for a month and divide that number, divide the number of outcomes that the business actually accepted. So the top of the fraction has six different parts that we measured. Model tokens, the tool, and the runtime agent calls. the data and the storage, controls, logs, the people that had to review it, the human review component, and the team that has to support it and keep it running. Now, the bottom part of that equation is the accepted outcomes. For request B, that’s an order that actually made it into the ERP correctly and didn’t need to be fixed. Three things that make this honest. One, failed attempts and rework, they stay in the cost. If an agent tried to go through this three times, you paid for three. Second, you count what was accepted, not what was attempted. Three, don’t. Only report the average. If you report the typical case and the worst case, you kind of miss like the large share of the bill. If most runs, if most of the runs that you’re doing cost a dollar and one in 50 costs $40, then the average really hides the thing that you need to see the most. So and that point’s not theoretical. McKenzie looked at its own internal usage and found that 10% of the users consume 65% of the tokens. All right, so let’s go back. Let’s run request B through this equation. A lot of numbers here. Same workflow, 10,000 documents a month, list prices that I used for this, you know, pretty much as of this week. Let’s start with the model. $1,320 a month. That’s 13 cents a document. And that’s the one line that the sponsor that came to us that They actually scaled that one out correctly. Now let’s add the search service at three replicas, storage, a data catalog, the logs, about $1,500 more. Now we’re at $2,800 a month, and so far that sponsor’s number looks pretty reasonable. Maybe a little bit under budget, but still kind of reasonable. Let’s add that reviewer. Say a person looks at 10 to 20% of the documents for 5 minutes each. And that’s pretty reasonable. You know, like in terms of 80% flow through rate, that comes out to 5,400 to almost $11,000 a month, depending on the wage. And so now you can see that that bar just tripled and that’s probably in the low end. So let’s add on top of that, the team that keeps it running. Let’s make that maybe half a person or to one person. That’s another 6 to $12,000 a month. So fully loaded, this workflow costs somewhere between 14 and $26,000 a month. And let’s round that, let’s like, you know, call that maybe 174 to $314,000 a year against the original request, which was only for 25,000. Now look what it’s made out of. Tokens are only 5 to 9% of that bill. People are making up more than 80% of the cost. Now, per document, the sponsor’s priced at 13 cents, and the business is trying is buying something that cost a dollar and forty-five to $2.00 and sixty-one cents per document. Now… I don’t want you to write down these numbers necessarily. The ratio that belongs to this workflow at this volume with this review rate is going to be different than something that you come up when you actually measure a pilot yourself. Now, and then the last thing I would say is that there’s also one other cost that’s not included in this. And that’s the one-time build, the testing and the training that needs to account being here as well. So… Here’s the one number that I’d like you to take out of this section. Someone on your team is going to tell you that the fix is to go to a cheaper model, and they might be partly right. There’s a smaller model in Microsoft Foundry that costs about 95% less per token than the premium ones. On request B, this instance, switching saves 12 and a half cents a document. Now, suppose that small model is slightly less accurate, not a lot, and suppose 2.3 more documents out of every 100 needs a person to look at them for another 5 minutes. Now, that’s it, and that wipes out the entire savings. That efficiency rate made more of a difference. than the model cost. So if a review goes from 10% of documents to 20, that’s $65,000 a year in additional cost. And the whole token bill at this math is only $16,000. So when somebody brings a plan to cut your AI costs and it’s around the model, Ask them what that actually does to the review rates so that you can calculate what the impact is going to be. All right. So what does the committee do with the request number with the request B? It doesn’t kill it. And I want to be pretty clear about that because the number that grew that much sounds like bad news. It isn’t necessarily bad news. This is the only request that we brought to the table with a working pilot. a defined outcome and a sponsor who knows the process. So a higher cost could still be a pretty good investment, but what changes now that you can see what the money is actually paying for. So the red pen comes out, $25,000 is gone, cross that out. The real figure is 174, maybe 300,000. a year and the decision could be to Invest and the approval, though, would carry a condition. The review rate gets measured and it gets reported every single month. And the metric that you’re going to use is the cost per accepted order. So that’s one request. We repriced it. Let’s look at the biggest one that we had on the request. All right, so request A, that’s more than half the money that was on the request. So let’s go through a real similar treatment. Here’s the math that was submitted. 