Before hiring an AI consultant, ask what work will change, what is included, who owns the result, how it will be tested, and what happens when it fails. Also settle data access, human approval, training, ongoing cost, and exit arrangements. At Applied Margin, those questions belong in the buying decision before the build. A convincing demonstration is useful evidence, but it is not a complete proposal.
Use the twelve questions below with every provider, including us. Ask for specific answers tied to your workflow and written into the appropriate scope or agreement. A good answer names a person, a deliverable, a test, or a limit. If the answer depends on information the provider does not yet have, it should identify how that uncertainty will be resolved before you commit to the affected work.
You're the owner. The office manager needs relief and the demo looks convincing. Test it against an ordinary Tuesday: incomplete inputs, competing priorities, and someone responsible for correcting the result.
What should you be able to compare before signing?
Compare proposals using the same categories. Different solutions are fine; their scope and responsibilities need to be clear.
| Compare | Evidence to request | What a gap means |
|---|---|---|
| Work and scope | Named workflow, deliverables, exclusions, and dependencies | You cannot yet compare the quoted prices |
| Responsibility | Client owner, provider owner, reviewers, and handoff | Important work may fall between the two teams |
| Evidence | Baseline, representative tests, and acceptance decision | The demo may be the only definition of success |
| Full cost | Service fee, software, usage, upkeep, and change terms | The first invoice is not the full commitment |
| Control | Accounts, data access, retained materials, and exit | Changing providers may be harder than expected |
For the difference between provider models, read consultant, agency, freelancer, or managed AI service. Then ask the same practical questions regardless of the label.
Which twelve questions should you ask?
1. What exact job are we improving, and why this one?
A good answer sounds like: “We are reducing the rework in this internal draft, for this team, using these inputs.” The provider should connect the proposal to your operating problem and explain the alternatives. For an illustrative HVAC office, “AI for sales” is too broad; preparing a checked estimate draft from approved job notes is a job you can inspect. Ask what would justify leaving it alone.
2. What is included, excluded, and still uncertain?
A good answer sounds like: “These users, sources, outputs, and review steps are included. This additional integration is not.” Confirm preparation, cleanup, training, testing, and launch support. The proposal should identify any unconfirmed system access and explain how changes affect the fee. A price without those boundaries cannot be compared fairly.
3. Who owns the accounts, work product, and usable handoff?
A good answer sounds like: “Here are the accounts you control, the materials you receive, and the third-party or proprietary components covered by their own terms.” Ask about instructions, configurations, source files where applicable, test cases, and exports. Confirm ownership and usage rights in writing; paying an invoice does not explain what you can retain or reuse.
4. Which number will tell us whether the change helped?
A good answer sounds like: “We will record the current review time and missing-information rate, then compare similar work after launch.” Include the human checking effort and exceptions. Hours released are capacity; cash benefit requires an actual use or avoided cost. Ask who collects the evidence, when you review it, and what result would lead you to change or stop the workflow.
5. What information will you access, and through whose accounts?
A good answer sounds like: “These approved materials are needed; these other records are excluded. Access is limited to named people and the agreed purpose.” Ask about vendors, permissions, retention, deletion, and offboarding. Use sanitized examples during initial evaluation. A dental practice or law firm needs the appropriate qualified review before sensitive material enters a proposed workflow; a business subscription alone does not settle the question.
6. Where does a human approve the result before it has consequences?
A good answer sounds like: “The system prepares a draft; this reviewer checks these facts and approves the next action.” Customer-facing messages need human review before sending, and consequential decisions need the appropriate qualified person. Ask the provider to show the approval point, not merely describe it. Check that a failed or skipped review cannot quietly turn into an automatic send, booking, filing, or commitment.
7. How will you test ordinary cases, exceptions, and failure?
A good answer sounds like: “We agree representative cases and written acceptance tests before release, including missing inputs and unavailable sources.” For an illustrative multi-location business, include a record with conflicting location instructions. Ask who decides whether the system passed and what evidence you receive. Do not expose real sensitive data to make a sales demonstration convincing.
8. Who inside our business must participate, and for how much work?
A good answer sounds like: “This operator supplies examples, this manager resolves process questions, and these people practice before handoff.” Ask the provider to state the expected preparation and decision load. Plan time for that work. Your busiest employee cannot simply absorb every project responsibility between customer calls.
9. What will our team be able to do without you?
A good answer sounds like: “The operator can run it, verify the result, stop it, use the fallback, and escalate with the right information.” Training should use the actual workflow and include a recovery example. Ask what written instructions remain, who can update them, and how a new employee learns the process. Attendance at a presentation is not evidence that the team can operate the system.
10. What will it cost after launch, and what support is included?
A good answer sounds like: “Here are the separate software and usage costs, the support scope, the coverage limits, and the work that needs a new proposal.” A monthly fee does not automatically include monitoring or unlimited changes. Ask what happens when a vendor changes a field, model behavior shifts, or your volume increases. Read AI consulting costs for a way to separate the spending categories.
11. What happens if the agreed work does not pass?
A good answer sounds like: “These written tests govern acceptance; these correction steps and commercial terms apply to this engagement.” Separate a delivery commitment from a revenue prediction. Ask for any guarantee's scope, exclusions, timing, and remedy in the agreement. Do not carry a guarantee from one offer into another.
