# Before you hire an AI team, ask for these five things

Canonical page: https://aiteamproof.com/note/before-you-hire-an-ai-implementation-partner/

AI Team Proof editorial — 2026-10-06

A practical first-call checklist for business owners: define the workflow, inspect the evidence, and agree on what happens after the demo.

## Start with one workflow, not a transformation promise.

An impressive demo is easy to understand. A proposal for work inside your business is harder: it must account for the people, systems and exceptions that the demo leaves out. Before you ask a provider for a price, write down one workflow you want to improve. Identify its input, the person responsible today and the output your business needs.

Our suggested first conversation is an evidence request. Ask the provider to bring the five items below. A smaller team with clear answers can be easier to evaluate than a larger company with a broad capability deck. These are editorial procurement questions, not a certification test.


## 1. A map of the work and its exceptions

For an invoice workflow, the happy path might be extracting fields and sending them to your accounting system. The real questions concern duplicate invoices, missing purchase orders, unfamiliar suppliers and disputed amounts. Ask where a person reviews the result and which actions the system is allowed to take without approval.

Anthropic’s guidance distinguishes predefined workflows from systems that choose their next steps dynamically, and recommends matching complexity to the problem. Use that distinction in your buying conversation: ask why your workflow needs an agent at all, and what a simpler approach would leave unsolved.

Source: [Anthropic: Building effective agents](https://www.anthropic.com/engineering/building-effective-agents)

## 2. A relevant delivery example you can examine

A logo tells you less than a description of the actual work. Ask what the provider built, what existing systems it connected to, whether the deployment is still operating, and what role its team played. If the client cannot be named, ask what can be shared without exposing confidential information.

Treat a provider’s own case study as provider-reported evidence. It can still be useful. Keep separate the facts that a customer or other source has confirmed. When someone quotes a percentage improvement, ask for the measurement period, starting point, sample and exclusions. A result from a different workflow is not a forecast for yours.


## 3. An acceptance test based on your work

Before a build starts, agree on examples of acceptable and unacceptable outcomes. Include ordinary cases and the exceptions that cost your staff the most time. Decide who grades the result, how many mistakes are tolerable and which mistakes require stopping the system.

Anthropic’s evaluation guidance separates an agent’s stated success from the actual resulting state. For a buyer, that means checking whether the invoice was stored correctly or the ticket reached the right queue, rather than whether the assistant said it completed the task. Ask to retain the test cases so later changes can be checked too.

Source: [Anthropic: Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents)

## 4. A handover and operating plan

Ask who will have access to the code, prompts, integrations and monitoring. Name the person who can revoke access, pause the workflow or switch back to a manual process. Clarify who investigates failures, what support includes and what becomes your responsibility when the engagement ends.

If the provider runs the service for you, ownership may look different from a custom implementation. The useful question is whether the arrangement fits your operating capacity. A business without an internal engineering team should understand the support it is actually buying.


## 5. A price with a boundary

Request separate figures or assumptions for discovery, implementation, ongoing support and usage. Ask what happens when volume grows, a model changes, or an integration breaks. A fixed project price is only comparable when the deliverables and exclusions are equally clear.

Finish the call with a short written record: the workflow, evidence received, unanswered questions, acceptance criteria and next decision. If a critical question remains open, make it a condition of the next step. You do not need every answer immediately; you need to know which commitments have actually been made.


## Sources and editorial approach

This is AI-assisted editorial guidance reviewed for this directory. It does not describe a project we delivered or endorse a particular provider.

- [Anthropic: Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) — checked 2026-10-06
- [Anthropic: Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) — checked 2026-10-06