It is the last session of the strategy offsite and a board member asks the question everyone has circled for two days: "Where is all this going to be in three years?" Two people answer. The first, an enthusiastic function head, says the work will be unrecognizable by 2028 and the organizations that hesitate now will not catch up. The second, a well-respected operations director, says the hype has run ahead of reality and the sensible move is to let the dust settle and buy the winning technology once it is obvious. Both answers land. Both sound like judgment. And both are the same mistake in different clothes: an attempt to answer a question nobody in the room, in the industry, or in any research lab can answer, followed by a multi-million-dollar commitment justified by that answer. This lesson is about the third posture, the one used by every serious planner working under genuine uncertainty, and about the artifact it produces: a portfolio of investments that pay off no matter which future arrives.

The Two Useless Postures

Every readiness leader gets handed the same two scripts, usually by sincere people. Breathless certainty says the curve is exponential, the capability is arriving, and the only rational response is to commit hard and now. Its output is overcommitment to whatever is fashionable: an org chart redrawn around a technology shape that is eighteen months old, a five-year platform contract signed at the peak of a vendor's story, a headcount plan built on a capability demonstrated but not deployed. Certainty feels like leadership. It produces decisions extremely sensitive to being wrong.

Dismissive caution says the hype always overshoots and the disciplined move is to wait for clarity. Its practical output is the organization that wakes up structurally behind: not behind on tools, which can be bought in a quarter, but behind on everything that takes years to build. Caution also feels like leadership. It is frequently the more expensive of the two.

Notice what they share. Both answer the question "how capable will these systems become, and how fast?" That is not answerable by you or reliably by anyone; the people building the systems disagree in public. Any plan whose value depends on getting it right is a coin flip dressed as analysis.

There is a well-worn discipline for this situation, and it did not come from technology. It came from capital-intensive industries committing billions to assets with thirty-year lives under uncertain prices and regulation, and it starts by giving up the thing everyone wants. You stop forecasting. Then, in order: write down the small number of facts that are already determined and require no prediction; sketch a handful of plausible scenarios without ranking them by likelihood; and sort every proposed investment by whether it holds value across all of them.

That last category has a name: no-regret investments. They are the ones you would make even if you knew for certain that the most boring scenario was the real one, and they are where a readiness leader's planning energy belongs. Here is the encouraging finding this lesson will earn rather than assert: nearly everything in the no-regret bucket is something this program has already taught you to build. Not a sales pitch, a structural consequence. Process documentation, instrumented baselines, certified data, redesign capability, an evidence layer, and clear decision rights are properties of the organization, not of the technology, and they do not become obsolete when a model release changes what is possible. They determine whether you can use anything at all.

What Is Already Determined

Start with the short, honest list. Admission is strict: an item belongs only if it is true regardless of how the technology develops. Anything requiring a view about capability is a forecast wearing a fact's clothes and belongs in the scenario section instead. Four items pass.

First: the regulatory calendar has dates on it

You do not need to predict regulation. Some is already written, with a published timetable. The European Union's AI Act, on the post-Digital-Omnibus calendar agreed in May 2026 and still pending formal adoption, sets out a sequence:

• Since August 2, 2025: obligations on GPAI (general-purpose AI, the broad foundation models sitting underneath most enterprise tools) are in force.
• December 2, 2026: transparency obligations for AI-generated content.
• December 2, 2027: obligations for high-risk systems listed in Annex III, covering employment, credit, essential services, and similar consequential decisions.
• August 2, 2028: obligations for Annex I systems, meaning AI embedded in regulated products.

Carry the caveat honestly: this calendar has already moved once and may move again. What does not move is the direction. Nobody credible proposes that documentation, risk classification, and human oversight requirements will be withdrawn. Whatever the dates settle at, an organization that can produce a risk classification, a record of who approved what, and evidence that a human oversaw a consequential decision is better positioned under every version of the calendar.

Add the observable trend, labelled as observation rather than projection: other jurisdictions and sector regulators are adding requirements and converging on similar demands. And add the forcing function most organizations feel first, which is not a regulator at all. It is a customer. Enterprise procurement questionnaires ask today what AI is in this product, what data it touches, who reviews its outputs, and what happens when it is wrong. That arrives in your sales cycle long before any inspector arrives at your door, and a deal timeline is a sharper deadline than a statutory one.

