One dashboard across EA, CSP, Azure, and Copilot consumption, chargeback that changes behaviour, and the forecast. Three knowledge checks along the way, and 1 clip from a senior cloud advisor.
The presenter in this session is an AI generated avatar. The curriculum and guidance are real, produced by Redress Compliance analysts from our consulting engagements and market network.
This is a taught session, not a talking head. The instructor works through analyst grade slides, and three times the video stops on a question with four options on screen. Pause, commit to an answer, and the next slide explains which option is right and why each of the others is wrong. Once in the session the frame splits and a senior cloud advisor gives the view from inside real Oracle negotiations, and the instructor picks the clip apart when the slides return.
The full narration of this session, section by section, for reading and reference. Guest analyst clips are marked.
Welcome back, session thirty four of forty. The last three sessions built visibility: compliance reconciliations, the three way join, and the gap ledger. Today we put a financial frame around all of it, which is what FinOps means in this context. And I want to open with a question that sounds trivial and is not. What does Microsoft cost your organisation per person? In most enterprises nobody can answer that. Not because the data is hidden, but because there are five separate spend lines with four different owners, four different billing shapes, and four different reporting rhythms, and no single view that contains all of them. So the organisation manages five lines separately and optimises none of them against the others. Which means the largest problem in any given quarter is invisible unless it happens to be in the one line somebody was looking at.
Five takeaways. One, one view and five lines: EA seats, CSP subscriptions, Azure consumption, Copilot subscriptions, and the metered AI layer, and most estates report on some of those and none of them together. Two, your real unit price: fully loaded cost per active user ran one point eight to three point four times the thirty dollar headline once base licences and idle seats were counted. Three, the idle rate: thirty to forty five percent of Copilot seats had no sustained weekly usage at ninety days, renewed by inertia wherever nobody measured. Four, the overage cliff: allowances ran out one to two quarters earlier than forecast once agents were switched on, and buyers with no alert found out at invoice time. Five, what terms cap it: a ceiling, rollover, and a hard alert threshold cut effective overage cost by twenty to thirty five percent, and they belong in the order form.
Why one dashboard, three reasons the pieces never meet, and this is an organisational problem that produces a financial one. Different owners: seats sit with IT or procurement, Azure sits with infrastructure or cloud engineering, and the AI consumption line sits with whoever built the agent, which is four owners, four reporting rhythms, and no shared denominator anywhere. Different billing shapes: an annual commitment, a monthly subscription, a consumption meter, and a credit pool do not add up naturally, and anybody who has tried to reconcile them into one figure has found the effort is in the normalising rather than in the arithmetic. And different time bases: a true up lands annually, a CSP term renews on its own date, Azure bills monthly, and credits deplete continuously, so a quarterly view of one line tells you almost nothing about the total. The consequence is that nobody can answer the per person question, and the dashboard is what makes it answerable.
The five spend lines, and read the third column as a checklist of everything this course has already covered. EA seats: annual commitment with true up at anniversary, failing as shelfware per session thirty three, growing quietly and correcting once. CSP subscriptions: monthly or annual terms per seat, failing as committed seats you cannot release, which is next session. Azure consumption: metered against the MACC commitment, failing as shortfall or overrun per session twenty one. Copilot subscriptions: per user per month on top of a base, failing as idle assignment at thirty to forty five percent per session twenty. And the metered AI layer: credits and messages with no headcount ceiling, failing as the overage cliff later in this session. The FinOps contribution here is not new analysis. It is putting all five failure modes in one place so that the largest one this quarter is the one you see.
First check. Finance asks what Microsoft costs per employee. What do you need to answer it properly? A, the EA invoice divided by headcount. B, all five spend lines normalised to the same period and divided by active users rather than headcount, because four of the five do not bill per employee at all. C, the total Microsoft spend divided by headcount. D, seat costs only, since consumption is an infrastructure matter. Pause it, and as you think, count how many of those five lines have any real relationship to headcount.
The answer is B, and the count is one. Azure bills for consumption, the AI layer bills for work performed, and CSP bills for committed seats which may or may not match your current people. Dividing any of those by headcount produces a number that moves when your headcount moves and tells you nothing about whether the spend underneath is efficient. A covers one line and calls it the answer, which is how most organisations currently report this to their boards. C at least captures the total and still divides by the wrong denominator, and it misleads in a specific direction worth naming: it improves when you hire, which is precisely backwards. D is the split that causes the whole problem, because separating seats from consumption is exactly how an agent costing more than a department goes unnoticed, per session eighteen. Normalise the period, sum the lines, divide by actives.
