←
AI-Ready Foundation
Strategic · M3 · lesson 3 of 5 · queued
Preview — browse every lesson free. Enroll to mark lessons complete, open partner links and save your progress. Login & enroll →
Measuring AI ROI Honestly: Productivity Evidence, Attribution, and Portfolio Management
📖
now learning

Measuring AI ROI Honestly: Productivity Evidence, Attribution, and Portfolio Management

15 min

AI ROI measurement is where most scaled AI programs quietly lose the CFO. The slide says the coding assistant saves six hours per engineer per week, worth a million a year. Headcount is flat, delivery dates have not moved, and the token bill has doubled. This lesson shows you why that happens, and gives you a workshop, the Value Ledger Review, that replaces the slide with a number finance will sign.

The AI ROI Measurement Problem: Where Is the Million?

"Where is the million?"

Picture the finance business partner asking it quietly, three slides into a quarterly AI steering committee. The head of AI has just shown the headline: the coding assistant "saves 6 hours per engineer per week, worth $1M a year". The number comes from a survey of the engineers. They were asked how much time the tool saves them, they answered honestly, and somebody multiplied the answer by headcount and salary.

She is not attacking anyone. She wants to know which budget line went down, or which revenue line went up. Nobody in the room can point to it. The meeting moves on, and something has broken that will not show up in the minutes: finance has stopped believing the AI numbers. Every later business case now gets a silent discount.

That is the expensive problem. It is not that AI has no value. It is that you cannot tell which of your initiatives have value, so you keep funding all of them a little, and none of them enough.

What happened when researchers timed developers instead of asking them?

They found that the developers felt faster and were measurably slower. This is the best single story for understanding why survey numbers fail, so it is worth telling slowly.

METR, a research group that evaluates AI systems, ran a randomized controlled trial (an experiment where chance decides who gets the treatment, so the two groups are comparable). The results were published in a paper on arXiv in July 2025. Sixteen experienced open-source developers worked on 246 real tasks in mature repositories they had known for about five years on average. Each task was randomly assigned to be done with AI tools allowed or not allowed. When they were, the developers mainly used Cursor Pro with Claude 3.5 and 3.7 Sonnet. The work ran from February to June 2025.

Before starting, the developers forecast that AI would make them 24% faster. After finishing, they believed it had made them 20% faster. The stopwatch said something else: with AI allowed, tasks took 19% longer, with a confidence interval of 2% to 39% longer. Outside experts did no better at guessing. Economists had predicted a 39% speedup and machine learning experts 38%.

These were skilled people, reporting in good faith on their own work, and they were wrong about the direction of the effect. If you had surveyed them, you would have built a slide claiming a 20% gain.

Does that mean AI slows developers down?

No, and the second half of the story is what makes it honest. The study covered a small sample and one population: experts working in large codebases they knew deeply. It says nothing reliable about juniors, new projects or unfamiliar code.

METR then tried to repeat the measurement. In an update published on 24 February 2026, it described a follow-up begun in August 2025 with 57 developers, 143 repositories and more than 800 tasks. Only 10 of the 57 were returning developers from the first study; the other 47 were new recruits from a wider set of repositories. This time the estimates pointed toward speedups: 18% less time for the returning developers (interval from 38% less to 9% more) and 4% less for the new recruits (15% less to 9% more). Both intervals include zero, and METR itself calls the new results unreliable. Many developers refused to work without AI. Between 30% and 50% withheld tasks they did not want to do without it. Agents changed which tasks got attempted at all, and running several agents at once broke the time tracking. METR's own conclusion is modest: developers are probably more sped up in early 2026 than in early 2025, with only very weak evidence of how much.

Two lessons come out of this, and you need both. First, do not trust how fast people feel. Second, measurement is hard enough that a specialist research group says its own careful design can no longer do it cleanly. Your internal survey certainly cannot. What you can do is measure the outcome of the workflow, which is the method this lesson teaches.

What does the gap cost across the market?

The pattern on that steering committee slide repeats at scale. All four sources below are self-reported surveys, so read them as what executives say, not as audited results.