1000 knowledge workers, $30 a month, $30 a month each for Microsoft 365 copilot. and $15 a month each for Mike for Agent 365. Now, if you don’t know what that is, that’s Microsoft’s control plane for agents. Now, 12 months, $540,000. Today, this company is spending 108,000 on 300 copilot seats. So if you do the math on that, this is about a five-fold increase. Now, the CIO, his logic is, you know, was one that I hear pretty consistently, and it isn’t, it’s pretty well thought out. People who have Copilot, they like it, they use it, and the agents that they’re seeing are multiplying, and somebody needs the governance framework to be able to manage them. Now the renewal date, that’s coming up. So why not do everything all at once and just get it off the table? Now that arithmetic is right. The question is whether times everyone is the right way to buy this. So Microsoft AI pricing gets a lot easier once you see that it has three different levers, three different meters, and they don’t convert into each other. So they’re different. The first one is per person who actually needs a license. Copilot and Agent 365 both work. Both work as a per seat, per person license. Now, it’s predictable and finance typically likes that, but it’s the easiest one that you could overbuy because the default answer that most people use is something like, let’s just give it to everybody. Now, the second meter is per use. What does the work actually consume? Agents that are built in Copilot Studio, they burn credits at about a penny a penny for each request. Agents in Microsoft Foundry, those burn model tokens and the credit that does it, it’s not just a message. One question to an agent could actually set off several charge steps. And that’s how the meter scales with activity. So generally, what I’m seeing is that this is the one that gets under budgeted the most. And that under budgeting is actually what happened on request B as well. Now, the third meter is prerequisites. Now, what actually has to be in place before the thing that you bought is going to work? It’s a base suite, a security suite, data capacity, those types of things. Now, request A is entirely on the first meter. It treats the whole decision as it’s just a set of seats. And then it left that third meter completely off the page when they gave us that calculation. So, let’s look at let’s look at the $15.00. Now, a small little surprise that could actually work out in your favor: Agent 365 is licensed per person, not for agent. So, one sponsor could, you know, you can look at it as like it scales it linearly. with the individual. Now, the part that’s not in the request, Agent 365 does have a prerequisite. You need to have Microsoft, you have to have an E5 license or E3 with two additional security suites. Now, this company, fictitious, this fictitious company, they’re on E3. And so the cheapest way for them to qualify for Agent 365 would be to do a step up to E5. And that’s $21 a user. So going back to the original line item, that says $15, but now we have to add that $21. So the invoice increase on per user is actually 36. So multiply that over 1000 people and that step up is $250,000 per year that nobody actually budgeted for. So request A, bought the way that it was submitted, cost $792,000. Now, that’s with just those raw math, those raw licenses, but there’s a slightly cheaper way to to buy the same thing. Microsoft has a bundle that’s called E7 that came out this year, and that includes E5, Copilot, and Agent 365. So for this company that comes to $720,000, a little bit of a savings, not a ton, but it’s a little bit of a savings. But either way, whether you look, however you look at those numbers, it’s not It’s not $540,000. And I would also say that this is list price, so you can always negotiate with Microsoft to get, you know, to get a little bit better price. It’s probably not going to be 720, but that’s on you and on Microsoft to help. I will give you a little bit of a hint. that Microsoft is pushing E7 this year, so you’re likely to get some sort of a discount. Now, I’m not telling you that Agent 365 is a bad purchase. For the people who own agents and sponsor agents, it’s the right control plane. The question is how, the question is which users need to have it. And that’s what I want to talk about a little bit more. So here’s how the committee should be thinking about buying. Start with the 300 seats this company already has. Look at who’s actually using them for recurring week every week in a workflow that you can name. In this company, that was 220 people. The other 80 seats, we’re going to get, we’re going to reassign them to people that are actually going to use them. And then ask yourself, where do the next seats go to? What 2 roles have a measured workflow where Copilot might actually offer some sort of real savings? Let’s hypothetically say that’s another 130 people. So we get 350 seats, not a whole 1000. Everyone still has copilot chat. Everybody in your organization still has that. It’s included with the subscription, the E3 subscription that they already have. But Agent 365, that’s going to go to the people who need it. The 60 people who own and sponsor agents. instead of just like everybody. And that’s the pattern that we used a little bit earlier. 