Second: the documentation and evidence expectation only rises

Across every regulatory regime, customer questionnaire, insurer's renewal form, and internal audit charter, the same demand escalates: show us how this decision was made, by whom, on what basis, and how you would reconstruct it a year later. No plausible future reverses this. There is no scenario in which the world asks for less traceability about consequential automated decisions.

That makes the evidence layer a rare thing in strategy: an investment whose value is monotonic. It gets more valuable if regulation tightens, if AI use expands, or if customers get more demanding, and it holds value if everything stalls, because internal audit and your own incident reviews want the same artifacts.

Third: your own process debt is exactly where you left it

This is the most reliably predictable fact in the lesson, and the least glamorous sentence in this program. The undocumented processes, unmeasured baselines, and unowned data that exist in your organization today will still be there in 2028 unless a named person is funded to fix them. No model release fixes them. No platform migration fixes them. They do not decay, resolve, or quietly improve while everyone is looking at the demos.

They also constrain what you can adopt, whatever becomes available. A capability that requires clean, owned, documented inputs cannot be used by an organization that lacks them, however good it is. Gartner's finding that 63 percent of organizations lack or are unsure about AI-ready data practices is not a story about data teams. It describes the binding constraint on two thirds of the market's ability to use whatever arrives next.

And note the arithmetic. Suppose the most aggressive capability scenario proves right. Who benefits? Not the organization that spent those years watching, but the one that can point a new capability at a documented process with a measured baseline and a named owner and be in production in weeks. Process debt is the determined-list item fully within your control, which is why it most rewards attention now.

Fourth: the human capability lag does not compress

Building redesign density, the number of people who can take a process apart and rebuild it around a new capability rather than bolting the capability onto the old shape, takes years. It is learned by doing, on real engagements, and there is no way to buy the reps. This holds whatever the technology does, and becomes more binding the faster capability moves, because a fast landscape demands more redesign per unit of time, not less.

BCG's 10-20-70 rule (10 percent of the effort in algorithms, 20 percent in technology and data, 70 percent in people and process) is the arithmetic behind this. McKinsey's State of AI work supplies the evidence: 88 percent of organizations use AI regularly, only about 39 percent report any EBIT (earnings before interest and taxes) impact, and the strongest differentiator among high performers is fundamental workflow redesign. That is a fact about people, not models, which is why it sits on the determined list.

The Scenario Sketch, Handled as Reasoning

What follows is not a forecast. These are three plausible shapes the next few years could take, sketched so you can test investments against them. They are not exhaustive, not ranked by likelihood, and no serious planner claims to know which is arriving. Their job is to be stress tests. If you find yourself arguing about which is right, you have misused them.

Shape one: capability plateau with continued diffusion

Models improve incrementally, and the action moves to diffusion: integration into workflows, better tooling around the same core capability, steady adoption. Here the constraint on value is almost entirely organizational, and this is the world the current evidence base describes. MIT's finding that roughly 95 percent of enterprise generative AI pilots produced no measurable profit-and-loss return, with only about 5 percent of custom tools crossing from pilot into production, is a diffusion failure, not a capability failure. In this shape BCG's 70 percent dominates, and every discipline in this program is exactly, unglamorously right.

Shape two: rapid capability expansion

Systems become substantially more capable, more work becomes agent-shaped (multi-step, tool-using, executing rather than only drafting), and the frontier moves faster than most organizations can absorb. The interesting thing is what this does to the constraint: it does not remove the organizational bottleneck, it relocates it. The binding constraints become oversight capacity (how many autonomous workflows humans can meaningfully supervise), permission and identity infrastructure (what each non-human actor may touch), containment (how fast you can stop one), and absorption bandwidth.

One data point behaves like a preview. Gartner's stated prediction is that over 40 percent of agentic AI projects will be canceled by the end of 2027. It is a prediction, cited here as one, and what makes it worth citing is not the number but the implied mechanism: projects get canceled when the organization cannot absorb them, not when the model cannot perform. In this shape, the readiness bar from the previous lesson stops being an advanced topic and becomes the operating floor.

Shape three: consolidation and commoditization

AI capability arrives as a standard feature inside the software your organization already runs, rather than as a separate category of thing you evaluate and buy. Here is the observable-fact part: this is already partly happening. Capabilities that were standalone products two years ago now ship inside office suites, service desks, customer relationship platforms, and enterprise resource planning systems most organizations already license. That is a description of the present, not a claim about the future.