  • McKinsey's State of AI survey, published 25 August 2026 with 1,719 respondents, found that 80% of individuals say AI improved their own productivity, while 37% of organizations attribute at least some EBIT impact to AI. That is essentially unchanged from 39% in the 2025 survey. High performers remain at about 6%.
  • McKinsey's March 2025 report, based on a July 2024 survey of 1,491 respondents, found that tracking well-defined KPIs for generative AI solutions had the strongest correlation with EBIT impact of the 12 adoption practices it tested, and that fewer than one in five organizations did it. (The same report found workflow redesign had the largest effect on EBIT impact of 25 attributes tested: a different cut of the data, which the operating model lesson builds on.) The 2026 survey adds that high performers are twice as likely to have defined impact-measurement processes.
  • BCG reported on 24 October 2024, from a survey with a sample of 1,000, that 74% of companies had yet to show tangible value from AI, and that leaders pursue about half as many opportunities as their peers.
  • The State of Tokenomics 2026 survey from the Tokenomics Foundation (a Linux Foundation project run in partnership with the FinOps Foundation), published in September 2026, found that proving value was the most-named challenge at 43%, and that organizations with defined ownership of AI economics were 3.7 times more likely to show value to the CFO.

Read those together. Individuals feel the gain. Organizations mostly cannot find it in profit. The practice most associated with finding it is measurement, and most firms skip it. The firms that do well place fewer bets, not more.

You will have seen the headline that 95% of AI pilots fail. Quote it only with its caveat, which the FAQ below explains; the McKinsey figures above are steadier.

Where does this lesson sit on the ladder?

This is Level 4, AI Scaled, lesson 4 of 5. In the previous lesson, AI observability: tracing, continuous evaluation and incident response, you built traces, continuous evaluation and quality scores for systems in production. Earlier in this level, the gateway from the internal AI platform lesson gave you cost per use case. Those are today's two raw inputs. This lesson turns them into a decision about money. The next lesson, the AI operating model: workflow redesign, change management and the 10-20-70 rule, takes what the ledger reveals about which initiatives pay and turns it into an operating model.

How Honest AI ROI Measurement Works: Vouchers, Tills and Three Traps

Start with five plain definitions, because the jargon hides simple ideas.

  • ROI (return on investment) is the value gained minus the full cost, divided by the full cost.
  • Attribution means showing that the change was caused by the AI system and not by something else that happened at the same time.
  • A baseline is the measurement you take before rollout, so you have something to compare against.
  • A guardrail metric is a quality measure that must not get worse while the main metric improves.
  • Portfolio management means reviewing all AI initiatives together and deciding, on evidence, which to scale, which to fix and which to stop.

Why are hours saved not the same as money?

Because hours saved are vouchers, not cash. Think of a kitchen drawer full of shop vouchers. Opening the drawer feels like wealth. But nothing has actually happened until you carry one to a till and redeem it for something.

An hour saved by an AI tool is a voucher. It is redeemed in one of three ways: the freed time was spent on other work that shipped, a cost was avoided, or a hire was not needed. If none of those happened, the voucher is still in the drawer. The hour went into a longer coffee, a slower afternoon, a queue further down the line, or simply into the same work done with less strain. That last one may be worth having, but it is not a million dollars.

The value ledger you will build applies one rule: only redeemed vouchers go in the "realized" column. Everything else goes under "projected", in pencil.

Trap one: what is the perception gap?

The perception gap is the distance between how fast people feel and how fast they are. METR's developers believed they were 20% faster and measured 19% slower. In Google Cloud's DORA 2025 report, published 23 September 2025 with nearly 5,000 respondents, over 80% said AI had raised their productivity while only 24% said they trust AI output "a lot" or "a great deal". People feel more productive using output they do not fully trust. That suggests checking is happening somewhere, and checking takes time nobody put on the slide.

What to do about it: treat a survey as evidence of sentiment and adoption. Never treat it as evidence of value.

Trap two: what is local-to-global leakage?

Local-to-global leakage means a gain at one step of a process is absorbed by the next step, so the whole process does not speed up. Imagine widening one lane of a motorway just before a toll booth. Cars reach the booth sooner and then wait longer.

Faros AI's report of 23 July 2025 shows the pattern in telemetry from more than 10,000 developers in 1,255 teams. Teams with high AI adoption completed 21% more tasks and merged 98% more pull requests. Review time rose 91%, pull request size rose 154%, and bugs per developer rose 9%. Faros found no significant correlation with company-level throughput or quality. Two caveats belong in the same breath: Faros is an engineering analytics vendor, and the data is correlational, so it shows what moved together, not what caused what.