10% of the users drove 65% of the consumption, if you remember that. Usage is never really ever spread evenly across the company, across every user in the organization. And that’s why headcount’s the wrong way to multiply this. So word of caution, you’re going to see studies that quote usage rates at 23%, 65%, 80%. Don’t adopt any of those as your cutoff. They measure different things. They measure things like daily use, daily use, monthly use, maybe some survey answers. Use the data that’s in your tenant. and slice that up by role so that you can understand like exactly who is using it. So, and that’s kind of what the rule is. Renew it by role and by demonstrating recurring work. And then when you see the next group of people that are going to adopt it, then you’ll have a little bit more of an informed decision. All right, so… Every prepaid offer, AI offer, pretty much is the same trade. You get a discount and in return you promise to use it. So the discount that you get is really kind of a bet on your own utilization, what you’re actually planning and doing. Let’s take a scenario like Copilot Studio credits. There’s a pay as you go and a credit cost a penny. Or you could go and you could buy a pack. That’s $200 a month for 25,000 credits. So that’s a pretty good deal if you use more than 20,000 of them. But if you use fewer than $20,000, the pay as you go is actually cheaper. And the thing about it is that these credits don’t roll over. They reset every month. Now they have annual prepaid pools and they work pretty much the same way, but they have higher stakes. Now the smallest tier is typically about 5% off. So you have to use more than 95% of the capacity that you bought just to break even. And remember, What you don’t use, it expires. There’s no refund, there’s no exchange, and most of these things renew automatically unless you turn it off. And I guess that’s kind of something that surprises a lot of people. In Azure, when you have a budget, you know, you get an alert and it tells you that you’ve overspent something, but what it doesn’t do is it doesn’t actually stop the spending. And none of this, none of this that I’m walking you through to saying like, don’t actually build a commit, don’t do that. It means commit to the baseline that you’re pretty sure of and leave the uncertain part, you know, flexible. All right, so request C has a $150,000 capacity commitment. Um, in it and like, I’m gonna I’m gonna use that #2 as we kind of move along. But while we’re on renewal, I want to use a calendar that’s not on somebody’s request. So remember, first, your base Microsoft 365 price went up in July 1st. So for a lot of people, the E3 went from $36 to $39 if you weren’t on like a longer term renewal or commitment. The E5 actually went up from 57 to 60. And so at your renewal time, you’re going to actually see those dollars hit. Now for this company, this fictitional company, it’s $36,000 a year that nobody asked for. Microsoft just increased the invoice. And you can’t really decline it because that’s where you’re at. So remember that aspect of it. And then also remember that things retire. So a preview feature, like preview features retire, something like Foundry workflows, that’s going away in December. Classic Foundry agents are going to retire at the end of March. Prompt flow is going to retire in April, and there’s models that retire, you know, they get switched off, you know, on a published kind of schedule. Now, why do I bring that up? Because each of those things, that’s a migration. Somebody has to move the workload, not only just move the workload, but they have to test it, and they have to make sure that it’s still giving the right answers. Now, the token might not, the token cost or the credit cost might not actually change, but you have to plan for the cost of doing that work. So that’s still real money. Now, I haven’t seen anything announced about like AI products, you know, for 2027. No price increases across the board, no price decreases across the board. But that doesn’t mean that there isn’t going to be one. So just make sure that you’re kind of keeping that flexibility and And then on this calendar, like it’s a calendar that you don’t actually own. It’s one that’s being forced on you. A really, really good example of that would be Microsoft’s cowork pricing that changed with very little notice and very little things that you could do about it other than just turn it off if you weren’t comfortable with it. All right. So let’s go back to the board and take a look at the update on the requests. So request A came in at 540,000. Its true cost, though, was north of 700. So the committee approves, but the committee resized it and approved 152,000. Now that’s not a cut. right? It’s a cut from the ask. It’s $44,000 more than the company is actually spending today. So it’s an expansion, but it’s an expansion that’s been sized to the people who’ve shown that they’re going to actually use it. And the condition that they set on this is that the seats follow rules with measured reoccurring work and that the next cohort gets released when the evidence, you know, actually supports it. All right, so that’s two of them. Let’s do the one with the great demo. Request number C, as submitted, the autonomous quoting agent. And I want to give this one, you know, a little bit of credit because it’s a good idea. It’s actually one that we do here pretty often. Today, a quote at this company takes two or three days. Somebody has to go and look up pricing, they go and check the inventory, they read the customer’s history, and then they write it all up. Now the demo did all that in about four minutes and on top of that it sent the quote out on its own. So if you’re a VP of sales and you watched, you know, you probably just watched your win rate go up in real time, two days, 4 minutes. Now the request is $450,000. There’s that’s broken into $300,000 to build it and $150,000 a year