Why Most AI Projects Don't Even Get to the Pilot Stage

Harvard Business Review's February 2026 synthesis of industry surveys finds that only 24-27% of organizations have adequate AI-ready talent, IT systems, or regulatory preparedness. For the smallest firms, the number drops below one in five. The pilot-to-production gap that most coverage focuses on, the 80% of pilots that never reach production, is the second failure point. The first one happens earlier, when the project that should have become a pilot never gets scoped because the readiness for it is not there.

What HBR actually measured

HBR's February 2026 article pulls from AICPA/CIMA's 2026 benchmark and adjacent enterprise surveys covering the talent, infrastructure, and regulatory readiness axes. The 24-27% range is consistent across the three axes, which is the load-bearing finding. It would be one thing if firms were strong in two and weak in one. They are weak in all three at roughly the same rate, which means the underlying constraint is operational maturity, not any single missing piece.

For sub-5-employee firms, the picture sharpens. Fewer than 1 in 5 have what they need across the readiness axes. 82% cite the belief that "AI isn't applicable to my business" as their reason for non-adoption. 29% of SMEs name training gaps as their biggest obstacle, yet only 12% are investing in any AI-related training. The asymmetry between the named obstacle and the investment to address it is the readiness gap in compressed form.

The "AI isn't applicable to my business" finding is the one worth holding closest. Most of the firms saying this are firms whose work has heavy admin overhead, recurring patterned tasks, and a founder operating well past their capacity ceiling. AI is not just applicable to those businesses, it is the most direct route to recovering the founder's time. The gap between the actual fit and the perceived fit is itself a readiness problem, because the operator cannot scope a project they cannot see.

Why the readiness gap is structurally upstream of the P2P gap

Most of the AI failure discourse, including Radiant Work's own thesis on the pilot-to-production gap, focuses on the 80% of pilots that never reach production. The HBR data adds the layer underneath: even the pilots that fail are the ones that got scoped. The majority of organizations never reach the pilot stage at all because the readiness preconditions are not in place.

This matters because the remedy for the two gaps is different. A failed pilot can usually be saved by better context engineering, better integration into the source of truth, or a tighter scope on what the agent actually owns. A failed readiness assessment is upstream of all of that. No amount of model selection or prompt design fixes a firm whose source of truth lives in three Slack channels and a notebook, whose policy is verbal, and whose team has never written a workflow specification.

The order of operations is the whole point. Address the readiness gap first, then the pilot can be scoped against real ground. The firms that try to skip the readiness work and go straight to a pilot tend to be the ones who end up in the 80% later, having spent budget on a tool that the operational substrate could not support.

The three readiness axes, made operational

HBR's three axes generalize well, but they have to be made concrete to be useful. Here is what each axis looks like at the scale of a 5 to 50 person creative firm.

Talent readiness is not whether the team can write Python. It is whether someone in the firm can specify a workflow precisely enough that an agent can run it. That means naming the trigger, the inputs, the steps, the decision points where a human stays in the loop, and the success criteria. Most firms have nobody who has ever written that down for any workflow, AI or otherwise. The first sign of talent readiness is the existence of a single workflow specification, written out, in plain language. Most firms cannot produce one. Building the first one is usually the highest-leverage 90 minutes a small firm can spend on AI.

Systems readiness is whether a firm's data lives somewhere an agent can read it. Project records in one tool, client communications in another, contracts in a third, and the source of truth in the founder's head is the modal small-firm setup. No agent can work against that substrate, regardless of capability. Systems readiness usually means consolidating to a single source of truth, with a clean schema, and ensuring the tools that produce content write back into it. The work is unglamorous and structural; it is also the precondition for everything downstream.

Regulatory and policy readiness scales to the firm. A 10-person studio does not need a SOC 2 program. It needs a one-page AI policy that names approved tools, prohibited uses, data handling rules, and an exception path. Cloud Security Alliance research suggests a written policy alone reduces shadow AI usage by roughly two thirds. The policy is the artifact that moves AI use from invisible to named. Without it, the firm is governing an exposure surface it cannot see.

Radiant Work's audit-first methodology maps a firm against these three axes before any agent is built. The Radiant Work FAQ covers how the readiness work sits inside the Operations Audit and where it leads into an Architecture Sprint when systems readiness is the binding constraint.

What lightweight readiness looks like in practice

The remedy for sub-5-employee firms is not enterprise tooling. It is four lightweight artifacts and a recurring review cadence.

A single workflow specification, written out for the workflow most likely to be the first agent. Trigger, inputs, steps, decision points, success criteria. One page. The act of writing it surfaces every assumption that has been living in the founder's head.

A single source of truth, even if the source is a Notion database and not a CRM. The criterion is not the tool. The criterion is that one place holds the canonical version of the data the agent needs, and the other tools either feed it or read from it.

A one-page AI policy. Approved tools, prohibited uses, data handling rules, an exception path. Two-thirds of shadow AI usage disappears when the policy exists in writing.

A monthly review. Thirty minutes, with the specification, the source of truth schema, and the policy open. New workflows named. Drift flagged. Investment in training updated.

These four artifacts move a firm from the bottom 20% of readiness to the top 25-30% in roughly two to four weeks of focused work. That is the leverage point. The firm is not yet running agents in production, but it has the substrate that makes the first agent ship-able and scope-able.

What the 24-27% number actually predicts

If HBR's data is even directionally right, three quarters of organizations are spending budget on AI tools they cannot operationally support. The line item shows up; the operational gain does not. The firms that get to the gain are not the ones with the bigger AI budget. They are the ones with the readiness work done.

The forecasting implication is straightforward. Over the next 18 months, the gap between the firms with readiness and the firms without will widen, not narrow. The firms with readiness will deploy a second agent, then a third, then a sequenced agent layer that compounds. The firms without will buy more tools, run more failed pilots, and conclude that AI does not work for their business. The conclusion will be wrong, but the evidence supporting it will look real.

This is why the order matters. Audit before automate. Make the firm legible to an agent before you bring an agent into the firm.

What to do next

HBR's data is a structural argument for sequencing the readiness work first. The firms that skip it and go straight to a pilot become the 80% later. The firms that do it first become the 24-27% who can actually deploy AI usefully.

If you want to know where your firm sits across the three readiness axes, schedule a conversation. The audit will tell you which axis is the binding constraint, what two to four weeks of focused work would close it, and which workflow should be the first agent once the substrate is in place.