What to do about it: measure at the level of the whole workflow, from request to delivered result, not at the step where the tool sits.

Trap three: why do hours not turn into money by themselves?

Time saved becomes value only if the capacity is redeployed to something that earns, or the cost base changes. That is the voucher rule stated in finance language. A team that saves 500 hours and has no named place to spend them has produced 500 vouchers. And "redeployed" must pass a test of its own, set out in step 5 of the workshop.

Are any of the gains real?

Yes, and they are uneven, which is why averages mislead. Brynjolfsson, Li and Raymond studied 5,179 customer support agents in "Generative AI at Work", NBER Working Paper 31161, April 2023, later published in the Quarterly Journal of Economics in 2025. Issues resolved per hour rose 14% on average and 34% for novice and low-skilled workers, with minimal effect on experienced workers.

Dell'Acqua and colleagues, in "Navigating the Jagged Technological Frontier" (September 2023), studied 758 BCG consultants. On tasks inside the frontier of what the model could do, AI users completed 12.2% more tasks and finished 25.1% faster. On a task outside that frontier, they were 19 percentage points less likely to produce a correct answer.

Put those next to METR and a consistent picture appears. The gain depends on who is doing the work and what the task is. Novices on routine tickets gain a lot. Experts in familiar territory may gain nothing. So a single company-wide productivity percentage is close to meaningless. You have to measure initiative by initiative.

What does an honest method look like?

It has six parts. Each one closes a specific trap.

  1. A baseline before rollout. Without it you are comparing today with a memory.
  2. A control group or staggered rollout. A control group is a comparable team that does not get the tool yet. A staggered rollout gives the tool to one site or team first and the others later, so the late group serves as the comparison. This is how you get attribution.
  3. An outcome metric at workflow level, such as cycle time, cost per ticket, error rate or revenue, plus a quality guardrail.
  4. Full cost counted: tokens, review time, platform and change management.
  5. A finance-signed value ledger with realized and projected value in separate columns.
  6. A portfolio review that stops or scales on evidence.

The Workshop: Run a Value Ledger Review in Seven Steps

The Value Ledger Review is a working session that ends with one document, the AI Value Ledger, signed by finance. Budget a half-day to set it up the first time, then 90 minutes each quarter.

Who needs to be in the room, and what do they bring?

This is a finance meeting with AI people in it, not an AI meeting with finance invited.

  • The finance business partner, who signs, and who owns the budget lines any banked saving will come out of. If finance is not in the room, you are producing another slide.
  • The head of AI, who chairs the session and owns the ledger between reviews.
  • The business owner of each initiative, who brings the operational metrics the business already reports.
  • The receiving manager for every redeployed benefit: the person whose team or backlog took the freed time. A two-line written note from them will do if they cannot attend. Without it, a redeployed claim cannot pass step 5.
  • A data analyst, who checks the comparisons and the baselines.
  • The platform lead, who brings cost per use case from the gateway and traces and quality scores from observability.

What are the seven steps?

  1. List every AI initiative with its owner, the one workflow outcome it is meant to move, and that outcome's baseline. One outcome, not five. If the owner says "productivity", ask: "Which number, on which report you already send, should change?" Then record the baseline and the date it was taken. If the initiative went through a pilot charter from the lesson on AI pilot design, baselines and the go or no-go decision, its baseline is already written down; use it. If there is none, write "unmeasured" and start a baseline today. Never reconstruct one from memory. An initiative with no named outcome is a hobby, and it goes on the list anyway so everyone can see it.
  2. Choose the comparison. In order of strength: a control group, a staggered rollout, or at minimum a before and after with other changes noted. "Other changes" means anything else that could have moved the metric: a new hire wave, a pricing change, a seasonal peak, a process redesign. Write them in the ledger so a reader can judge.
  3. Add one quality guardrail per initiative, and its floor. The floor is the level below which you pause, whatever the savings look like. Set it before you see the results, the same way you set a budget before the shopping trip.
  4. Count the full cost. Four lines: tokens and licences, human review time, platform share, and training and change effort. Review time is the one people forget. If a person checks every AI output, that person's time is a cost of the AI system.
  5. Classify each claimed benefit, and put every redeployed one through the CFO's test. Three classes only. Banked: a cost was removed or revenue was booked. Redeployed: capacity was moved to named work, with a name on it. Unbanked: time was saved and has no destination. If the AI system's own output removed the cost (the assistant handles the tickets, so contractor seats lapse), it is banked. If freed people do other work that removes a cost (paralegals take back work from outside counsel), it is redeployed, and it passes the test below. Then ask the question a finance partner asks before counting redeployed capacity as realized value: does the work that received the time have a value of its own on the ledger? An agency fee avoided, an outside counsel budget released, a renewal saving booked: yes. "The team got further through its backlog": no. Redeployed benefits that pass count as realized. Those that fail go in a separate realized capacity sub-column, recorded in hours with an owner, never added to cash.
  6. Finance signs the ledger, and someone reads the bottom line aloud. Realized and projected sit in separate columns; unbanked benefits are only ever projected. The signature carries a name and a date. Then compute two numbers for the whole portfolio, realized value as a share of full cost and realized ROI, and say them out loud. If the portfolio does not yet pay for itself, the minutes say so in one sentence.
  7. Decide scale, fix or stop for each initiative, and set the next review date with one measurement gap to close before it. Aim for fewer, larger bets. BCG's October 2024 survey found that leaders pursue about half as many opportunities as their peers, so a shorter list after the review is a sign the method is working. One gap per initiative: a list of ten will not get done.

