Stop Buying Agents: The RFP Is the Pilot That Never Ends

A request for proposal is how a financial institution buys technology: specify it, tender it, contract it. Agentic AI does not hold still long enough to be bought that way.

The upside is well documented. Inverto, a BCG company, puts the potential at 30 to 40% cost reduction by 2030 for banks that move from task-level tools to agent-run workflows, and up to 30% profitability improvement. Yet in the initiatives it reviewed, none had delegated any decision-making autonomy to AI. The pilots pass. Production does not follow.

Inverto's diagnosis is structural: 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 standard contracts never contemplated. We see the same pattern. Where we part company is the fix. Inverto's 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 it is built inside your stack, by engineers you can name, into code you own.

You can download the full field note as a PDF, or read the argument below.

Pilots pass. Production does not follow.

The obstacle is not the model. It is the shape of the purchase. An RFP takes months to write and answer; agentic systems change between weekly releases, so the requirement is stale before the contract is signed. Budgets run out before scale-up. The result is controlled experimentation that never industrialises, which is the same plateau we describe from the workflow side in who owns the wait?.

We cover the three procurement gaps, cadence, accountability and one-size sourcing, and why they are symptoms of one decision (buying agentic AI as if it were software) in Why agentic AI pilots never industrialise.

Accountability falls between vendors because there are vendors

"Who is responsible when an agent's autonomous output causes harm?" In a stack that spans a model provider, an orchestration layer and a monitoring tool, the honest answer is nobody in particular. Underneath sits a data problem: a Stanford CodeX analysis found 92% of AI vendors claim data-usage rights beyond what the service needs, against a 63% norm for standard SaaS, and most defaults permit training on client operational data unless you exclude it.

Inverto lists five components of an AI-ready contract: responsibility mapping, audit and reporting rights, data usage and control, continuity and exit, and conduct safeguards. Each is a clause that tries to obtain from a vendor something a build gives you by construction. A clause is a request; a property of the system is a fact.

More in Accountability, contracts and data: what you negotiate vs what a build gives you.

The higher the consequence, the less you can buy off the shelf

Inverto's governing principle is that the closer an agent sits to a client or a regulated decision, the tighter and more bespoke the sourcing must become. Read the recommendation for the two highest-exposure tiers, front office and middle office: co-build and strategic partnership, not a product purchase. That is an embedded applied-AI engagement described from the buyer's side of the table. Standardised sourcing is the right call for the back office, and it is also where the least value sits.

Their fourth bottleneck is incentives: fixed scopes discourage iteration, IP restrictions block co-development, and there is no post-launch reward, so systems stagnate before they are worth what they cost. The vendor's reward ends at launch. Yours does not.

We work through the risk tiers, the phased gates and the incentive problem in Risk tiers and incentives: why high-consequence AI is co-built, not bought.

Funding gated to outcomes is good procurement. Autonomy gated to outcomes is good engineering.

Inverto's three procurement phases (governed proof of concept, minimum viable product, scale development) are the same shape as our three engagement stages (assess, build, prove and scale). The difference is what each gate produces. Their gates release funding to a vendor on validated performance. Ours release autonomy to a workflow on validated performance, and leave the validation record behind as a working part of the system. Both match what supervisors expect. Only one leaves you holding the code.

This is the same argument we make in our operating model for AI where the output has consequences, and why the second workflow costs less than the first is the compounding logic of the company brain.

The way through

Start narrow and high-frequency. Keep autonomy constrained until governance is in place, and put that governance in the substrate so widening autonomy is a threshold change with a record, not a committee. Treat data segregation as foundational, because the build runs in your perimeter. And involve procurement early: bring the CPO into the first fortnight, when the decision is build, buy or automate, not after.

Is your agentic AI programme stuck in procurement? Bring us the workflow the RFP has stalled on. We will map it in a fortnight, tell you honestly whether to build, buy or automate, and if it is a build, deliver it inside your stack with the evidence attached. Download the field note, or send us the workflow.

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