The unit that matters, five things about cost per active user. It is not the list price: fully loaded cost per active user ran one point eight to three point four times the thirty dollar Copilot headline once base licences and idle seats were counted into it. The denominator is sustained use, meaning sustained weekly actives rather than assigned seats, because any other denominator flatters the figure and will not survive its first challenge. The numerator is the full stack: the subscription, the base licence it requires, the metered layer beside it, and the deployment and governance work, all of it or the number is not a cost. It is the only comparable figure, working across vehicles, across years, and across products, which nothing else on the invoice does. And it is what you negotiate from, because estates that arrived at renewal with documented weekly active usage negotiated from their own telemetry while estates without it negotiated against the vendor's adoption narrative.
Guest analyst The most useful twenty minutes I have ever spent in a CFO's office involved one division sum. A pharmaceutical company, and the CFO had asked a perfectly reasonable question, which was whether Copilot was good value at thirty dollars a head. His team had prepared a paper arguing yes, and it was a decent paper. And I asked whether I could add one line to it. So we took the full Microsoft 365 Copilot spend for the quarter, which was the subscription line. Then we added the base plan upgrades that had been bought specifically to make those people eligible, because about nine hundred of them had been on a frontline plan. Then we added the metered consumption line, which lived in a different system and which nobody in that room had seen before. That gave us the numerator. Then, instead of dividing by assigned seats, we divided by the number of people who had used it in at least three of the previous four weeks. And the answer came out at just over ninety dollars per active user per month. Now the CFO's reaction is the reason I tell this story. He did not say cancel it. He said, at ninety dollars I need to know which roles are getting ninety dollars of value, and at thirty I did not have to ask. That is exactly the right response, and it is the response you can only get to if somebody has done the division. The programme continued, at about a third of the seat count, aimed at the roles that could clear the higher bar.
Ninety dollars per active user, and a better question from the CFO because of it. Do the division. Second check.
Check two. Your Copilot line looks like thirty dollars per user per month. What is the realistic figure per active user? A, thirty dollars, that is the contracted rate. B, somewhere between one point eight and three point four times that, once the required base licence, the idle assignments, and the metered layer are counted against sustained actives. C, slightly above thirty, allowing for administration. D, below thirty, because volume discounts apply. Pause it, and ask yourself what else had to be bought for that thirty dollar line to work at all.
The answer is B, and three things drive that multiple, all of which have appeared earlier in this course. The base licence is a prerequisite, from session sixteen, and on a frontline population it is the larger of the two lines, which is exactly what the pharmaceutical company found. Idle assignment sits at thirty to forty five percent with no sustained weekly usage at ninety days, which means the same spend gets divided across far fewer real users. And the metered layer bills separately, from session seventeen, usually in a system the seat owners never look at. Across the deployments benchmarked, fully loaded cost per active user landed at one point eight to three point four times the headline. A reports the rate and calls it the cost. C accounts for the smallest of the three factors and misses the two that matter. D is true of the rate and irrelevant to the multiple.
The overage cliff, three things about the moment an allowance ends, and the framing is that before the cliff AI feels free while after it every action is a line item. It arrives early: included allowances are sized for interactive use, and agentic runs consume five to ten times what an interactive prompt does, so allowances ran out one to two quarters earlier than the vendor forecast once agents were switched on. It is invisible until the invoice: buyers with no alert threshold discovered the cliff at invoice time, when it was already too late to renegotiate the rate, and nothing in the daily experience signals that an allowance is depleting. And three terms cap it: a ceiling, rollover, and a hard alert threshold, where buyers who negotiated rollover and a spend ceiling cut effective overage cost by twenty to thirty five percent. All three belong in the order form rather than in a renewal conversation, because afterwards you are asking rather than negotiating.