of committed capacity, so that it can, you know, still deliver it that four minute that that 4 minutes. Of the three requests, this is the one with the most upside. It actually touches revenue, not just a cost factor. And everybody who saw the demo was pretty much blown away by it and said, let’s go do this. Let’s go ahead and score this one properly too. Now. I have come up with a better way to rank use cases and I’ve done it this way so that the budget that you allocate for it actually follows value instead of just like the enthusiastic demo that you saw. And I call this a use case value map. It’s got 10 different dimensions and you score each one of these from 1 to 5. Now, the thing I want to point out is that every one of these dimensions, every score has a written definition so that two people that are scoring the same idea land on the same value. Now, then I had the dimensions weighted. Five of them, five of these actually count as triple the other ones. And those five would be business value, strategic alignment, technical feasibility, data readiness, and owner readiness. Now, let’s start where this room started, strategic alignment. Request C gets only… One, five, request C, if you look at this, is the only one that gets 5 because it goes straight at the revenue and that’s what everybody felt when the demo was there. They weren’t wrong about that one. But look at the look at 2 rows that the demo didn’t show you. Specifically wants you to look at data readiness. This project, this C, it got a 2 on it. The sources are known, but access and permissions aren’t sorted out. And now the other one, owner readiness, that gets another two. There’s a sponsor, and nobody though that actually is going to be capable of running it. So we have We have a little bit of lack of readiness on this one. Let’s total these ones up. Request B, the one that was kind of quiet, that scores 101 out of 120 total possible points. Request A, if you take a look at it, scored 93, and request C, the flashy one, only scored 69 on this one. So the demo puts C first, but the map puts it at dead last. Now, it didn’t make that a worse idea. It just had a scoring across a common playground. So, what do I do with that? Um… Score is just a score and it’s got some weaknesses. A strong row can actually pay for a weak one. Picture a use case that scores a perfect 5 on 9 dimensions and a one on the data readiness. It still totals 108 out of 120. That’s 90% of the actual points and it goes to the top of almost every list. But data readiness is really important. And I’m going to tell you that if it only scores A1, it can’t go live. So three of the 10 dimensions should be working as gates, gates to your decision process. And a gate is going to hold the money until you can actually get that fixed. The first one that we already talked about is data readiness. At this company, the pricing lives in three different systems, and each one has different rules about who can see what. Now, the demo that we saw used just a single extract, and it was clean. Now, production can’t do that same thing. The second gate that I have for organizations is owner readiness. The VP of sales sponsors the agent, but when it sends a quote at the wrong price, who’s going to own that problem? Today, nobody owns that problem. Gate 3 is the risk tier. With request B, A mistake is an order that a reviewer actually catches. But in request C, a mistake is a price that gets sent directly to a customer. That’s the whole point of that automation. That becomes a commitment, and it means if this kind of commitment needs a full review before we can actually scale this one out. West C actually gets held at all three of these gates. And one of these would have been enough to hold any one of these things. Now, you’re on the committee. What does that mean? Does it mean that you say no? Not necessarily. So, there’s a rule, there’s a delay rule. Delay the commitment, but don’t delay like the evidence that to follow through on this. What do I mean by that for request C, that splits this money, the request into two. In 2027, the committee funds $120,000. That pays for a finite piece of work. Fix the pricing data and the permission that this use case is going to need. Name the person who actually is going to own the outcomes and write the tests that say what an acceptable quote is going to look like. The rest of it, you’re pushing that out to 2028. That’s the scale build. and it’s the $150,000 capacity commitment that I asked you to hold on to if you remember earlier. Don’t prepay for capacity in a workload that can’t actually go live. And it comes with a specific reopening condition. All three of those gates that we had in the previous slide, they have to pass. Now someone You’re probably going to ask the obvious question. If models keep getting cheaper, why not wait on everything? Because 2 things are true at once. The price of a fixed level of AI capability has been falling by almost half every single quarter. And Gartner forecasts that the cost of an agentic workflow is going to rise by more than fivefold in 2028. because we keep asking agents to do harder and harder things. So falling prices are a reason to keep your commitments short. They aren’t a reason to wait on work that’s already been proven to work. Every month that request B isn’t running, that’s a month of benefit that you don’t get back. So lesson here is delay what’s unproven, but. Don’t delay what’s working and don’t delay anything that has a retirement date on it. All right, at the