Procurement in the AI era: Adopt fast, pivot cheap

Procurement in the AI era: Adopt fast, pivot cheap

In my advisory work with CxOs, I’ve lost count of transformations that begin with immaculate intent and end with merely immaculate paperwork. Somewhere between a glossy RFP and the first meaningful release, momentum ebbs. The pattern is familiar: big bang scope setting, heroic milestones, one supplier to “own the outcome”, and a contract that treats change as a nuisance rather than a certainty. The question I now pose at the outset is disarmingly simple: if we’re wrong, how cheaply can we change our minds?

That question reframes procurement from a race to the lowest rate card into a race-to-value and cost-to-pivot. In an age where AI, data platforms and cloud architectures evolve monthly - leaders will not be judged on who squeezed the last 2% from the day rate. They’ll be judged on whether they bought learning and optionality, without the typical supplier handcuffs.

Where programmes falter (and what we recommend instead)

I have observed a handful of patterns recur across sectors; these are the ones I now address head on:

• Precision masquerading as accuracy: Thick specifications create comfort but not clarity. They age fast. We consistently recommend anchoring the first 8–12 weeks around what will be learnt, not a list of deliverables. The horizon is short on purpose.
• Paying for activity, not acceleration: Supplier teams hit documentation milestones while nothing changes for the customer. We put time to first value on the scorecard and track learning velocity, with assumptions tested, iterated or killed per sprint.
• Binary assurance: Heavy pre award checks followed by cruise control. In high change domains we recommend progressive assurance i.e. lighter touch at pilot, heavier when at scale, evidence observable at each gate.
• Monolithic contracts for modular systems: Technology architecture is fashionably composable; however, many technology contracts aren’t. We help clients make delivery interfaces contractual artefacts with acceptance tests and payments tied to conformance. Lock in shrinks when interfaces are real.
• Incentives that assume certainty: When outcomes are uncertain, micro-managed deliverables don’t save you. Decision rights at the point of action and gain share on realised impact work much better.

Reflections from the field

While advising on the set up of an automation factory for a PE backed SaaS client, the default would have been appointing one technology partner, defining a 2–3-year scope, and hoping for the best. We recommended a different line. The client ran two parallel discovery tracks for 8 weeks with a hard cap and a common interface contract into the client’s case management system. The rules were explicit:

• Gate 1 judged time-to-first-value in production, not demo show
• Data and model portability were non negotiable; exit was pre priced
• Option pricing allowed rapid scale on what worked - and graceful stop on whatever didn’t

We scaled one vendor and off-ramped the other with no drama, because the contract expected that outcome right from the start. Six months later, value was live in two markets, and the board conversation had shifted from “which supplier?” to “which use-case next?” - a healthier question in my opinion.

So, here’s a practical playbook that should serve as a guide rather than a prescription.

Contract for learning and options

• Write short horizon SOWs (8–12 weeks) that articulate hypotheses and interfaces; treat specifications as promises about how parts will talk, not predictions about everything that will happen
• Attach option prices for scale-up paths so you can move fast without renegotiating from scratch
• Preserve competitive tension where uncertainty is highest - parallel discovery runs are a tool, not a doctrine

Measure acceleration, not paperwork

• Put TTV (time to first value) and learning velocity on the supplier scorecard. Examples: time to first deploy to a safe prod slice; time to first measurable saving; hypotheses tested and killed per sprint
• Tie a small accelerator to early value (or a shared savings component once outcomes are measured) to align behaviours

Make assurance progressive and visible

• Stage controls: Pilot → Limited Production → Scale. Evidence expectations grow with blast radius
• For AI work, ask suppliers to map controls to recognised frames and show evidence. Keep it lean early, complete later

Engineer for reversibility

• Treat interfaces/APIs as assets. Pay on passing interface tests. Publish them internally so another team, or supplier, can compete at the boundary
• Mandate data portability and an SBOM where software is delivered. Price exit and transition assistance up front
• Keep IP clean: client owns domain assets and data derivatives; supplier keeps generic tooling with fair licensing back where needed

Put decision rights where action lives

• In uncertain work, tighten governance and loosen micromanagement. Give go/stop/reshape decisions to the product trio (business, tech, commercial) at gates; escalate only on money or principle

Caveats and constraints

A balanced word of caution: this approach is not a silver bullet, and it will not fit every context out of the box.

• Regulatory and assurance intensity: In highly regulated domains, pilots may still demand heavyweight controls. Progressive assurance must respect statutory obligations; shorten cycles where safe, but don’t dilute duty of care.
• Supplier market depth: Parallel discovery presumes credible alternatives. In thin markets or where a platform ecosystem is effectively a single route, competitive tension will be limited. Over modularising in such contexts can create friction without choice.
• Operating model maturity: Short horizons and options create more decision points. If product governance is weak, this adds noise rather than speed. You may need to build decision cadence before you scale the model.
• Interface tax: Treating interfaces as first class artefacts is powerful, but it does add upfront work (contracting, testing, documentation). Net gains arrive over multiple cycles, not day one.
• Accountability drift: Splitting scopes across modules can blur ownership if you don’t define clear outcomes and decision rights. Someone must still be answerable for the integrated result.
• Commercial overhead: Option pricing and stage gates require procurement, legal and finance capacity. If your team is bandwidth constrained, start with one or two critical buys rather than broad adoption.
• Cultural readiness: Some executive teams prefer long, predictable plans over adaptive bets, even when those plans are fiction. You may need to prove the model with a contained success before it generalises.

If you can’t go all in yet, try this:

• Apply the model to a single, well bounded initiative with a visible sponsor and contained risk
• Use soft competition (e.g., a credible challenger held in reserve) if parallel discovery is impractical
• Keep the interface/portability clauses even in single supplier deals; they are your future optionality
• Start with observability of time to first value before introducing gain share and option menus

Closing thoughts

Speed without structure is almost gambling. Structure without speed is certainly bureaucracy. The job is to design commercial systems where being wrong is affordable and being right scales quickly. When we do that, procurement stops being the brake and becomes the flywheel.

‍