Developers are already living in a world of 25× productivity. An AI-native competitor is quietly rebuilding your industry with a fraction of your headcount. And most executives are still talking about Copilot.
AI is producing 25× productivity gains for the companies that have figured out how to capture them. It is producing essentially nothing for the 89% of firms that haven’t. This first issue maps the gap, the five phases of AI implementation, where most organisations get stuck, and why 85% of AI use creates no business value at all. Part 2 covers how to break through, including the investments that compound, the teams that deliver, and what boards must decide now.
Lies, Damned Lies and Statistics.
Last year Anysphere, the company behind Cursor, passed $1 billion in annualised revenue with a team small enough to fit on a single Zoom call. It is the fastest B2B company in history to that mark.1 Salesforce stopped hiring software engineers altogether after its own AI agents lifted engineering productivity by more than 30%.2 Shopify’s CEO leaked a memo in April 2025 declaring “reflexive AI usage” a baseline expectation, with no new humans hired until teams proved AI could not do the work.3 Klarna shows the other side of the ledger: having loudly replaced 700 customer-service agents with a chatbot in 2024, it was quietly hiring humans back by mid-2025 after admitting the all-AI approach had degraded quality.4
There is noise on both sides. But strip out the headline failures and the underlying signal is real, named, and citable. Set that against NBER Working Paper 34836,5 published in February 2026 and drawing on surveys of nearly 6,000 senior executives across the US, UK, Germany, and Australia. It finds that 89% of firms report no measurable productivity impact from AI over the past three years. The average cumulative gain reported is 0.29%. So is AI just hype? I think not, but it does come down to interpretation.
My issue with these findings is not that they are wrong. They report what executives told the survey. My issue is that they describe a deficiency in strategy, visibility, and focus, and the technology gets the blame.
BCG reported that firms scaling AI in software development are already seeing about 25% productivity improvement, with expectations of 44% at full scale.6
Outside of dev teams, knowledge workers are researching more effectively than search engines, admin staff are using Gen AI to compose and respond to emails, finance staff are writing AI assisted spreadsheets, but none of these efforts are showing up in tangible ways to the higher echelons and the gap between what AI can really do and what most organisations are capturing from it is not closing. There are many reasons for this, but I would propose two major factors.
1. AI is being applied to activity, not to output. Faster admin, faster collateral, faster emails: these are personal productivity, not business productivity. Example: automating a manual sales process frees up the sales team’s time but does not stimulate demand or convert more deals.
2. Time saved is absorbed, not redirected. Hours freed by AI are absorbed into the operational baseline unless they are deliberately reassigned to a revenue line or a cost line before deployment. Example: AI compresses the time to produce a sales presentation. The collateral lands faster, but unless those recovered hours are explicitly redirected to a measurable activity, they are invisible to the executive.
The point is without specific focus, time saved is not converted into output the board can see, unless it corresponds with a reduced headcount. The roadmap below sets out the five phases of AI implementation and shows the points along the way where most organisations get stuck.7
Converting AI Capacity into Board-Visible Revenue: An Executive Roadmap
AI saves time. The Board sees revenue or cost, not time. Time saved is invisible to the Board unless it converts into more units produced, more sales closed, or genuinely lower cost per unit. We exclude cost reduction from headcount cuts or scrapped vendor contracts: that is AI washing, not AI value. Otherwise, AI value is invisible to the Board.
Phase 0 — Executive Alignment: decide upfront on “more output” or “more sales”. Capture baseline KPIs. Define decision authority.
Phase 1 — Strategy & Workflow Design: identify which work is to be redesigned. Pre-assign hours saved.
Data Readiness Reality: assume data auditing and cleaning take twice as long as budgeted.
Expert System Replication: encode top-performer logic so every team member has access to the best decision-making.
Internal LLMs & Data: commercial models connected to proprietary company knowledge.
Automated Agents: purpose-built agents for high-volume, repeatable tasks.
Productivity gain: 18%–25% (BCG). Recovered time MUST flow into production or sales.