If this shape dominates, the buy-versus-build calculus shifts further toward buy, much of the tooling question resolves itself without a procurement exercise, and the differentiator becomes entirely the process and data work, because everyone holds the same capability inside the same platforms. The only thing left to compete on is whether your processes are documented, your data is trustworthy, and your people can redesign work faster than the competitor with the identical license.

Two disciplines for using these. They are not mutually exclusive: a world where core language capability commoditizes while agentic execution expands rapidly is entirely coherent. And different functions may experience different shapes at once, with finance operations in a plateau while customer service lives through an expansion. Planning that assumes one weather system across the enterprise will be wrong somewhere.

The Artifact: The No-Regret Portfolio

Here is the lesson's deliverable. The No-Regret Portfolio sorts every investment in your plan into three buckets, using one test anyone in the room can apply out loud.

The sorting test. For any proposed investment, ask three questions in sequence:

1. Does it retain value if capability stalls and the next three years bring incremental improvement plus slow diffusion?
2. Does it retain value if capability accelerates and much more work becomes agent-shaped?
3. Does it retain value if capability commoditizes into platforms you already own?

Three yeses: no-regret. Two: pays under most shapes, fund it with a normal business case. One: it is a bet. Bets are legitimate, and a portfolio with no bets has given up on upside. But a bet must be named, sized so being wrong is survivable, and staged so it has review points where you can stop. The failure mode this test prevents is not betting. It is a bet smuggled into the plan wearing the word "strategy."

If an investment only pays off in the future you predicted, it is not a strategy. It is a bet, and bets get sized, staged, and named.

Bucket one: no-regret

These pass all three questions. Each carries its reason, which makes the test repeatable rather than a matter of taste.

Bucket two: pays under most shapes

These clear two of three: strong under expansion and commoditization, merely useful under a plateau. Enough to fund on a normal business case, with the honest note that the plateau case is thinner.

• Permission and identity infrastructure for non-human actors. Under expansion it is the difference between operating and stopping. Under commoditization it is required anyway, since embedded capability still acts under some identity. Under a plateau it is prudent hygiene.
• Containment capability, the tested ability to stop, roll back, and scope-limit an automated workflow quickly. Vital if execution expands, good incident practice otherwise.
• The pattern-review skill, the shift from reviewing individual outputs to reviewing behaviour across a population of them. Essential at agent volumes, useful at any volume, over-engineered if volumes stay small.
• Platform portability work, keeping prompts, evaluation sets, process logic, and data contracts in forms that survive a vendor change. Pays handsomely under commoditization or consolidation, costs a little under a plateau.

Bucket three: genuine bets

These pay under one shape. Again: legitimate, and sometimes correct. But sized, staged, named.

• Deep commitment to a single vendor's ecosystem, where switching means rebuilding rather than reconnecting. Pays if that vendor's shape wins.
• Large internal builds of capabilities the market may commoditize. The base rate is unkind: MIT found externally partnered solutions succeeded roughly twice as often as internal builds. Building something that ships inside your existing suite in eighteen months is a bet on commoditization not arriving.
• Org structures designed around one technology shape. A reporting line built entirely around an "agent operations" function is a bet on expansion; an AI team folded wholly into IT is a bet on commoditization. Structures are expensive to unwind and set expectations you manage for years.
• Any multi-year plan whose value depends on a specific capability arriving on a specific timeline. If the plan collapses when that capability slips two years, it is a bet regardless of how the deck labels it.

One discipline makes this bucket work: every bet gets a cap, an owner, and a written review date. The cap is what you are willing to lose. The review date is when you look again, with evidence, and continue, resize, or stop. A bet with a review date is a managed position. A bet without one is a funded belief.

Worked Example: The Annual Horizon Review

The numbers below are illustrative, built to show the shape of the exercise rather than to promise results. They follow the enterprise storyline used through this level: a mid-sized organization, a central AI function, roughly thirty use cases over three years. The review takes a half day, once a year, with the program lead, the data owner, the risk lead, and two function heads.

Step one: check the determined list. Three items had moved. The December 2026 transparency milestone was now inside the planning year rather than comfortably beyond it, converting a policy discussion into a dated delivery item. Two enterprise customers had sent AI-specific procurement questionnaires that quarter, a first, both asking for named human oversight and a description of the data touched. And the process-debt inventory showed 3 of 6 in-scope functions still had no current process documentation, defined as reviewed within eighteen months.

