Why Agentic AI Pilots Never Industrialise

Agentic AI pilots never industrialise because financial institutions try to buy them the way they buy software, and the two do not fit. The upside is documented and large, the technology works well enough to pass a pilot, and yet almost nothing reaches production with real autonomy. The obstacle is not the model. It is the shape of the purchase.

That gap is worth stating precisely. Inverto, a BCG company, in "Scaling Agentic AI in Financial Services: Procurement as a Strategic Enabler" (2026), puts the prize at 30 to 40% cost-reduction potential by 2030 for banks that move from task-level tools to agent-run workflows, and up to 30% profitability improvement over the same horizon. Against that, its own finding: none of the initiatives it reviewed had delegated any decision-making autonomy to AI. The demonstrations run. The delegation does not.

This is the argument we set out in full in Stop buying agents. Here we take the first part of it: why the pilot passes and production never follows.

The upside is documented. The autonomy is not being delegated.

A pilot is designed to prove feasibility under controlled conditions. It runs on a bounded dataset, with a human checking every output, in a sandbox that nobody depends on. Passing it tells you the model can do the task. It does not tell you the institution can let the model do the task unsupervised, at volume, in a regulated process where a wrong output has consequences.

Inverto's diagnosis is structural, and it matches what we see. Procurement and governance frameworks built for stable, deterministic software cannot accommodate systems that change weekly, spread accountability across several vendors, and need a level of traceability that standard contracts never contemplated. The result is controlled experimentation that does not industrialise. It is the same plateau we describe from the workflow side in who owns the wait?.

Where we part company with Inverto is the fix. Its answer is to make procurement good enough to buy agentic AI safely. Ours is that most of what makes it hard to buy disappears when the system is built inside your stack, by engineers you can name, into code you own.

Three procurement gaps, one root cause

Inverto names three procurement gaps. We would call them three symptoms of one decision: treating agentic AI as if it were software. The useful move is to look at each gap twice, once as it appears in a purchase, and once as it appears when the work is built rather than bought.

• Cadence. Procurement assumes stable requirements. An RFP takes months to write and answer, while agentic systems change between weekly releases, so the requirement is stale before the contract is signed and budgets run out before scale-up. When the work is built, the requirement is not stable because the work is not understood yet, and that is treated as the starting condition, not a failure. An embedded team maps the workflow first, prototypes in days, and there is no RFP to outlast.

• Accountability. Contracts assign responsibility to individual vendors, but the workflow spans a model provider, an orchestration layer and a monitoring tool, so failures fall between them. When the work is built, one accountable party builds the workflow end to end, inside your environment, with named roles on your side before go-live. There is no gap between contracts because there is one engagement, not three.

• Sourcing. Every use case runs through the same category, threshold and template, whatever its regulatory exposure. When the work is built, controls are engineered to the consequence of the output. A KYC triage tool and an internal drafting assistant get different checks because they get different code, not different paperwork.

None of this says procurement is wrong to care. It says the function is being asked to paper over a structure that was chosen before it got involved. The gaps are real. They are gaps in a purchase, not in a build. We work through the accountability and contract side of this in Accountability, contracts and data, and the risk-tier and incentive side in Risk tiers and incentives.

What changes when the requirement is allowed to move

The single assumption doing the damage is that you can specify an agentic workflow up front, tender it, and hold the specification still while a vendor delivers against it. You cannot, because the thing you are specifying changes faster than the procurement cycle that governs it.

Building removes the assumption rather than working around it. Mapping the workflow with the people who run it replaces the RFP with something you can inspect in a fortnight. Prototyping on your own documents replaces the promised capability with a demonstrated one. And because delivery happens inside your perimeter under one accountable party, the traceability that standard contracts never contemplated is emitted from the first run instead of negotiated for. The pilot stops being a thing you pass and start over from. It becomes the first stage of the system you keep.

This is why the plateau is not a technology problem waiting on a better model. The models are already good enough to pass the pilots. What is missing is a delivery shape that lets a validated capability turn into a delegated one, with the evidence attached, inside an environment the institution controls.

If your agentic AI programme is stuck at the pilot, the workflow the RFP has stalled on is the place to start. Download the field note for the full argument, or send us the workflow and we will map it, tell you honestly whether to build, buy or automate, and if it is a build, deliver it inside your stack.

Figures and framing are from Inverto (a BCG company), "Scaling Agentic AI in Financial Services: Procurement as a Strategic Enabler", 2026.