What does a filled-in review look like?

Illustrative only: picture a 1,500-person business services firm reviewing four initiatives. Every figure below is invented to show the method. The old steering committee slide claimed $3.2M of projected annual value across the four.

InitiativeOutcome metric and baselineComparisonGuardrail and floorBenefit classDecision
Support assistantCost per resolved ticket; $8.40, taken the prior quarterStaggered rollout across two sitesCustomer satisfaction must not fall below 4.3Banked: two contractor seats not renewedScale
Coding assistantLead time for changes; 9 days, from delivery dataBefore and after, review time trackedDefects reaching production must not riseUnbanked: survey says hours saved, lead time flatFix: address the review bottleneck
Sales email draftingMeetings booked per rep per month; 11.4Control group of repsUnsubscribe rate must stay under 0.5%None shownStop
Contract reviewTurnaround days per contract; 6 daysBefore and afterErrors found in sampled legal review: no more than 2 per 50 contractsRedeployed: paralegal time moved to a backlog with its own value lineScale with caution

Now walk through the room, one initiative at a time.

Support assistant. The site that got the assistant first is at $7.10 per resolved ticket. The site still waiting is at $8.30 in the same quarter. That second number matters: it shows the baseline barely moved without the tool, so the drop is attributable. Satisfaction held at 4.4 against a floor of 4.3. The benefit is banked: two contractor seats were not renewed, worth $140,000 a year, and finance can see the line. Full cost is $95,000 a year: $38,000 tokens, $22,000 for quality sampling by team leads, $20,000 platform share, $15,000 training. Decision: scale to the second site.

Coding assistant. The survey says 6 hours saved per engineer per week. Across 60 engineers and 46 working weeks at $60 an hour, that is the famous $1M. But lead time for changes was 9 days before and is 9 days now, and the time a pull request waits for review has grown from 1.5 days to 2.6. This is local-to-global leakage in your own data. The engineering manager argues the hours went into the tech-debt backlog. The finance partner applies the step 5 test: that backlog has no value line of its own, so even if the claim is true it is realized capacity, not cash. Full cost is $310,000, including $60,000 of extra reviewer time. She writes $0 under realized and $1M under projected, in pencil. Nobody is accused of anything. The decision is "fix": one quarter to attack the review bottleneck, then re-measure lead time.

Sales email drafting. Reps with the tool booked 11.2 meetings a month. The control group booked 11.4. The unsubscribe rate in the AI group rose to 0.7%, through the 0.5% floor. No benefit, a breached guardrail and $60,000 of cost. Decision: stop. The owner is thanked for running a clean test, because a clean "no" saved next year's budget.

Contract review. Turnaround fell from 6 days to 3.5. Sampled legal review found 1 error in 50 contracts, inside the floor. The freed paralegal time, about 0.8 of a full-time role, was moved to the supplier renewal backlog, and its receiving manager is in the room. The step 5 test passes: that backlog used to go to an outside firm under a $55,000 annual outside counsel budget, and legal has released the budget. So finance records $55,000 as redeployed and realized. Full cost is $70,000 a year: $24,000 tokens and licences, $26,000 for the lawyers' sampled review, $12,000 platform share, $8,000 training. Contract review therefore does not yet pay for itself on realized value. The case for scaling rests on turnaround. Because the comparison is only before and after, the decision is "scale with caution", and the gap to close is a staggered rollout to the second legal team.