start I told you that there was a request that was that was missing from the packet. I’m not going to do the poll again, but look at what each one of these requests quietly assumed. Request B needed a team to keep it running, and when we rebuilt its cost, we charged B for that whole team. Now, that’s probably unfair to B, because the second and the third workflows, they’re going to need similar people, probably the same people. Request C needs data and permissions work for it can pass the gate. The renewal calendar needs someone to migrate workloads and retest them when things actually retire or get moved out of the system. And request A needs somebody to drive adoption and measure whether the seats are actually paying off. Nobody submitted any requests for any of that. Each sponsor assumed it was somebody else’s problem and somebody else’s budget. And I know some of you are, you know, some of you that are on this webinar today, we’re talking about governance committees and we’re talking about staffing. And that’s, you know, that’s what I’m talking about here. And this is what the committee should do. It writes a request for D itself. That’s the shared foundation. The team runs, the run team that we’re talking about moves out of request B and into D, where it’s funded a single time for all the use cases. Let’s call that somewhere around 75 to $150,000. And what that does is it brings request B’s ask down to 99 to 100 and sixty-five $1000 on its own. Now, the rest of the, I don’t know, I can’t give you a complete number. There’s no published range from migration or data work that I would, you know, want to put my name on. It has to get scoped out properly, and that’s on you and your organization to figure that out. It’s two ways that we can get this wrong. Charge the first workflow for the whole platform, or leave the platform out and every workflow looks like it’s cheaper than it actually is. So request D needs a little bit of a shape, though, or it’s going to wind up turning into more of a slush fund. And here’s something that I’ve been calling the AI operating system, and it’s got six different layers. You’ve already used the first three of them today in this webinar, leadership and decision rights. That’s this committee. Demand and portfolio, that’s the value map on the board. And then trust and risk, that’s the gates that we established. Request D is the other three. The Microsoft control plane, which is where the data and the permission work lives. Delivery and release, which is the migrations and the retesting. and the operations and value, that’s the run team adoption and the measurement. And here’s why that matters for a budget. Each layer is funded once and it’s shared by every use case that goes on top of it. So each layer gets one owner and one line. Now next year, when request #5 shows up, you don’t have to rebuild anything. You just ask which layers it needs to lean on, and that’s how the that’s how the 10th use case ends up cheaper than the first ones that you went through in your in the system. I do see that there’s some things in the chat. Anybody ask a question? No, that was just the looks like it was just the poll. Um… All right, board update. Now, this time we’ve got the board kind of filled up, right? Request C came in at $450,000. The committee funds 120 for the milestone and then the rest of it, we’re going to push out to 2028. And we’ve got a clear condition for when that actually comes back. Request D, That takes the empty role. Nobody asked for it, but the committee said we’re going to fund that anyways. Request B, that got a little bit cheaper because it’s no longer carrying the whole run team that we had allocated in the first pass. And then the most exciting request waits, at least the glamorous one gets funded. And that’s what the That’s what a healthy committee output looks like. That’s what it looks like when you guys are doing your job. And then there’s one promise that we haven’t touched yet, and that is, how do you prove that any of this actually paid off? So… You’ve all seen the headlines, AI returns $3.70 for every dollar that you spend. I think I copied that from Microsoft’s front page, but I’ve seen that number, heard that number quite a bit. Now, can that go into your business case? Probably not. I think it’d be dishonest just to take a number like that. The most recent report behind that page puts it at 2.8 for generative AI and 2.3 for agents. Now, those are estimates at very large companies and estimates that they gave about themselves. There’s no audits behind them. It’s just a description that somebody at the company made. Now, in fairness, the scary headline has the same problem. You’ve heard that 95% of AI pilots return nothing. That came from a 2025 study that mixed different samples and different definitions of what the return means. It isn’t the odds that that’s going to be what your next pilot rates at. So throw those numbers out, and here’s a pair of numbers that I would keep instead. McKinsey surveyed about 1,700 people this summer. 80% that AI improved their own productivity. 37% said it had a positive effect on the company’s earnings. 