Pricing and Buying
What to Ask an AI Consultant Before You Hire One
Most AI projects do not fail because the model was wrong. They fail because the buying decision was wrong. RAND found that more than 80 percent of AI projects fail, twice the rate of failure for information technology projects that do not involve AI. MIT’s Project NANDA put a number on the business side: its 2025 report found that 95 percent of organizations are getting zero return from enterprise generative AI.
Those two numbers describe one pattern. Companies buy a demo instead of a process. The chat interface performs beautifully in the sales meeting, then meets your real data, your real approval steps, and your real staff.
The person you hire decides which side of that split your budget lands on. Here are the eight questions I ask before any engagement starts, and the answers that tell you to walk away.
Key Takeaways Before You Interview an AI Consultant
- Process before technology: A serious consultant studies how your business runs before naming a single tool.
- Enterprise data readiness is not optional: Expect an audit of your documents, records, and systems before anyone builds an agent.
- Financial metrics, not adjectives: Every deliverable ties to cost removed, revenue added, or hours returned.
- Security and governance are design decisions: Ask in writing how your proprietary data stays out of public training sets and who approves each new use case.
- Handoff beats dependence: The engagement ends with your team running the system, not renting it.
Where Enterprise AI Projects Stall Before They Reach Production
Company Data Lives in Five Systems and Agrees in None
Language models need clean, reachable information. Most companies keep procedures, customer records, and financial history locked inside separate platforms with conflicting formats. When an agent pulls answers from that mess, output quality collapses on day one.
A capable architect builds the retrieval layer before the application, because RAG architecture is only as good as the documents underneath it. Skip that step and you get confident, wrong answers. Teams then spend more time correcting the software than the software saves them.
Automation Multiplies a Broken Workflow
Technology amplifies the habits you already have. Automate a disorganized sales follow-up and you get disorganized follow-up at higher speed. Leaders often expect software to solve what is actually a management problem.
Fix the process first. Map the human steps, remove the friction, then automate the predictable remainder.
Feature Chasing Replaces Business Outcomes
New model releases land every few weeks and pull executive attention with them. Voice agents and custom copilots get built because a competitor built one. That produces expensive software for hypothetical situations instead of real customer demand.
Strong deployments target repetitive, high-volume work that drains staff hours. If a project does not cut cost or raise output, it is decoration.
Nobody Owns the System After Handoff
Models drift, prompts rot, and source documents change. Projects die quietly when the consultant leaves and no internal owner holds the keys. Name that owner during scoping, not during the postmortem, and write the AI governance rules that decide who approves a new use case once the consultant is gone.
Security and Compliance Arrive Last
Legal and IT review usually starts after the prototype impresses someone. That sequence kills projects late and loudly. Access rules, retention terms, and audit logging belong in the first design conversation.
Shadow AI Spreads Faster Than the Approved Tool
Staff waiting on a slow internal system paste customer records into whatever public chatbot loads fastest. That is shadow AI, and it moves proprietary information outside your control without a single approval step. Ask how the rollout plan gives employees a sanctioned option that beats the unsanctioned one on speed, and what interim policy covers the gap.
Vendor Lock-In Turns a Pilot Into a Permanent Bill
Proprietary platforms hold your prompts, your embeddings, and your evaluation history in formats you cannot export. Switching providers later means rebuilding the retrieval layer from scratch at full cost. Keep the orchestration layer, the vector store, and the prompt library inside your own repository so the model underneath stays replaceable.
Cost and Latency Surprise the Finance Team
Token spend scales with usage, and a chatty agent gets expensive fast. Multi-agent workflows compound it, because every handoff between agents adds another round trip and pushes response time past the point where staff route around the tool. Ask for a cost-per-transaction estimate and a latency target before the build starts.
Eight Questions to Ask an AI Consultant, at a Glance
| Question to Ask | What a Strong Answer Sounds Like | Red Flag |
|---|---|---|
| How do you connect AI to my existing workflows? | Maps the order, quote, or ticket flow before naming any tool | Pitches a platform in the first meeting |
| How do you evaluate enterprise data readiness? | Audits documents, systems, and retrieval quality first | Says the model will figure it out |
| How do you measure return on the build? | Names the metric, the baseline, and the review cadence | Promises productivity without a number |
| Custom models or existing APIs? | Orchestrates proven APIs and customizes context and instructions | Proposes training a foundation model from scratch |
| How is my proprietary data protected? | Zero-retention settings, role-based access, private deployment option | Cannot describe the provider’s retention terms |
| What happens when the model is wrong? | Guardrails, evaluation sets, human review on high-risk output | Claims the system does not make mistakes |
| How does my team take the system over? | Documentation, runbooks, live training, one named internal owner | Keeps the keys and bills indefinitely |
| Do you work in generative engine optimization? | Structures content and entities so LLMs cite the brand | Sells keyword placement only |
Question 1: How Do You Connect AI to Existing Business Workflows?
A reliable consultant starts by examining how your business makes money. They map the operational bottlenecks before recommending any cloud architecture. Ask them to walk you through their discovery process step by step.
You want a framework that keeps human oversight in place and defines the handoffs between software and staff. If they pitch a vendor before asking how your team processes orders today, end the conversation. Automation without a process map is failure at higher speed.
The right partner spends real time interviewing your frontline staff, because that is where the friction actually lives.
Question 2: What Is Your Process for Evaluating Enterprise Data Readiness?
