AI governance
By 2028, a third of enterprise applications are forecast to act on their own. Write that into next year's contracts now
Gartner's own prediction is that a third of enterprise applications will include agentic AI capable of semi-autonomous action by 2028. That is a forecast, not an observed result — and the honest way to use it is to plan the procurement language for machinery you haven't bought yet.
Executive summary
33%
of enterprise applications forecast to include agentic AI by 2028, per Gartner (Oct 2024)
<1%
had this capability at the time Gartner published the forecast
Core conclusions
- Most AI governance conversations are retrospective; the more useful one for capex planning is forward — what decision rights are you prepared to grant equipment you haven't bought yet.
- Treat the 33 percent figure as a planning horizon, not evidence anything has already changed on your floor.
- Write a named action boundary, a role-based escalation path, a logging requirement, and a stated fallback state into the next equipment contract cycle.
Most governance conversations about AI in industrial settings are retrospective: a system is already deployed, and the question is how to oversee what it is already doing. The more useful conversation for an operator planning capex over the next few equipment cycles is the forward one — what decision rights are you prepared to grant a system you haven't bought yet, and when.
Gartner's often-cited prediction, published in October 2024, is that by 2028 a third of enterprise applications will include agentic AI capable of taking semi-autonomous action, up from less than 1 percent at the time of the forecast. Treat that figure as exactly what it is: an analyst's projection about software in general, not a claim about what any specific machine on your floor will be authorised to do. The use of it here is as a planning horizon, not as evidence anything has already changed.
What to write into a 2027-cycle equipment contract now
- A named boundary for autonomous action: what the equipment may decide and execute without a human step, stated as a specific list, not a general capability claim.
- An escalation path with a person named by role, not a queue — the same requirement this shelf has argued for agentic software applies directly to agentic control of physical equipment.
- A logging requirement sufficient to reconstruct any autonomous action taken, retained for a period set by consequence rather than by vendor default.
- A stated fallback: what the machine does, specifically, in the interval between an anomaly and a human response — not merely that it 'stops safely', but what state it stops in and what has to happen before it resumes.
Why this belongs in procurement, not just governance
A capability that is not in the contract does not become negotiable later on favourable terms — it becomes a change order, priced by the vendor who already has the equipment on your floor. The 2028 forecast is a reason to write the boundary now, into the next contract cycle, while it is still a specification decision rather than a retrofit.
None of this requires believing the 33 percent figure precisely. It requires accepting the direction it describes and pricing the governance work into the equipment budget before the equipment arrives already deciding things on its own.