80 and 37. People are saving time, and most companies can’t find that, though, in the financials. That gap is the whole problem and let’s find out what is actually happening with that time. So take 100 of those copilot seats from request A. Suppose that each person saves 30 minutes a week. That’s a pretty believable number, something that I think is pretty well established that you could get out of it. Over 50 weeks at $60 an hour, that’s $150,000 of saved time. This is the figure that shows up in the business case that most people throw out. that most people put in front of you, I should say. Now ask what the people did with those 30 minutes. Some of it went into useful work and some of it went into lunch hours. Suppose that 30% was redeployed to something that actually matters. That’s $45,000 of real capacity, not the original number that we had. Now I ask how much of this is cash. Did a budget line go up or down? Did you cancel a contractor? Did you skip out on a hire? Did you ship more products with the same number of people? If the answer is no, then that cash number is actually zero. $150,000 was promised. 45,000 was actually redeployed. but you still get to zero. Same people, same 30 minutes, you know. And I’d say that this isn’t my opinion. There was a randomized study that covered 66 companies. and a little bit more than 7000 workers. The people were given, the people given the tool spent about two hours less on e-mail every week, and the researchers, they couldn’t detect any change in how much work was actually getting done. Now, the time is real, but it turns into money only when somebody changes how they operate. It changes only if you don’t make a hire or the overtime that you stop paying or an order that you can take now that you couldn’t do before. So the question for every request is which operating change is this going to produce and who’s on the actual hook for it? So, here’s how I’d ask Finance to keep the books on this. Create 4 registers. The first is cash. A budget line actually went down. Now, request B belongs here if the higher it avoids really don’t get made, right? The second is capacity. That’s the time freed and you can show where only when you can show where it went. That’s request A. It’s a real benefit, but it still isn’t cash. And the third is margin. You sold more or you shipped more because of it. And that was the promise of request C when it was ready. Then there was the 4th one, that’s risk. You reduced an exposure or you met an obligation. Now, that’s, nobody should pretend that the team, that that run team has a payback period. It has a rationale, and that’s a totally different thing. There’s the cost of failure that I’ve talked about in other talks, and I don’t have a ton of time to go through that one, but that’s what you’re trying to reduce here is cost that would happen if you get a mistake wrong. So one rule holds all this together. Never count the same hour twice. If you book the time as capacity, then you can’t also book that revenue that the people earned it with. So when a business case starts to mix all four of those into one ROI percentage, that’s your cue to go ahead and Kick that one back out. All right, so how do you prove it? How do you decide what proof looks like before you spend the money? And here’s three things that you can write down. One is the accepted outcome. I go back to that. A single sentence with a quality limit and a time limit. For request B, that was an order in the ERP. an order that was correct within an hour and it had no rework. If the sponsor can’t write down that sentence, then the request isn’t ready for you to measure it. Two, we’ve got the baseline. What does today look like? Volume, minutes per order. What was the error rate? What’s the cycle time? Measure that a few weeks before you launch, and then you’d be surprised how often nobody actually knows what that number is. And 3rd, the comparison. You need something to compare against, so you can tell AI from a slow month. The simplest version is just a staggered rollout. Two order desks go first. to wait. Now you’ve got a control group and you’ve got, you know, you’ve got a control group that you can measure against the other aspect of it. And set expectations on timing. Gartner asked finance leaders how long this kind of work takes to reach its expected value. For document extraction and payables automation, The answer was actually 9 to 10 months. So plan to prove it across the year with checkpoints along the way. Don’t expect to get those numbers in the first month. The last piece of the proof section, and it’s the one that gives the rest some real teeth. Prove the full amount if you like, but release it in four different steps. Step one is feasibility. It can do the task. That’s what the pilot proves. And it’s all that the pilot proves. So step 2 is production performance. It does the task at an acceptable cost per acceptable outcome. And that’s where request B is headed to next. Step 3 is operating change. Something in the business starts to change. It’s different. That hire wasn’t made. The overtime was reduced. And then step 4 is replication. It works a second time in a second place. Now, each step has a date and each date has a decision attached. Continue, re-scope, or stop it. And here’s what I’m seeing mostly in practice. AI programs get funded at step one, and nobody ever asks about step three. The money just kind of keeps flowing because the demo was good and nobody wrote down what it would take to make them actually rethink that decision or stop. So evidence, I would say, is use that evidence to earn the right to scale things out. Now, let’s go back to the vote. What did the board decide? This is what you as the committee, you know, finally, finally decided. Request B to order intake workflow. It came in, it came in and a request $25,000. You’re investing 99 to 99 to 100 and sixty-four 1000. It’s