Now put the four on one page.

InitiativeFull cost a yearRealized valueProjected, in pencilRealized minus cost
Support assistant$95,000$140,000 bankedNone+$45,000
Coding assistant$310,000$0$1,000,000-$310,000
Sales email drafting$60,000$0None-$60,000
Contract review$70,000$55,000 redeployedNone-$15,000
Portfolio$535,000$195,000$1,000,000-$340,000

Add it up. Realized value covers about 36% of full cost this year, which is a realized ROI of about minus 64% on the definition from earlier. It is the honest starting point of a J-curve, the common pattern in which cost arrives before value, and it is why only one initiative, the only one that pays for itself today, is scaled without caution. The $3.2M slide is gone. The committee puts new money into two initiatives instead of four, gives one a quarter to fix its bottleneck, and stops one. It is a harder meeting than the old one, and the first one where the CFO's partner nods, because the numbers finally include the bad news.

Four quarters later: when pencil became ink

Picture the same committee four quarters on, still illustrative. The coding team fixed its review queue, lead time fell from 9 days to 7, and the delivery head named two planned hires the firm no longer needs: some of the pencil became ink. Contract review earned its "scale" in a staggered rollout. Two rows were stopped last quarter without drama, and every new initiative arrives with a baseline because owners know the first question. The finance partner brings the ledger to the budget cycle herself, and realized value has passed full cost.

What changes when the ledger has forty rows instead of four?

The method stays the same; the meeting cannot. Tier the rows. Initiatives above a cost threshold your finance partner sets get the full walk-through every quarter. The rest get a written row and review by exception: they come to the table only if a guardrail was breached, a baseline is still "unmeasured" after two reviews, or realized value has not moved. New requests enter through the same value, feasibility and risk scoring you used when choosing your first AI use case and avoiding pilot purgatory, and they arrive with a baseline already written. The ledger grows; the meeting does not.

What template do you take away?

The artifact is the AI Value Ledger: one row per initiative, with these fields. Copy it into the spreadsheet your finance team already uses.

FieldWhat to writeWho supplies it
OwnerOne named business ownerHead of AI
Workflow outcome metricOne number the business already reportsBusiness owner
Baseline value and dateThe pre-rollout figure, the pilot charter figure, or "unmeasured"Data analyst
Comparison methodControl group, staggered rollout, or before and after with other changes notedData analyst
Guardrail metric and floorThe quality measure and the level that triggers a pauseBusiness owner
Full cost by lineTokens, licences, review time, platform share, change effortPlatform lead and finance
Benefit by classBanked, redeployed or unbanked, each with an amountBusiness owner, challenged by finance
Value line for redeployed workThe ledger line that shows what the receiving work is worth, or "none"Receiving manager, checked by finance
Realized valueBanked, plus redeployed that passes the value testFinance
Realized capacity (not cash)Redeployed hours that failed the value test, with an ownerFinance
Projected valueEverything not yet redeemedFinance
Finance sign-offName and dateFinance business partner
DecisionScale, fix or stopSteering committee
Next review dateThe date, and the one measurement gap to close before itHead of AI

At the foot of the sheet, add one summary row: total full cost, total realized value, realized value as a share of full cost, and realized ROI. That row is the sentence the minutes record.

A ledger with three "unmeasured" cells, a negative bottom line and a finance signature is worth more than a slide with a big number and no signature.

Failure Modes, How to Spot Them, and the Metrics to Track

Which failure modes should you watch for?