⚠ The 10-Hour Leak: Knowledge workers using AI complete tasks roughly 25% faster, equivalent to up to 10 hours of recovered capacity per week for a 40-hour role.8 Those hours are rarely assigned to a measurable activity. The Board sees no change.
The Golden Rule of Redirection: every hour saved by AI must be assigned to a specific output BEFORE deployment.
Sales Acceleration: focus on four board-visible numbers. Units per week · Sales per rep · Revenue per employee · Cost per unit.
Governance & Scaling: avoid single-vendor dependencies. EU AI Act compliance.9 Regulatory obligations built in, not bolted on.
AI gives a literal form to a familiar business idea (Jim Collins, Good to Great, 2001).10 More output produces better operational data, which trains smarter models, which generate more output. A self-reinforcing cycle.
Phases 0 & 1: Alignment and Infrastructure
Before any AI deployment can generate board-visible output, the organisation should answer three foundational questions:
1. What metric are we trying to improve?
2. Who decides what gets automated?
3. Is the data ready?
Many organisations skip this phase entirely, and spend the next twelve months wondering why their AI initiatives produce activity but no results.
The roadmap begins here deliberately. Phase 0 is executive alignment: deciding what “more output” means in specific, measurable terms: not “productivity” in the abstract, but which units, which revenue lines, which cost centres. This is where the Board Visibility Test is applied for the first time. If the initiative cannot trace to a number the CFO tracks, it should not be launched. That part is important.
Phase 1 is infrastructure: Strategy & Workflow Design. The trap to avoid is automating the existing process. Phase 1 starts from the desired outcome and asks of each workflow: should this still exist? Some are eliminated. Some are simplified. Some are redesigned. Only the survivors are benchmarked and prepared for AI deployment. The Data Readiness Reality check in the roadmap is blunt: assume data auditing and cleaning will take twice as long as budgeted. Organisations that skip this step discover the cost later, when their AI models produce unreliable outputs from unreliable inputs.
On the infrastructure side, the failure mode is the mirror image. Organisations buy GPU capacity, platforms, and vendor licences before they have decided which workflow each piece is meant to serve. The capacity arrives. The workflow does not. Infrastructure must follow the workflow, never lead it.
Phase 2: Custom AI Deployment
Phase 2 is where the roadmap shifts from preparation to execution. The framework targets three parallel workstreams: Expert System Replication: encoding top performer logic so every team member has access to the best decision-making in the organisation; Automated Agents: deploying purpose-built AI agents for high-volume, repeatable tasks such as lead qualification and contract routing; and Internal LLMs & Data: using commercial models connected to proprietary company knowledge to resolve routine queries without human intervention.
Expert System Replication is the most undervalued of the three workstreams. Every organisation has a small number of people whose judgement is disproportionately good. The rep who closes 30% above team average. The analyst whose forecasts hit. The lawyer who never misses a clause. Their logic, captured and made available through an internal AI that any team member can consult in natural language, becomes a baseline rather than an exception. The aim is not to clone people. It is to stop the best instincts in the organisation from being a single point of failure.
The critical requirement at this stage is that recovered time must flow into production or sales. The roadmap is explicit: time freed by automation that is not pre-assigned to a revenue-generating or cost-reducing activity will be absorbed. It will not appear on any dashboard. The pre-assignment happens in Phase 1. The enforcement happens in Phase 2.
“Automating a sales process will not create more sales. You still need the demand.”
— Darryl ClaretThe Invisible Productivity Trap
The Trap is not a phase. It is the failure mode that Phases 0 and 1 exist to prevent. When Phase 0 and Phase 1 are skipped and AI is deployed in Phase 2 anyway, the result is consistent: AI is genuinely saving time, but that time is being absorbed invisibly rather than redirected into business output. Most organisations are here.
The trap is mundane. Each person on the team finds an hour or two a day they did not have before. Some of it flows into deeper work. Most of it disappears into a longer lunch, an extra coffee, an easier pace. The hours arrive. The output does not. No one designed the deployment to redirect those hours to a specific measurable activity before the rollout began.
