The biggest red flags are promised savings before anyone has looked at your work, a tool recommendation before your processes are mapped, builds that live in the consultant's own accounts, vague answers about where your data goes, undisclosed commissions, and no handover plan. One flag deserves a hard question; two or three together are a reason to walk away.
Apply these checks to anyone you consider hiring, including me. Most flags have an innocent version: a new consultant may have few references, and a commission is fine once it's disclosed. What separates a flag from a deal-breaker is the reply when you ask about it: a straight answer with evidence, or a change of subject.
Warning signs in the sales conversation
The first call tells you a lot, because nothing has been written down yet and people talk the way they work. For the broader list of questions to put to any candidate, see how to choose an AI consultant; the four signs below are the ones worth ending a call over.
1. Savings promised before they've seen your work
It sounds like: "Businesses like yours save 20 hours a week with our AI system." Nobody can know your saving until they've counted your volumes and timed your tasks, so a figure quoted in the first conversation is marketing, not analysis.
Compare it with an estimate built from your own numbers: "You send about 600 ready-for-collection messages a month at roughly a minute each, so automating them saves around 10 hours a month, less maybe an hour checking the exceptions." That second version can be wrong, but you can see why, and you can check it.
Ask: "What would you need to measure to estimate our saving, and will the proposal show the sum?" The innocent version: a consultant describing ranges from their own past projects, clearly labelled as past results and not a promise for yours.
2. A tool recommended in the first conversation
A garden centre owner mentions customer emails, and within ten minutes the consultant is recommending a particular chatbot platform. When a month of emails is finally counted, 212 of 300 turn out to be about opening hours, delivery charges or whether a plant is in stock: a clearer contact page, a stock note on the website and an automatic reply would have handled most of them for nothing. Tool-first advice fits your business to the tool, often the one the consultant knows best or earns from.
Ask: "What would make you recommend a different tool, or no new tool at all?" A good consultant can name the conditions. The innocent version: if you've told them your whole business runs on Microsoft 365, mentioning Copilot early is reasonable. The flag is the same product coming up whatever you describe.
3. Pressure to sign, or fear used as a sales tool
Watch for "this price is only valid until Friday", "your competitors are already automating and you'll be left behind", or "AI will replace your front desk anyway, so you may as well start now". Manufactured urgency pushes you to skip the checks in this list. Fear-driven projects also start from the wrong goal: replacing people, instead of removing the work nobody wants to do. A consultant who talks about your staff as a cost to be cut will struggle to get their help, and staff knowledge is usually where the best automations come from.
Ask: "If we started in two months instead of now, what would change?" The innocent version: a genuine calendar constraint stated plainly, such as "I can start on the 1st, or not until March."
4. Claims you can't check
This covers "our proprietary AI" that turns out to be a thin layer over a public model, case studies with no business named and nobody you can call, and technical claims that are out of date. One that still circulates: "we set the temperature to zero so it never makes mistakes." Anthropic has deprecated the temperature setting on its newer Claude models and OpenAI's reasoning models don't support it; consistent output now comes from fixed instructions, examples, structured outputs and review steps. A consultant repeating the old line isn't necessarily dishonest, but isn't current either.
Ask: "Which model and whose account does this run on?" and "Can I speak to a client whose project is live?" The checks in spotting a ChatGPT wrapper and vetting case studies and references go further. The innocent version: a newer consultant with few references who offers a small paid pilot instead, so you can judge the work directly.
Warning signs about your data and accounts
These three are the flags most owners miss, because they sound technical. They're really about control: who can switch your automations off, who can see your customers' messages, and whose interests shape the advice.
5. Everything built in the consultant's own accounts
The automations run in the consultant's Make organisation or Zapier account, with your Gmail connected. The AI calls go through an API key on their card. It works, until they stop paying, get busy elsewhere, or you part ways, and then your workflows stop with no warning. A typical way this shows up: the card behind the consultant's API account expires, the AI step starts failing, and the first anyone hears of it is a customer asking why nobody answered their enquiry last week.
Moving things later is possible but not free. Zapier can copy Zaps to another account, but the copies arrive switched off with placeholder connections that must be re-authorised. Make's exported blueprints carry the modules and settings, not the connections. Every connection has to be rebuilt by someone who knows what it was for. More detail is in who owns the AI workflows a consultant builds.