Garbage in, garbage out still governs applied machine learning. Ask how they plan to clean and structure your internal documents. They need a repeatable method for outdated PDFs, inconsistent spreadsheets, and legacy databases.
The answer should cover how they index your content and stand up retrieval-augmented generation inside your security perimeter. A good architect delays the application build until the data foundation holds. They should be willing to tell you your databases are not ready yet.
Question 3: How Do You Measure the Return on an Implementation?
Refuse vague promises about productivity and innovation. Ask for the specific numbers they track after launch. Response time, cost per resolved ticket, customer acquisition cost, and direct revenue lift all qualify.
My own practice runs on process-first execution: fix the data pipeline, then layer answer engine and generative engine optimization on top of it. That sequence produced a $12 million revenue lift for one client and more than $5 million in annual sales impact across engagements. Ask any consultant to tie their deliverables to a number you already report to your board.
Question 4: Do You Build Custom Models or Use Existing APIs?
Training a proprietary foundation model is almost never the right call for a small or midsize company. It burns compute budget, senior talent, and quarters of calendar time. Ask whether they plan to build on established APIs from providers such as OpenAI, Anthropic, or Google.
A practical consultant handles LLM orchestration with frameworks like LangChain and vector stores like Pinecone, then layers agentic workflows on top. The customization belongs in your context, your instructions, and your retrieval layer, not in the base model. That keeps costs predictable and the system upgradable.
Question 5: How Do You Handle Security and Proprietary Data?
Pushing financials or customer records into a public model creates compliance exposure you cannot undo. Press hard on their privacy standards. They should know how to configure zero-data-retention terms with API providers and produce the documentation to prove it.
They should also deploy open-weight models on private infrastructure when the sensitivity demands it. Ask specifically about role-based access control and how they stop an employee from querying restricted records. One leak costs more than the entire project.
Question 6: What Happens When the Model Gets It Wrong?
Language models produce confident, incorrect answers. You need the consultant’s specific plan for containing that. Ask how they build guardrails, evaluation sets, and validation layers into the application.
The answer should include prompt testing, structured output constraints, and human review on anything high-risk. An experienced architect assumes failure and designs so errors get caught before a customer sees them. Anyone claiming an error-free system has not shipped one.
Question 7: How Do You Handle Staff Training and Adoption?
Launch day is the halfway point. If your employees route around the new tool, the project failed regardless of the code quality. Ask how knowledge transfers to your internal team.
Expect documentation, written procedures, and hands-on sessions with the people who use the system daily. The goal is internal capability, not a permanent retainer. They also need a straight answer on change management and on the job-security anxiety that automation creates.
Question 8: Do You Work in Answer Engine and Generative Engine Optimization?
Buyers now find answers inside ChatGPT, Gemini, and Perplexity before they ever reach a blue link. A current consultant builds for that reality. Ask how their strategy handles answer engine optimization and entity structure.
If organic traffic feeds your pipeline, your advisor must know how to structure site data so language models cite you as a source. Keyword placement alone no longer earns the citation. This is the layer most AI consultants skip entirely.
How to Choose an AI Consultant From the Answers You Get
Score every candidate on three things: how long they spend on your process before mentioning a tool, whether they name a metric without being prompted, and whether they plan their own exit. Consultants who lead with architecture diagrams are selling the part that is easiest to build.
Three kinds of firm answer these questions differently. A global practice such as Accenture brings bench depth and a procurement cycle to match, while a specialist product firm such as Analytics AIML, which I co-founded, ships narrower systems faster, and an independent consultant gives you direct access to the person doing the thinking. Match the shape of the firm to the shape of the problem rather than to the size of the logo.
The strongest partners challenge your assumptions, audit your workflows, and demand a working data foundation before writing code. They also talk in financial terms rather than technical vocabulary. That single habit predicts whether the project survives its first budget review.
If you want a strategist who fixes the process before writing code, review my AI consulting services, then send me the workflow costing you the most. I will tell you honestly whether AI belongs in it.
Frequently Asked Questions (FAQs)
What does an AI consultant actually do?
An AI consultant audits how your business operates, finds the work that drains hours, and designs generative AI and agentic workflows around it. The role connects your business goals to the technical build and the data work underneath it.
Do I need clean data before hiring an AI consultant?
No. You need the willingness to fix it. A good consultant audits your systems first and gives you a sequenced data readiness plan before any generative tool goes live.
What does hiring an AI consultant cost?
Three variables set the price: the state of your data, the number of systems the solution must reach, and whether the consultant advises or also implements. I publish no rate card because those variables change the number dramatically. Ask every candidate to price the discovery phase separately so you can compare like with like.
How long does an AI implementation take?
Duration follows scope, not calendar convention. A single workflow with clean source data moves fast, while a build that spans multiple systems, security review, and staff training runs considerably longer. Ask for a phased plan with a working deliverable at the end of each phase.
What is the difference between an AI consultant and an AI automation agency?
An AI automation agency sells build capacity: it wires tools together against a scope somebody else defined. A consultant defines that scope, decides which work should be automated at all, and owns the governance, data, and handoff questions around it. Hire the agency when the decision is already made, and the consultant when it is not.
Will autonomous AI agents replace my enterprise support team?
No. Current agents handle routine questions and triage well, and they lack human judgment on anything unusual. The implementations that work equip your existing staff to resolve hard cases faster instead of cutting headcount.