12. How do we pause, exit, or move to another provider?
A good answer sounds like: “Here is the notice process, what remains available, what can be exported, and how access is removed.” Also ask who restores the manual fallback and who handles outstanding work. A proprietary managed service may not transfer as software you can run yourself. That can be acceptable if you understand the replacement effort and the retained records before signing, rather than during a dispute.
Which answers should make you pause?
Pause when a provider will name a tool but not the job it improves. Pause when “we handle everything” has no named reviewer, support hours, or exclusions. A demonstration that works only with ideal inputs needs more evidence. A claim of guaranteed revenue or immediate staff replacement needs scrutiny, not a faster signature.
Be particularly careful with promises of autonomous customer-facing communication or consequential action. Ask to see the human approval path and the failure behavior. “The model is accurate” does not answer either question.
Ask what a case study measured, over what period, and what the client authorized the provider to disclose. An anonymized example can still explain its limits. If evidence cannot be shared, ask for another evaluation of capability. Impressive numbers without a method do not prove anything about your business.
“We need to inspect that integration before quoting it” can be a useful answer. Honest uncertainty is different from hiding an unverified dependency inside a confident promise.
How does Applied Margin answer these questions?
Our three primary offers have different purposes. The $1,500 fixed AI Workflow Sprint covers one team of up to five people: a short leadership kickoff, two hands-on working sessions, two to three reusable workflows, written instructions, a prioritized opportunity map, and one follow-up session. Session lengths and scheduling are agreed before booking. Custom software, integrations, and subscriptions are separate; the fee does not automatically credit toward a build.
Implementation starts at $3,500 for the defined company-knowledge use case with agreed materials and users. Larger systems receive a separate fixed proposal. Ongoing Partnership starts at $1,500 per month, with a three-month opening term, for two 45-minute working sessions and one lightweight improvement monthly. Broader delivery and monitoring are scoped separately. Current services and terms explain those boundaries.
The optional Margin Scan is $1,500 for seven business days when deeper discovery is useful. Its report is yours. The Scan's Clarity Guarantee is at least three relevant opportunities, one ranked first move, and a written implementation decision by business day 7, or a full refund. Its full $1,500 credit applies to an Applied Margin implementation authorized within 14 days. Those terms belong to the Scan, not the Sprint or other services.
The proposal specifies the system, access, tests, training, ownership, handoff, and human review. This public page helps you ask questions; it does not replace that agreement.
What should you print or bring to the call?
Use this checklist with the person who will operate the workflow. Mark unanswered items and decide who will resolve each one before committing.
- Named job and reason to improve it
- Included work, exclusions, and dependencies
- Accounts, retained materials, and usage rights
- Baseline, measure, and review decision
- Approved information and access boundaries
- Human approval before customer-facing or consequential action
- Representative tests and failure cases
- Internal participants and preparation time
- Operator practice, instructions, and fallback
- Full cost and support limits
- Acceptance, correction, and applicable guarantee terms
- Pause, exit, export, and access-removal process
If the purchase is hard to explain afterward, use the spouse-or-CPA conversation to expose what is still unclear.
FAQ
Do I need technical knowledge to evaluate an AI consultant?
You need to understand the job, the business constraints, and who will own the result. A provider should explain its proposed workflow and limits in language you can assess. Bring a qualified technical or compliance reviewer where the work warrants one. You do not need to compare model internals to insist on clear scope, tests, access, and accountability.
Should I ask for a free proof of concept?
Ask what evidence is appropriate for the uncertainty you need to resolve. A demonstration, a reference, or a small scoped evaluation may answer different questions. Do not assume unpaid work is the only proof of competence. Any evaluation should use agreed inputs, boundaries, and success criteria, with sensitive information excluded unless its handling has been approved.
Is a provider refusing to guarantee ROI a red flag?
No. A provider should distinguish the delivery it controls from business outcomes affected by demand, staffing, pricing, and adoption. Ask for realistic assumptions, a baseline, and a decision process if the evidence disappoints. Any guarantee should specify exactly which engagement, deliverable, time limit, and remedy it covers. A broad promise is less useful than an enforceable written scope.
Does the Margin Scan guarantee also cover the AI Workflow Sprint?
No. The Margin Scan's Clarity Guarantee and 14-day implementation credit apply only to that optional assessment. The $1,500 AI Workflow Sprint is a separate engagement with its listed sessions, workflows, instructions, and opportunity map. It has no automatic credit toward implementation. Review the current services page and the written proposal for the terms of the work you choose.
What should I bring to a first conversation?
Bring a plain description of the work, who does it, what makes it difficult, and what you want to change. Existing tools and a rough sense of volume are useful too. You do not need a finished process map. Keep customer, employee, patient, and client details out of the initial inquiry; agree any later information sharing and access first.
Start with the questions, not the demo
Book the free introductory conversation. Bring the checklist and the person closest to the work. If the immediate goal is helping one team use AI on real tasks, ask about the $1,500 AI Workflow Sprint. If a solution is already understood, we can discuss implementation scope directly.