Chargeback that changes behaviour, and what has to be true for it to work at all. Attributable: spend traced to a team rather than to a cost centre average, because an average nobody caused changes nobody's behaviour. Timely: monthly rather than at the annual true up, because a cost eleven months old is history rather than a decision anybody can still influence. Actionable: the receiving team can genuinely change it, since charging for something somebody cannot influence breeds resentment and nothing else. Understandable: credits and messages translated into money, because nobody manages a unit they cannot price. And paired with the ledger, alongside the assigned against active gap from session thirty three, so that cost and waste appear in the same view. Azure meters pay as you go credit spend, which makes attribution available if you set it up at the start, and attribution changes behaviour where a central budget line never does.
Last check. Your AI consumption is rising steadily. Where does the control belong? A, in a policy asking teams to be mindful of consumption. B, in the order form as a ceiling, rollover, and alert threshold, plus monthly attribution, because nobody experiences running an agent as spending money. C, in an approval process for every agent. D, in the annual budget, reviewed at the true up. Pause it, and as you think, ask what a person or a scheduled job actually notices at the moment the consumption happens, because the answer is nothing at all.
The answer is B. A asks people to be careful about something they cannot see and are not billed for, which is why consumption governance built on awareness fails in every organisation that tries it, and it is not a failure of the people. C is closer and overcorrects into a bottleneck, because an approval gate on every agent slows exactly the work the platform exists to enable, and the arithmetic from session eighteen belongs on large or scheduled agents rather than on all of them. D discovers the problem eleven months after it began, which is the invoice time failure I described a moment ago. B works because each element attaches to the mechanism rather than to anybody's intentions: the ceiling bounds the worst case, rollover stops unused allowance evaporating, the alert fires before the drop rather than after it, and monthly attribution puts the number in front of the team that generated it while they can still do something about it.
The FinOps method, three things to build in this order, each about a week of work and each making the next one possible. One, the five line view: EA, CSP, Azure, Copilot, and the metered layer, normalised to a monthly figure in one place, and it is a table rather than a tool, because the value is that it exists and that one person maintains it. Two, cost per active user: full stack spend divided by sustained weekly actives, per product, which is your real price and the number to take into every renewal, every business case, and every conversation with finance. Three, the caps and the alerts: ceiling, rollover, and alert threshold on every metered line, negotiated into the order form, then monthly attribution to the teams generating the spend. And the compounding matters: this feeds the gap ledger, the Copilot case, and the counter quote, so the same measurements serve three jobs.
Session thirty four, three sentences. One: five spend lines with four owners and four billing shapes means nobody can answer what Microsoft costs per person, so the dashboard is the precondition for the question rather than a reporting nicety. Two: cost per active user is your real price, and it ran one point eight to three point four times the Copilot headline once base licences, the thirty to forty five percent idle assignment rate, and the metered layer were counted properly. Three: the overage cliff arrives one to two quarters early once agents are running, and the controls are structural rather than cultural, being a ceiling, rollover, and an alert threshold in the order form, worth twenty to thirty five percent of effective overage cost. Next session closes module seven with the problem of running more than one purchasing vehicle at the same time.
Homework, about an hour, and this week you build the five line view. One, find all five numbers: EA, CSP, Azure, Copilot, and metered AI, for the same month, and note which ones took more than a phone call to obtain because that difficulty is part of the finding. Two, name the owner of each: one person per line, and if two of those lines share no reporting relationship whatsoever then you have located the structural problem. Three, compute one unit cost: Copilot full stack spend divided by sustained weekly actives, compared against thirty dollars, and write down the multiple. Four, check for a ceiling: does any metered line have a cap, rollover, or alert today, and most estates find the answer is no on all three. Five, set one alert: a threshold with a name attached, on the largest metered line, which is the cheapest control in this course and still usually missing.
Five reads before next session, all free on redress compliance dot com. First, the Copilot true cost analysis, which carries the realised multiple, the idle rate, and the measurement method behind this session's central number. Second, capping the AI consumption overage cliff, on the cliff itself, the steepest shapes, and the three terms that cap it. Third, the Azure FinOps cost governance framework, for the consumption side with attribution and forecasting. Fourth, Copilot Credits and the Azure commitment, on how two of your five lines interact on a single meter, which is easy to miss. And fifth, Microsoft SAM and licence optimisation, for the seat side of the same dashboard from session thirty two. Next session closes module seven: governing multiple vehicles. EA plus CSP plus MCA in one estate, the contract register, the calendar, and the co termination strategy. See you there.