I own an enterprise cost-analysis workflow for upstream projects. Until recently it was cumbersome. It ran every cycle and on every ad-hoc update. With so many people working on it and source files arriving from successive iterations, reconciling whether the numbers were even valid had become a workstream of its own. Every cycle, new and more ingenious errors crept in.
The Trap move was the obvious one: ask Copilot to help me with the spreadsheet, save an hour or two, feel productive. Instead I removed the spreadsheets from the equation. Python and an LLM now read the unstructured source files directly and feed a JavaScript dashboard for scenario analysis. A final step publishes the working model back into Excel with sense-check proofs that derive the same numbers using Excel formulas in parallel, for those that are uncomfortable and yet to trust the “black box”. The reconciliation work is gone. The errors are gone. The spreadsheet is gone.
What matters is what came next. We did not absorb the saved time into longer lunches. We redirected it, deliberately, into activating our resources from administration to actual project development and standing up the projects. The team is also visibly happier. Removing manual, error-prone work tends to do that. Part 2 returns to why that matters financially. The Golden Rule of Redirection is not aspirational. You design where the time goes before deployment, or it disappears.
The Golden Rule of Redirection: Every hour saved by AI must be assigned to a specific widget or a specific sale before deployment. Time savings must have a pre-assigned destination. Without this, AI value is zero on the board’s spreadsheet, regardless of how much individual productivity has improved.
This is why the Board Visibility Test at the top of the roadmap is binary: “Same output, faster” is board-invisible: it registers as zero revenue change. Only “more units produced” or “faster revenue closed” registers. An AI deployment that cannot answer this test before go-live is, almost by definition, destined for the Trap.
Most AI Use Creates No Business Value
Most AI use never reaches the business. MIT’s NANDA report, published in August 2025, put the failure rate of corporate generative AI initiatives as high as 95% against $30-40 billion in enterprise spending.11 The five most common AI activities, rewriting emails, quick code fixes, using AI as a search replacement, fixing Excel formulas, and summarising documents, are valuable to individuals and essentially invisible to the business.
Usage metrics, how often employees open the tool, how many seats are activated, are vanity metrics at best. The meaningful signal is which workflows have been transformed, automated, or eliminated, and what business output that created. A task automated 500 times a month generates compound value. A task where someone opened ChatGPT for five minutes generates none. Phase 3 of the roadmap (covered in detail in Part 2) targets four specific numbers: units per week, sales per rep, revenue per employee, cost per unit. These are the metrics that move boards to act.
What does the minority that gets it right look like?
In the mid-2010s I sat with the CTO of a fund administrator who had reframed his customer services and related operations as an “open kitchen”: the partners being onboarded could see straight through the wall to where their work was, or wasn’t, being done. We designed a portal that exposed every step of the flow, every blocked document, every waiting signature. It was elegant. It was also only a design. The technology existed in pieces, but doing the work behind the glass, the document validation, the chasing, the routing, the judgement calls, was prohibitively complex. We could open the window onto the kitchen. We could not yet staff the kitchen.
I revisited the same idea recently with a team of students, on a partner-onboarding workflow for joint ventures: a sophisticated, document-heavy, legally and financially sensitive process that most partners detest because it is opaque. The working prototype puts agentic employees inside the kitchen. They guide each partner through the steps in plain language, validate documents in seconds, chase what is missing, and keep everyone in the loop without anyone having to ask. Humans stay on the judgement calls and the relationship. The agents handle the choreography.
The point is not the time saved per partner, although it is significant. The point is that the partner can see the work happening, and so can the board. If your AI is invisible to your customer, it is almost certainly invisible to your board too.
So the paradox resolves itself. AI is working. The value is invisible. Organisations are stuck in the Trap because they skipped the prep work in Phases 0 and 1 that would have made saved time visible to the board.
A version of this conversation will look baffling to executives ten years from now. We did not ask for an ROI justification before giving a knowledge worker a computer. We did not calculate the productivity delta of providing email access. These are infrastructure, the baseline conditions under which modern work happens. Companies will continue to survive without AI. They will simply do it on an uneven playing field they did not choose, against competitors who treated AI as infrastructure from day one.
Part 2 takes on the question most boards are not yet asking: what if the most valuable AI returns are the ones that never reach the P&L?