Ask: "Can you build in accounts we own, with you added as a user?" The innocent version: building a prototype in their own sandbox during discovery, with a written plan and date for moving it into your accounts.
6. Vague answers about where your customer data goes
Ask a simple question and listen to the shape of the answer. Here is the same question answered two ways:
Question: When the AI drafts a reply, where does the customer's message go?
Weak answer: Don't worry, it's all secure. We use enterprise-grade AI and everything is encrypted.
Strong answer: The email text goes from your Gmail to Zapier, then to OpenAI's API, which doesn't use API data for training by default. The draft comes back as a Gmail draft for your team to approve. Apart from Zapier's run history, nothing is stored anywhere else, and I'll put that in a one-page diagram.
The difference matters because plans differ. Consumer ChatGPT and Claude accounts have a model-training switch in their privacy settings; business plans such as ChatGPT Business and Claude Team don't train on business content by default. A consultant who can't say which kind of account your data passes through hasn't thought about it.
Ask: "Can you draw the data flow on one page, including every service that stores a copy?" The innocent version: "I need to check that vendor's current terms and I'll confirm in writing." Vendors change their terms often, so that answer is a good sign.
7. Commissions or referral fees they haven't mentioned
Many software companies pay the people who bring them customers, and the sums are real. Make's affiliate programme pays 35% of a referred customer's subscription payments for 12 months. HubSpot's Solutions Partner programme pays partners 20% of net revenue from a qualifying sale for up to 36 months. Microsoft's CSP partners buy licences at wholesale and set their own price, so a partner quoting you Copilot seats may earn a margin on each one. Nothing is wrong with partner status: many good consultants are certified partners precisely because they know a tool well. The flag is not being told, because a commission can tilt a recommendation between two tools that would both do the job.
Ask: "Do you receive any payment, discount or credit from the tools you're recommending?" A straight answer, even "yes, 20% for three years", is a good sign. Disclosure done well reads something like: "I'm a Make affiliate and earn a commission if you sign up through my link. I've recommended Zapier here because your team already uses it; if you'd rather use Make, I'll build it the same way." The fuller set of checks is in checking whether your AI consultant is independent. The innocent version: disclosed partner status, with a recommendation that would stand without it.
Warning signs in the written proposal
A proposal is the first thing you can hold a consultant to, so read what's missing as carefully as what's there.
8. Running costs left out of the quote
The proposal prices the build and says nothing about what you'll pay every month afterwards. AI automations have three kinds of running cost: subscriptions, per-task automation fees and model usage. Subscriptions such as ChatGPT Plus or Claude Pro don't cover API calls, which are billed separately per token. An "AI by Zapier" step uses 1, 3 or 5 tasks per run depending on the model tier chosen. On Make, each scheduled check uses a credit even when nothing has arrived, so a scenario checking every 15 minutes uses about 2,880 credits a month before doing any work, well above the free plan's 1,000.
Ask: "What will this cost to run each month at our volumes, and at double our volumes?" The innocent version: a range with its assumptions written down, because exact usage isn't known until the thing runs. A good running-cost section can be three lines long: "Zapier Professional at the 1,500-task tier, $39 a month billed annually. Model usage for about 80 quote drafts a month, a few cents to a few dollars depending on the model chosen. Text messages at your SMS provider's rate."
9. No test plan or acceptance criteria
"We'll build an AI assistant for your enquiries" is a direction, not a deliverable. Without a definition of done, every disagreement at the end becomes a matter of opinion, and you tend to lose it. Compare: "Accepted when it drafts correct replies for at least 45 of 50 real past enquiries, flags the rest for a person, and never states a price that isn't in the current price list." A good test sheet lists each case with its expected result, for example: "Test 14: customer asks for a discount on a repeat booking. Expected: polite no, mention the loyalty card, no invented offer. Result: pass." Fifty lines like that settle arguments before they start.
Ask: "How will we both know it works before the final invoice?" The innocent version: criteria that are still rough at proposal stage, with an agreement to fix them after discovery and before any build starts.
10. Products that are retired, retiring or renamed
AI products change monthly, so check every named product in a proposal. Some recent examples: OpenAI is retiring custom GPTs, which stop running on 11 December 2026; Copilot Pro was withdrawn, with support for existing subscribers ending on 1 August 2026 and Microsoft 365 Premium replacing it; ChatGPT agent was retired on 9 July 2026 and replaced by ChatGPT Work; OpenAI closed the Sora API on 24 September 2026; Excel's COPILOT worksheet function was retired on 14 September 2026; and OpenAI removed its Assistants API on 26 August 2026, so anything a developer proposes to build on it is already obsolete. A proposal built on any of these either wasn't checked or was recycled from an older deck.
