AI Vendor Management for Executives: A Whitepaper

AI vendors behave differently from conventional software vendors. Models are deprecated on short notice, pricing changes materially, capability shifts between providers within quarters, startups disappear, and AI features appear inside products you already run without going through procurement. This whitepaper sets out how to manage the category: selection, contracting, concentration, ongoing management, and exit. It is general guidance, not legal advice.

What are the vendor categories?

Category	What they supply	Main risks	Management emphasis
Model providers	Inference via API or hosted models	Deprecation, pricing, terms, availability	Replaceability, contract terms, alternatives
Platforms	Gateways, agent frameworks, evaluation, observability	Lock-in, portability, viability	Standards support, export, exit terms
Applications with AI	Products whose value is an AI capability	Thin wrappers, model dependency, accuracy claims	Evidence, exception burden, durability
Embedded AI	AI features inside products you already buy	Unmanaged data flows, undisclosed changes	Inventory, review at renewal, notice terms

Most governance attention goes to the first category and most unmanaged exposure sits in the fourth. The questions your CISO will ask about AI guide covers the security dimension of each.

How should vendors be selected?

On evaluation against your own cases, not on demonstrations or published benchmarks. The process that works:

1. Define the task and the threshold before looking at vendors, so the bar is not set by what the first demo achieves.
2. Assemble fifty to a hundred real cases with known correct outcomes, including hard and awkward ones.
3. Run each candidate on the same set, measuring pass rate, cost per task, and latency.
4. Test the exception path: what happens to the cases the system cannot handle, and what does that cost in human time?
5. Assess the surrounding product: integrations, permissions, evaluation, monitoring, and what happens when the underlying model changes.
6. Check durability: what would a competitor with the same model access still lack?

A vendor who declines to be evaluated on your cases has answered the most important question. The red flags in an AI vendor pitch guide covers what else to watch in the process, and the how executives should evaluate an AI demo guide covers converting a pitch into a test.

Which contract terms matter most?

Six, and they are worth negotiating specifically rather than accepting standard terms:

Data use and training. Explicit prohibition on training on your data, with retention periods, deletion commitments, and subprocessor transparency. This is the term customers will ask you about, so it flows through to your own commitments.

Model change notice. A minimum notice period for deprecation or material behavior change, and the ability to pin a version for a defined period. Without this, a provider update can invalidate a validated system with no warning. The model deprecation risk management guide covers the operational side.

Portability and export. Prompts, configurations, connector definitions, evaluation assets, and logs, in documented formats, available on demand rather than only at termination.

Liability and indemnity. Proportionate to the risk, including intellectual property indemnity for output where relevant.

Security and breach notice. Defined controls, certifications, and notice periods that allow you to meet your own obligations.

Exit assistance. A contractual obligation with defined scope and duration, not a goodwill expectation. The how to sunset an AI vendor guide covers using it.

Consult counsel on drafting; the point here is that these are the terms where the difference between a good and a poor position is largest.

How should concentration be managed?

Deliberately, with limits set in the risk appetite rather than accumulated by default. Concentration in AI arises faster than in conventional software because a single provider can end up serving every agent through a shared gateway, and because capability leadership concentrates attention on one or two names.

The measures worth tracking: share of agents dependent on each provider; share of AI spend by provider; number of agents with a validated alternative; and time to switch for the most critical agents. The chief risk officer's guide to AI agents covers aggregation across the estate.

Mitigations are architectural more than contractual: a gateway that abstracts providers, evaluation sets that make alternatives testable, and connectors built to open standards so the integration layer is portable. The AI gateways explained for executives and agent interoperability explained for executives pieces cover both.

What about embedded AI?

The exposure most often missed. A CRM adds an assistant; a helpdesk adds automated responses; an HR system adds screening support. None goes through AI procurement, all process company data, and many are enabled by default.

The management approach:

• Add AI questions to vendor reviews and renewals: what AI features exist, what data they process, whether they are on by default, whether they can be disabled, what the training and retention terms are, and how model changes are notified.
• Inventory embedded AI alongside built agents, with the same risk tiering.
• Require notice of new AI features and material model changes in contracts.
• Assess at renewal rather than only at first purchase, because the product you bought may now include capabilities you did not assess.

