The Third Horizon Is One Sentence

Most multi-year plans I am handed have a property that ought to worry whoever signed them. The third year is specified as finely as the first. Same grain of detail, same count of bullets. It reads as rigour. It is close to the opposite, because a plan that already knows what will be true three years out has assumed, without saying so, that the first two years will teach it nothing.

An internal digital strategy proposal from a global pharmaceutical company handles its far horizon differently, and does it plainly enough that you can count it.

The document sets out an operational strategy and roadmap across three years, written for a governance forum, with one business objective per year and a single fixed statement of purpose above all three. Under each year it lists key priorities. The near year carries four. The middle year carries four. The far year carries one line: integrate the platforms scaled in the middle year, and work with the commercial operations function to run localised experiences across markets.

One line. Not a shorter list. A single sentence, where each of the two nearer columns carries four itemised commitments.

Count the itemised commitments before you read them

There is a diagnostic in this that needs no domain knowledge. Count the itemised commitments under each year of any multi-year plan, and read the shape of the counts before you read the content.

A plan that steps down from four to four to one is telling you its authors know the difference between a decision they can make now and one they can only make once something has happened. A plan whose three columns carry the same weight of detail is telling you that its far year was written in the same room, in the same week, by the same people, with exactly the information available on the first day, and that nothing in its structure allows the far year to change.

The second pattern is much more common, and it is rarely dishonesty. It is a formatting habit. Templates want symmetry. A steering group reading three columns finds one thin column uncomfortable, and the cheapest way to relieve the discomfort is to write more. The invented detail then hardens. It gets costed, resourced and repeated in a status pack, until somebody is defending a line item that began as a layout decision.

The far-horizon sentence is not vague, which is the part people skip past. Vague would be aspiring to an integrated future. This names which prior state must exist, which function must be in the room, and what the output is measured as. Three hard constraints on a year nobody can schedule yet. What it declines to supply is the work.

The near horizon earns its detail

The near year is where this plan spends everything, and the spending is unglamorous. Deliver on outstanding commitments to a platform already in flight. Run a set of named local projects, six of them. Enable indirect impact through guidelines and best practices. And, fourth of four, establish a process for managing future digital demand.

Five outcomes are named against those priorities, and one of them is labelled interim in the plan's own words: a stopgap capability across the main channels, written down as temporary, with its replacement already sitting in the middle year. Nobody labels a first-year deliverable interim by accident. It is an advance admission that the thing being built will be thrown away, which is the only way a stopgap ever gets thrown away on schedule.

The guidance backlog on its own page is the same instinct again. Fourteen assets, each named individually, sorted into two colour-coded groups with three of them ringed as the highest priority. That is a portfolio rather than an intention. It would also be absurd written for the far year: nobody can list the fourteen documents an organisation will need three years out.

Enumeration is cheap and accurate a year out. It gets expensive and wrong at three. This plan spends its detail where detail is still true.

What a far horizon is actually for

Technology, content and process are tracked at every horizon, so no stage can be recorded as done because a system went live. That discipline is what makes the thin far year defensible rather than lazy.

Follow the content layer across the three years and the design shows. The near year produces guidance on good content across channels, plus a plan for non-branded external material and a solution to sourcing it. The middle year produces a scaled internal and external library held globally with a process for localising it. The far year produces the same sentence as the middle year, word for word. The content layer is finished in the middle year and then deliberately held still.

The process layer in the far year restates the priority line itself. Only the technology layer adds something that was not already committed: the scaled platforms joined together, fed by a single spine of activity data.

Read as a work plan, a year whose content outcome is a copy looks like an unfinished slide. Read as a design, it is the most disciplined page in the document. The far year is not where new work gets invented. It is where the work of the two nearer years is required to converge, and the plan says so by refusing to add anything to it.

4 / 4 / 1

Itemised priorities under the near, middle and far horizons

14

Guidance assets named on the near year's publication roadmap

3

Outcome layers tracked at every horizon: technology, content, process

6

Named local projects, all of them in the near year

The plans crossing my desk now have the symmetry problem

Almost every AI plan I have been asked to read this quarter is laid out with the symmetry that this pharmaceutical plan avoided, and the exposure is worse, because the underlying capability is moving faster than the plan's own review cycle.

The shape is familiar. A three-year build-out costed by quarter. A retrieval-augmented assistant for one function in the opening months, more functions in the following year. A vector store selected in a named quarter. An evaluation set matured on a date. Copilots extended to further functions on a schedule. Then, in the final year, a line about carefully scoped autonomous execution inside a workflow that has not yet been measured once.

Every item in that final year was written using what is known today about what a model can do, what inference costs, and what an evaluation set has to contain before anyone will trust its verdict. All three of those are moving. The retrieval design I would specify for a document-heavy function today is not the one I would have specified a year ago, and the difference is not a refinement, it is a different architecture with a different cost base. Writing that choice into a third-year budget line does not make it stable. It makes it a commitment to being wrong on a date.

The ground moves in ways that have nothing to do with model capability, too. European legislators reached political agreement on an AI act at the end of last year. Nothing is in force, the text is still settling, and the classification that will apply to a given internal system is not yet readable off a page. A third-year compliance deliverable written this month is written against a document that does not yet exist in the form it will take.

Four priorities, four priorities, one sentence. The plan spends its detail where detail is still true, and then it stops. Symmetry across three columns is a layout habit that gets costed, resourced and defended long after anyone remembers it was a layout habit.

How to write a third horizon you will not have to defend

Fix the purpose and let the priorities decay. The statement of intent above this plan does not change across the three years; only the operating ambition escalates, and only the nearer years get itemised. A thin far horizon under a fixed purpose is a plan. A thin far horizon under a drifting purpose is an absence.

Make the far year name a dependency and a partner rather than a deliverable. What has to be true before that year can start, and whose agreement does it require. Both are answerable today. The work is not.

Give the far year a decision date instead of a schedule. Say when its itemisation will be written and what evidence from the nearer years will be on the table when it is. That turns an empty column into a scheduled commitment to think, which is what it always was.

Put the intake mechanism in the near year. The fourth of the four near-year priorities here is a process for managing future digital demand, built before the demand exists. Every organisation that runs a successful first assistant discovers within a quarter that every other function wants one, and whether that arrives as a portfolio or as a queue nobody designed was decided a year earlier. Building that intake ahead of the demand is the first block of work in a RealAI Platform engagement.

Write down what would make you rewrite the plan. One with a stated trigger for its own revision survives contact with a moving capability frontier. One without gets quietly abandoned instead, which costs the same and teaches nobody anything.

Priorities, outcomes, layers and counts are as recorded in an internal digital strategy and roadmap proposal presented to a governance forum at a global pharmaceutical company. That document states intent on a date and reports no results, so nothing in it is a delivered outcome. Reading its asymmetry as deliberate design rather than an unfinished slide is ours.