That third item was the least glamorous line in the review and it drove the year's largest investment: a funded, named documentation and instrumentation effort across the three lagging functions. Nobody found it exciting. Everybody could see it passed all three sorting questions, sat within their control, and that no plausible future made it unnecessary.

Step two: sort the portfolio. Fourteen proposed and in-flight investments ran through the three questions. Twelve landed where their sponsors expected. Two were reclassified, and those two are where the practice earned its half day.

The first was a proposed deep integration with one vendor's agent framework, presented as "our agentic strategy." It failed question three cleanly: if that capability arrives embedded in platforms already licensed, most of the integration work is wasted. It was not killed, because the expansion shape is plausible and being early would matter. It became a bet: capped at a defined pilot in one function, with a named owner and a nine-month review date carrying two pre-committed criteria. The sponsor's comment was the useful part: reclassification did not dent their enthusiasm, it moved the conversation from "do you believe in this" to "what would we need to see."

The second ran the opposite way: pause data-foundation remediation for two quarters pending "clarity on where AI is going." It sounds prudent. It failed decisively, because the foundation pays under all three shapes and has no dependence whatsoever on the clarity being waited for. The pause was rejected on that ground in under five minutes, because the test made the argument mechanical rather than political.

Step three: measure absorption speed. Time from charter approval to first verified movement in a business number had been 31 weeks on the first use case three years earlier. On the most recent it was 9 weeks, attributable to the flywheel's couplings rather than any tool: most new use cases now start on certified data and an existing instrumented baseline instead of building both.

The review adopted that figure as the board-level indicator for the year, targeting under 10 weeks as the portfolio grew. The framing used with the board is worth copying: we cannot tell you what capability will exist in 2029 and will not pretend to, but when something valuable appears this organization takes about nine weeks to get it into production and measured, against thirty-one three years ago, and that number is what we manage.

Step four: write down the beliefs. The program lead wrote a one-page note: what they believed about the next two years, why, what would change their mind, and a review date twelve months out. It was filed with the review pack, not circulated as a prediction.

The Failure Story: The Organization That Waited for Clarity

A mid-size manufacturer, roughly 1,400 people, decides in 2025 to pause AI investment until the landscape settles. This is not a stupid decision by unserious people. The leadership team is thoughtful, they have read the failure record honestly (S and P Global's finding that 42 percent of companies scrapped most of their AI initiatives in 2025, up from 17 percent the year before, is quoted in their own board paper), and they conclude being second is cheaper than being wrong. The reasoning is defensible except for one hidden assumption.

They pause. Their competitors do not.

Two years later they resume, and it goes badly in a specific, instructive way. The gap is not tooling; tooling they buy in a quarter, better and cheaper than what existed in 2025. The gap is everything with a multi-year build time. No current process documentation, so the first question ("what does this process cost per unit today?") has no answer. No instrumented baselines, so nothing they attempt can be proven to have worked. No data ownership, so the first use case spends four months establishing who decides what a customer record means. No trained practitioners, because redesign density is learned on engagements they never ran. No governance apparatus, so every approval is argued from first principles.

Their first serious attempt spends eighteen months building foundations before producing a measurable business result. A competitor of similar size, running the same tools bought at the same time, ships its fourth use case in that window, having spent the pause years documenting processes and certifying data.

The lesson is precise and uncomfortable enough to state flatly. Waiting for clarity about the technology was reasonable. The error was applying that wait to investments that had nothing to do with the technology. Process documentation does not depend on which model wins, data ownership does not depend on the pace of capability, and trained practitioners do not depend on the vendor landscape. The no-regret list was available the entire time, required no forecast, and was declined on the strength of a forecast about something else.

The Planning Practice, and the Number That Outlasts the Forecast

The portfolio is not a document you write once. It is a short annual cadence that should take less than a day.

1. Revisit the determined list. Has anything moved from bet to determined? A regulatory date that firmed up, a customer requirement that became standard in your sector's contracts, a capability that stopped being speculative and started shipping inside software you already own. That movement is one-way and the most useful thing the review detects.
2. Re-sort the portfolio. Run the three questions against every investment in the current plan, including the ones you sorted last year. Investments migrate between buckets as the world changes, and the migrations are the signal.
3. Re-ask the one question that matters most. What is our absorption time for a new capability, measured from decision to verified business result?

