Six hypothetical AI agent scopes show what to include in a B2B budget: connections, review, testing, running costs and acceptance criteria.
The six scenarios below are hypothetical planning examples, not delivered Afiniti Global projects, client results or approved offer prices. They show how scope changes when permissions, review and exception handling are made explicit.
There is no universal minimum price for a production agent. Start by describing one task, its users and the actions it may take. Compare written scopes on the same basis, including what is excluded and what happens if a required connection cannot be supported.
What changes the budget
Separate five questions. What may happen without approval? Which systems must be read or changed? How will correctness be checked? What data restrictions apply? Who handles failures and future changes? Ask suppliers to show the assumptions behind each answer. Framework names or a polished demo do not establish the effort required for your workflow.
Scenario 1: Inbound lead qualification
A hypothetical first scope reads one enquiry source, checks required fields and existing CRM records, then prepares a qualification recommendation for a salesperson. Include approved enrichment access, duplicate checks, routing rules and review of uncertain matches. Acceptance tests should include an existing customer, an incomplete form and a disputed score. Exclude autonomous outreach or rejection unless separately agreed and tested.
Scenario 2: Support reply preparation
A hypothetical support assistant drafts replies from approved help content and passes account-specific issues to an authorised colleague. Budget for content preparation, permission checks, source references, escalation and evaluation against real question types. Measure accepted drafts, edits and wrong-answer incidents. Do not assume a ticket-deflection rate or remove staffing costs before a controlled trial supports that decision.
Scenario 3: Compliance review preparation
A hypothetical document review workflow collects permitted material, highlights possible exceptions and prepares evidence for an analyst. It does not certify compliance or make a regulatory decision. Scope qualified review, data restrictions, retention, access controls and the record of who approved each finding. Confirm the applicable obligations with the organisation’s responsible advisers before estimating specialist work.
Scenario 4: Sales-call coaching drafts
A hypothetical coaching tool processes authorised recordings and prepares notes against an agreed rubric. Check recording permissions, transcription quality, retention and who can see the results. Include test calls with poor audio and ambiguous language. A manager reviews the notes; the system should not make employment or performance decisions from an unverified score.
Scenario 5: Customer onboarding coordination
A hypothetical onboarding workflow tracks documents, review decisions and account-setup requests. First check whether ordinary rules can handle the sequence. Budget separately for each supported connection, manual approval, partial failure and duplicate submission. Identity checks and account changes require explicit scope and responsibility; several named agents do not by themselves make the process reliable.
Scenario 6: Internal research assistance
A hypothetical research assistant searches permitted filings, internal notes and licensed sources, then drafts a briefing with references for review. Include source access, freshness, permissions, missing information and unsupported statements in the tests. References make checking easier but do not guarantee correctness. Measure reviewer effort and factual errors before projecting any increase in throughput.
Review current supplier costs
Obtain current rates for models, storage, hosting and software accounts. Estimate usage from representative tasks, including retries and testing. Avoid carrying old token prices or subscription allowances into a new proposal without checking them.
Reusable tools may reduce some implementation work, but they do not remove the need to confirm the task, permissions and expected behaviour. Ask which components are reused, which need custom work and which costs remain dependent on your data or existing systems.
Estimate payback with explicit assumptions
Use measured net benefit rather than an assumed share of salaries. In a hypothetical budget, a $24,000 build with $3,000 of monthly cash benefit and $1,000 of monthly operating costs has $2,000 net monthly benefit: simple payback is 12 months after that benefit begins. Include rollout delays and internal setup effort. If net benefit is zero or negative, this calculation does not produce payback.
Compare the complete scope
Request a proposal covering deliverables, access requirements, acceptance tests, handover, operating costs and exclusions. Record low, expected and high usage cases, and what would trigger a revised quote. These scenarios are prompts for scoping; none establishes a market rate, a delivery date or a promised business result.