the request that grew the most. But it also was the one that had the strongest case. On request A, the copilot and Agent 365, it came in at $540,000, and you resized it appropriately to 100 and 52, which is still more than what you’re spending today. So like I said before, it’s growth. And then request C was the quoting agent. It came in at 450,000 and you’re funding 120 this year to get it through the gates. And then the rest, you’re deferring that until 2028. Don’t forget about request D. That’s the Shared Foundation. Nobody submitted it, but you guys were smart and you funded the run team at 75 to 150,000. And you still got a little bit of scoping to do the rest. So invest, resize, delay, fund. Four requests, four different answers. And if your budget process can only say yes or no, then you’re missing the steps that I put out for you here. Take a look at the bottom line, though. Look at the totals. A little over $1,000,000 was submitted. Those same requests, bought the way they were written, would really have cost somewhere between 1.3 and 1.6 million. I walked you through the math, how to score those appropriately. You proved somewhere between 450 and 600,000, plus some of the work that still has to get scope. That’s roughly half of what was asked for. and about a third of what it would have cost if you would have just rubber stamped it. And more of this is actually going to reach production, because the one workflow that’s ready is fully funded, and so is the team that’s needed to run it. And the total though, it’s not really the point. Look at what the money is attached to now. Every approval on that board carries 4 things. An owner, a named person who’s accountable for the outcome, a person with a name, not a title. A measure, that’s the accepted outcome, and which of the four registers that the benefit gets booked into. Full cost basis, people included, all the prerequisites that we came up with included as well. And finally, a review date with a decision that should be attached to it. Are we going to continue with it? Are we going to rescope it? Or are we going to stop? Now, notice the things that aren’t on this list. We don’t have vendors, I don’t have models, tools. Those things can change during the year. And honestly, the way that we’re seeing it, they probably are. But these four things that are up on the screen, they shouldn’t change throughout the year. And if a request comes in and it doesn’t have those four, it’s not ready for you as a committee to go ahead and vote for. Send it back. Tell the sponsor exactly what’s missing. You’re going to be doing them a favor. Now, every deferral carries one thing, the evidence that reopens it. Request C wasn’t dead. It has three gates and dates. The VP of sales knows exactly what to bring back, which is far more useful than him hearing, we’re not going to do it this year. So I think this is pretty much the whole method. Now, if we go back and think about the same question I asked before, which ones of these would you fund first? And now remember, this time there’s a fourth option. And that’s pretty much this whole thing in one picture. So I started this off with six words. The pilot was cheap, production isn’t. I hope that sentence sounds a little bit different than when we started at the beginning of this. Production’s not cheap, and that wasn’t really ever the problem. Nobody at this company got into trouble because AI was costing money. They’d have been in trouble if they’d approved $1,000,000 without knowing what any of that was actually buying. So here’s the difference between spending and investing as plainly as I can put it. Spending is aligned with a price. It gets approved because it was there last year, or because the demo was good, or because everyone else is doing it. and those kind of spending, they get renewed by default. Now, investing different. It’s aligned with an accepted outcome, a full cost, an owner, and a review date. It gets released when the evidence actually shows up, and it gets stopped when it doesn’t. The same dollars can be either one. Request A was spending at $540,000, and it’s an investment at 152. Nothing about the software changed. Nothing about the tools changed. What changed is what you asked of it. You don’t need a bigger budget for AI in 2027, and you don’t necessarily need a smaller one. What you do need is 1 where you can actually say what the line items go for. All right, so three things that you can do before your next budget meeting. One, pick a single AI request and go back and reprice it. Ask the sponsor for the cost per accepted outcome with the people that are in it and what that review rate is. Nobody knows, you’ve learned something important for free. Second, read your Microsoft renewal with three questions in mind. Who actually needs a seat? What’s the prerequisite for anything new on that quote? And what’s retiring in 2027 that someone has to migrate? The third one is write the request that nobody submitted. What’s the run team look like? What’s the data work, the migrations, the adoptions, and put names next to those. If you don’t do these three things, you’ll walk into the meeting ahead of nearly everybody else. And if you’d like any help, that’s what concurrency can do. So here’s a QR code for some feedback. If you’d like. Would love to hear your feedback on this. But also, there’d be an opportunity for you to reach out to us and book a meeting where I can help you score some of these. And with that, I want to say thank you very much for your time today.