  • Survey ROI. The source of the number is "users told us". Spot it by asking for the operational metric behind the claim. If there is none, it belongs in the projected column.
  • Activity as value. Licences issued, prompts sent or "percent of code written by AI" appear on the value slide. These measure use, not results. Even the widely repeated figures of roughly 30% of code at Microsoft and Google come with a caveat: reporting by SiliconANGLE on 29 April 2025 noted it is unclear how either company measures them.
  • Bottleneck shift. Output rises at one step while queue time rises at the next. This is the Faros pattern. Spot it by putting the upstream and downstream metrics on the same chart.
  • Cost-only scorecard. There is no quality metric beside the saving. In February 2024 Klarna said its AI assistant handled two-thirds of its customer service chats in its first month, the work of 700 full-time agents. The 700 figure was Klarna's own claim. In May 2025, as Entrepreneur reported on 9 May 2025, CEO Sebastian Siemiatkowski said cost had been too dominant a factor and the result was "lower quality", and Klarna began recruiting human agents again: a rebalancing to a hybrid model, not an abandonment of AI. The fix is step 3, a guardrail with a floor set in advance.
  • Unbanked hours multiplied by salary. A consultancy example reported by CIO.com in December 2025 put an unnamed client's saving at about 29,000 hours and about $1M a year: an estimate of exactly this kind, until the capacity is redeployed.
  • Redeployment on paper. Benefits are marked "redeployed" because time went to "the backlog" or "higher-value work", with no ledger line showing what that work is worth. Spot it by asking the step 5 question for every redeployed row; anything that fails moves to the realized capacity sub-column.
  • Zombie initiatives. Nothing has been stopped in the last two reviews. Either every bet you made was right, or the review is not deciding anything.

Which metrics tell you the review is working?

MetricWhat it tells you
Share of initiatives with a pre-rollout baselineWhether new projects are starting measurable
Share of claimed value that is finance-signed as realizedHow much of the story is redeemed vouchers
Initiatives stopped or merged per reviewWhether the portfolio is being managed or only reported
Ratio of full cost to realized value, per initiativeWhich bets pay for themselves today
Portfolio realized value as a share of full costWhere you sit on the J-curve, and whether it is bending upward
Guardrail breaches per quarterWhether savings are being bought with quality

Expect the second metric to be uncomfortably low the first time. That is the starting point, not a failure. What matters is that it rises review after review.

Frequently Asked Questions

How do you prove AI ROI to a CFO?

Show a finance-signed ledger, not a slide. For each initiative give a baseline, a control group or staggered rollout, a workflow outcome with a quality guardrail, and the full cost, then let finance separate realized from projected value and state the portfolio's realized ROI, even when it is negative.

Can time saved by AI be counted as cost savings?

Only when it is redeemed. Saved hours become realized value when a cost is removed, revenue is booked, or the time moves to named work with its own value on the ledger, such as an agency fee avoided; otherwise they are projected value, or at most realized capacity, never cash.

What is the AI productivity paradox?

It is the gap between people reporting that AI makes them more productive and organizations failing to find the gain in throughput or profit. McKinsey's State of AI survey, August 2026, found 80% of individuals report higher productivity while only 37% of organizations attribute any EBIT impact to AI.

Is it true that 95% of AI pilots fail?

Not as a reliable fact. MIT NANDA's August 2025 headline counts as success only pilots with rapid, measurable P&L impact, and that definition is disputed, so quote it only with the caveat; the steadier figures are McKinsey's 2026 numbers of 37% reporting any EBIT impact and about 6% high performers.

How long does it take for an AI initiative to pay for itself?

Expect cost to arrive before realized value, so judge the trend rather than a single quarter. Plan for a J-curve: track realized value as a share of full cost for each initiative at every review, and treat one whose gap has not narrowed after two or three reviews as a candidate to fix or stop.

Key Takeaways

  • Do not trust how fast people feel. In METR's July 2025 trial, experienced developers believed they were 20% faster and measured 19% slower, and METR's February 2026 update says even its own design can no longer measure the effect cleanly.
  • Hours saved are vouchers. They count as realized value only when redeemed: banked as removed cost or booked revenue, or redeployed to named work that has its own value on the ledger. Redeployed time without that value line is realized capacity, not cash.
  • Measure the workflow outcome, not the step where the tool sits, and put a quality guardrail with a pre-set floor beside every saving.
  • Take the baseline before rollout, reuse the one from your pilot charter where it exists, and use a control group or staggered rollout so you can attribute the change to the AI system.
  • Count the full cost, review time included, and read the portfolio's realized value against it aloud, even when the answer is a negative realized ROI.
  • Keep one AI Value Ledger, signed by finance, with realized and projected value in separate columns, and use it each quarter to scale, fix or stop. Fewer, larger bets is the goal.
  • The initiatives that pay usually changed how the work is done, not only which tool is used. Lesson 20 turns that into an operating model.