Sources
- Bloomberg, “Anysphere, Hailed as Fastest Growing Startup Ever, Raises $900 Million” (5 June 2025). bloomberg.com. Cursor reached $100M ARR in January 2025, $500M by June, and crossed $1B by November 2025 with a team of roughly 150. ↩
- The San Francisco Standard, “Marc Benioff says Salesforce will hire no engineers this year due to AI” (27 February 2025). sfstandard.com. See also Salesforce Ben, “Salesforce Will Hire No More Software Engineers in 2025, Says Marc Benioff.” salesforceben.com. ↩
- CNBC, “Shopify CEO says staffers need to prove jobs can’t be done by AI before asking for more headcount” (7 April 2025). cnbc.com. Memo confirmed by Tobi Lütke on X: x.com/tobi/status/1909251946235437514. ↩
- Bloomberg, “Klarna Turns From AI to Real Person Customer Service” (8 May 2025). bloomberg.com. See also Fortune, “Klarna plans to hire humans again, as new landmark survey reveals most AI projects fail to deliver” (9 May 2025). fortune.com. The original 700-agent claim was made by Klarna in February 2024. ↩
- Yotzov, I., Barrero, J. M., Bloom, N., et al., “Firm Data on AI.” National Bureau of Economic Research Working Paper 34836 (February 2026). nber.org/papers/w34836. Survey of nearly 6,000 senior executives at firms in the US, UK, Germany, and Australia. 69% of firms report active AI use, but nine-in-ten executives report no measurable impact on employment or productivity from AI over the past three years; average cumulative productivity gain reported is 0.29%. ↩
- Boston Consulting Group, “How AI Is Paying Off in the Tech Function” (2026). bcg.com. Two-thirds of surveyed companies are using AI in software development; 36% are scaling or have fully deployed it. Firms at scale are seeing about 25% productivity improvement, with expectations of 44% at full scale. ↩
- Executive Roadmap Framework. Proprietary framework developed by Darryl Claret. The five-phase model spans Phase 0 (Executive Alignment), Phase 1 (Strategy & Workflow Design), Phase 2 (Custom AI Deployment), Phase 3 (Growth & Scale), and Phase 4 (the Compounding Flywheel). The Invisible Productivity Trap is positioned as the failure mode that Phases 0 and 1 are designed to prevent, not as a phase in itself. The framework’s anchoring concepts (the Board Visibility Test and the Golden Rule of Redirection) are part of an in-house framework used in advisory engagements. ↩
- Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., Lakhani, K. R., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” Harvard Business School Working Paper 24-013 (September 2023). hbs.edu. Field experiment with 758 BCG consultants found that for tasks within the AI’s capability frontier, consultants completed 12.2% more tasks, finished 25.1% faster, and produced output rated 40% higher in quality. ↩
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024, the European Union Artificial Intelligence Act. Published in the Official Journal on 12 July 2024; entered into force on 1 August 2024. eur-lex.europa.eu. Phased applicability: prohibited practices and AI literacy obligations from 2 February 2025; general-purpose AI model rules from 2 August 2025; full applicability from 2 August 2026; high-risk systems embedded in regulated products from 2 August 2027. ↩
- Collins, Jim. Good to Great: Why Some Companies Make the Leap... and Others Don’t. Harper Business, 2001. Collins’s flywheel framing describes enduring corporate greatness as the cumulative effect of many small, consistent pushes rather than a single decisive breakthrough. The metaphor was independently sketched by Jeff Bezos for Amazon’s strategy around the same period (Brad Stone, The Everything Store, Little, Brown, 2013). ↩
- MIT NANDA, “The GenAI Divide: State of AI in Business 2025” (August 2025). Reported in Fortune and Harvard Business Review. Based on 150 leader interviews, a 350-employee survey, and analysis of 300 public AI deployments. Found that despite $30-40 billion in enterprise spending on generative AI, only about 5% of AI pilot programmes deliver rapid revenue impact; the remaining 95% stall. Root cause identified as flawed enterprise integration and a learning gap rather than model quality. ↩