Buying agentic services: outcomes, ownership and escape routes

Three ways to procure capability

Agentic services range from packaged software with an AI operator to consultancies building bespoke systems. A managed outcome service handles a defined process—such as invoice intake or support triage—and charges by user, transaction or result. A custom implementation partner designs an agent around your systems. Staff augmentation supplies specialists while your organisation retains architecture and delivery ownership.

Managed products are fastest when the process is common and integrations already exist. Bespoke work suits differentiated workflows, unusual controls or legacy systems. Staff augmentation is appropriate when product ownership is strong but specialist skills are temporarily missing. Avoid commissioning custom AI merely to reproduce a mature software category; configuration is usually cheaper than invention.

Write the problem before the request for proposal

A useful brief describes the current workflow, transaction volume, systems, data classes, exceptions and measurable baseline. State the target in operational language: reduce median handling time from 18 minutes to 10 while maintaining 98% policy compliance. Do not prescribe a multi-agent architecture unless it is genuinely required. Ask bidders to explain the simplest design that can achieve the outcome.

Provide de-identified examples and require a paid discovery or proof of value using representative cases. A successful demonstration should show failure handling, audit traces and integration constraints—not only ideal outputs. Agree acceptance tests before implementation. These might include task success, false-positive rate, latency, cost per completed case and the percentage requiring manual intervention.

Understand the full price

Budgets generally include discovery, integration, model usage, platform licences, security review, testing, training, support and ongoing optimisation. Published software prices rarely represent the complete operating cost. Usage-based contracts should define the billable event and treatment of retries, failed jobs and provider outages. Ask for cost alerts, hard limits and monthly unit-economics reporting.

Fixed-price discovery gives both parties a contained way to reduce uncertainty. Production work can then use milestone pricing tied to tested capabilities. Outcome pricing may align incentives, but only if the outcome is objectively measurable and the supplier does not control which cases it accepts. Retain a portion until documentation, handover and operational acceptance are complete.

Contract terms that matter

Clarify ownership of prompts, workflow code, evaluation sets, connectors, generated data and fine-tuned artefacts. Your business should retain access to its data and run history in a usable format. Require disclosure of model providers and subprocessors, data locations, retention, training use, security controls and incident-notification periods. If personal data is processed, put an appropriate data-processing agreement in place.

Service levels should cover availability, support response, recovery objectives and degraded third-party models. Define who approves model upgrades and who pays for remediation if an upgrade reduces quality. Include an exit plan: export formats, deletion certification, transition assistance and a reasonable period to migrate. Model portability is valuable, but switching providers may still require prompt and evaluation changes.

Governance cannot be outsourced

A vendor may operate the technology, but your organisation remains responsible for customer impact, employment decisions, regulatory duties and brand risk. Name an internal product owner, process owner, security contact and executive sponsor. Establish thresholds for human review and specify prohibited decisions. Review performance by customer segment to identify uneven outcomes hidden by an overall average.

Request traces and dashboards that your team can inspect. Vendors should explain how they test prompt injection, data leakage, hallucinations, tool misuse and access-control failures. Independent penetration testing may be appropriate for high-risk deployments. Ensure support staff cannot casually view sensitive prompts or outputs.

From pilot to durable operation

Run a time-boxed pilot on one process with sufficient volume to expose exceptions. Compare against the baseline, including review time and correction effort. If the pilot succeeds, expand in controlled stages rather than adding departments and capabilities simultaneously. Train employees on when to trust, verify and override the system. Update process documentation so the AI is part of normal operations rather than an unofficial side channel.

Track benefit after launch. Useful measures include throughput, cycle time, first-time resolution, rework, customer satisfaction and cost per successful transaction. Savings are not realised merely because a task becomes faster; capacity must be redeployed or demand absorbed. The strongest service partners make these assumptions explicit and help the client build enough internal understanding to challenge them.