For regulated firms this is a compliance matter as much as a security one, since the embedded feature may be making or influencing decisions within a regulated process. The executive playbook for AI in regulated industries covers that dimension.

How should vendors be managed after selection?

With a cadence proportionate to their criticality:

Activity	Frequency	Purpose
Performance against evaluation set	Monthly or continuous	Detect quality change, including provider-side
Cost per task review	Monthly	Catch pricing or usage drift
Model version and change log review	On notification and quarterly	Maintain validation currency
Alternative validation	Quarterly for critical agents	Keep the switch option real
Contract and terms review	Annually and on change	Terms change; so do your obligations
Concentration review	Quarterly	Aggregate exposure
Vendor viability assessment	Annually, more often for startups	Anticipate failure

The first line is the one companies most often lack: without scheduled evaluation, a provider-side quality change is invisible until it reaches customers. The how to respond to an AI vendor failure guide covers the response when something does change.

How should early-stage vendors be handled?

Differently, not avoided. AI capability frequently sits with young companies, and refusing to buy from them cedes capability. The mitigations: keep the integration portable, hold your own copies of prompts, configurations, and evaluation assets, structure the commercial relationship so a failure is survivable, and assess viability annually with a realistic view. Where the capability is genuinely differentiated and the vendor is fragile, the honest options are to accept the risk with mitigations, to negotiate source or asset escrow where feasible, or to build the capability internally.

What should the build-versus-buy decision consider?

Vendor management and build-versus-buy are the same decision viewed from different sides. The layers worth buying are the ones that are standardizing and that no customer differentiates on: model inference, gateway and observability tooling, evaluation infrastructure, and general-purpose applications. The layers worth building are the ones specific to the company: the agents themselves, the specifications that encode how the business works, the evaluation sets built from its own cases, and the connectors into systems nobody else runs.

Two failure patterns follow from getting this backwards. Building generic infrastructure produces a maintenance burden on something vendors will commoditize within a couple of years. Buying a product that encodes your process knowledge puts your most durable asset inside someone else's platform, where it is neither portable nor yours. The agentic AI value chain piece covers the layer-by-layer logic, and the practical test for any purchase is what the company still owns if the relationship ends.

How does vendor management change as the estate grows?

At one or two agents, vendor management is a procurement activity. At twenty, it becomes an operating discipline: a provider's pricing change affects a dozen budgets, a deprecation triggers a coordinated revalidation programme, and an outage is a business continuity event rather than an inconvenience. The transition point is usually where the same provider becomes load-bearing for several business units, and it tends to arrive before anyone has assigned ownership for managing it.

Three practices mark the shift: a named owner for the vendor relationship rather than distributed account management; a consolidated view of spend and dependency rather than per-project budgets; and scheduled evaluation that covers provider-side change rather than only the company's own releases. Companies that make this transition deliberately negotiate better and absorb provider changes without drama; those that do not discover the dependency during the first material price change.

What does good governance of the category look like?

An inventory covering all four categories with owners and risk tiers; selection by evaluation with documented results; standard contract requirements applied consistently; concentration limits in the risk appetite with measured position; a management cadence proportionate to criticality; and exit readiness verified rather than assumed. Reported to the executive team quarterly and to the board where concentration or spend is material.

What should executives ask?

• Which providers are load-bearing, and for how many agents?
• Was the last vendor selected on evaluation against our cases, or on a demo?
• What do our contracts say about training, notice, portability, and exit?
• Which products we already buy have added AI features since we assessed them?
• Could we switch our main provider in a quarter, and has that been tested?
• Who owns vendor management for this category, and when did they last review it?

How can FISTA Solutions help?

FISTA Solutions runs evaluations of vendor systems on clients' own cases, advises on the contract terms that determine position, builds the gateway and portable connector layer that make providers replaceable, and delivers the alternative as production AI agents where building is the better answer. Its AI enablement practice helps executive teams set concentration limits, inventory embedded AI, and establish the management cadence. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries, with a 99.9% uptime record on production systems.

To assess your vendor position before a provider forces the question, talk to FISTA on WhatsApp, or read LLM vendor lock-in for the model-provider dimension.