That third question is the central takeaway of this lesson and arguably of the level. You cannot control how capable these systems become, or when. You can control, precisely and measurably, how long your organization takes to turn an available capability into a changed business number. An organization that absorbs something valuable in nine weeks rather than nine months wins under every scenario here: under a plateau because it extracts more from a slow-moving capability set, under expansion because absorption is the binding constraint by definition, and under commoditization because when everyone holds the same capability the only difference left is who puts it to work first.

Stop forecasting the capability. Start measuring how fast your organization can absorb one.

This is time-to-value from the flywheel lesson viewed through a different lens: there it was evidence the flywheel was turning, here it is your most durable competitive measure, because its meaning does not change whichever shape arrives. Put it on the board's page. It is more honest than a capability forecast and far more actionable than a maturity score.

The Intellectual Honesty Note

One last thing, offered sincerely, because this lesson has spent several thousand words urging care about confident claims and it would be absurd to exempt itself. Everything here is anchored to a dated record: MIT's 2025 finding, Gartner's and McKinsey's 2025 surveys, S and P Global's 2025 figure, an EU calendar agreed in 2026 and pending formal adoption. That record will change, some of these numbers will look different in three years, and a few of this lesson's judgments will prove wrong. The honest move is to say so in advance rather than quietly update later and pretend the older view was never held.

So apply the program's evidence discipline to yourself. Write down what you believe about the next two years, in one page: the belief, the reasoning, the supporting evidence, and the specific evidence that would change your mind. Put a review date on it, twelve months out, and file it where you will find it.

The value of this practice is not accuracy. It converts being wrong from an embarrassment into a learning event with a date attached. A leader who wrote "I expect agentic workflows to be routine in customer operations by mid-2028, because of X, and I would abandon that view if Y" and reviews it on schedule becomes a better forecaster and a more trustworthy voice. A leader who only asserts the current consensus has no record and learns nothing structural. Treat every projection as provisional, including this one, and say so out loud. That habit is the credibility the next lesson builds on.

What to Do Monday Morning

1. Write your own determined list, on one page, then audit it and cross out anything that is a forecast in disguise. Four categories to check: regulatory dates that genuinely apply to you, evidence expectations from customers and auditors, your process debt, and your redesign density.
2. Run the three-question sorting test on every investment in your current plan. Stall, accelerate, commoditize. Do it out loud with the sponsors present, and write down the reason for each verdict, not just the verdict.
3. Reclassify anything that only pays under one shape as a bet, and size it. Give each a cap, a named owner, and a written review date with pre-committed criteria. Say the word "bet" in the meeting; it moves the conversation from belief to evidence.
4. Measure your absorption speed for the most recent use case and the earliest you can reconstruct: decision to verified business result, in weeks. Put both on the board's page with the trend, and attribute the difference to your foundations, not your tools.
5. Write down what you believe about the next two years, with the reasoning, the disconfirming evidence, and a review date twelve months out. File it with your annual review pack.

Key Takeaways

• Reject both useless postures: breathless certainty overcommits to whatever is fashionable, dismissive caution produces the organization that wakes up structurally behind, and both answer a capability question nobody can answer.
• Separate what is determined from what is forecast, admitting only four kinds of item: dated regulatory obligations (GPAI since August 2, 2025; transparency December 2, 2026; Annex III high-risk December 2, 2027; Annex I embedded August 2, 2028, pending formal adoption), the rising evidence expectation, your own process debt, and the multi-year human capability lag.
• Treat process debt as the most predictable fact available: undocumented processes, unmeasured baselines, and unowned data will still be there in 2028 unless a named person is funded to fix them, and they constrain what you can adopt regardless of what arrives.
• Use the three shapes as stress tests rather than predictions (plateau with diffusion, rapid expansion, consolidation with commoditization), remembering they are not exhaustive, not mutually exclusive, and may hit different functions differently at once.
• Sort every investment with the three-question test: does it hold value if capability stalls, if it accelerates, and if it commoditizes? Three yeses is no-regret, two is a normal business case, one is a bet.
• Fund the no-regret bucket first: current process documentation with default instrumentation, certified data coverage, redesign density in people, the evidence layer, governance and decision rights, and the flywheel couplings that make capability compound.
• Name bets as bets and give each a cap, an owner, and a review date, because vendor lock-in, internal builds of commoditizable capability, and structures designed around one technology shape are legitimate positions but ruinous when smuggled in as strategy.
• Manage absorption speed, from decision to verified business result, as your most durable competitive measure, and write your beliefs down with a review date so being wrong becomes learning rather than embarrassment.