Ask: "When did you last check the status of each product in this plan?" The innocent version: a proposal drafted a few months ago that the consultant updates the same week you point it out.
Warning signs about what happens after go-live
The last two flags only bite months later, which is exactly why they need settling before you sign.
11. Customer-facing AI with no human check or disclosure
The plan has AI replying to customers on its own from day one, with no hand-off to a person when it's unsure and no mention of telling customers they're talking to a bot. If you sell to customers in the EU, the AI Act's transparency duties, including telling people when they're interacting with a chatbot, have applied since 2 August 2026. The same law has banned AI that infers people's emotions at work, including in recruitment, since 2 February 2025, so a consultant pitching mood detection on staff calls is selling something unlawful there.
Ask: "What does a customer see when the AI isn't sure, and who gets alerted?" The innocent version: fully automatic messages that are low-risk and templated, such as "your order is ready for collection". The flag is free-form AI answers going out unchecked. Picture how it goes wrong: a nail salon's unchecked assistant tells a customer that gel manicures are refunded if they chip within 14 days. The salon has no such policy, but the customer has it in writing, and the owner now has a dispute and a review to answer.
12. No handover plan, so only a retainer keeps it running
The pitch usually goes: "We look after everything monthly, so you'll never need to touch it." That's a comfortable promise and a lock-in. You should be able to pause, run and fix the basics without the consultant, which means owning the accounts, holding exported copies of the workflows, having the final prompts and a short runbook, and knowing the monthly costs. Ongoing support can be good value, but it should be a choice, not the only way your business keeps working.
Ask: "What exactly will you hand over, and could another consultant take over from it?" Use the AI consultant handover checklist as your list. The innocent version: a complex system where the consultant recommends a support arrangement and still documents it well enough for someone else to run.
Ten minutes of checking between the first and second call
Most of these flags can be tested cheaply before you commit to anything. A short routine to run on every candidate:
- Check the named products. Open the vendor's own page for each tool in the proposal and confirm it still exists under that name and plan. This alone catches flag 10.
- Ask for one redacted document from a past project. A runbook, a test sheet or a data-flow diagram with the client's details removed shows how they actually work far better than a case study does.
- Read their public writing against today's facts. If their recent posts still describe retired products as current, expect the proposal to do the same.
- Ask one reference a specific question. "Could you run and fix the automations after they left?" gets a more useful answer than "were you happy?"
- Put your key questions in writing. The six below take a consultant ten minutes to answer, and written answers can go into the contract.
None of this needs technical knowledge. It needs the habit of checking a claim against the source before accepting it, which is the same habit you'd want from the consultant.
Signs that look bad but are usually fine
Being even-handed cuts both ways. These often put owners off, and usually shouldn't:
- They won't quote a fixed price on the first call. Without seeing your processes, a fixed price is either padded or a guess. A priced discovery stage is the honest alternative.
- They charge for discovery. Mapping your work takes real time. What matters is that discovery ends with a written output you keep, even if you hire someone else to build.
- They tell you not to use AI for something. A consultant who talks you out of automating a job that only happens twice a month is saving you money.
- They recommend a cheaper tool than you expected, or one you already have. Often a sign of independence rather than inexperience.
- They say "I don't know, I'll check." Vendor features and terms change monthly. Certainty about everything is the worrying sign.
- They're a certified partner of a tool. Fine, as long as they told you before you asked.
Raising a flag without souring the conversation
You don't need to be confrontational. Framing a question as your own due diligence usually gets a better answer than an accusation. Wording you can adapt:
"Before I compare proposals, I'm asking everyone the same questions:
1. Which accounts would this be built in, and who owns them at the end?
2. Can you sketch where our customer data goes, service by service?
3. Do you receive any commission or credit from the tools you recommend?
4. What will it cost to run each month at our volumes?
5. How will we both know it works before the final payment?
6. What will you hand over, and could someone else maintain it?
Written answers are fine; I'd like to compare like for like."Good consultants tend to welcome this, because it separates them from the competition. If the answers come back vague, or not at all, you have your answer.
A dry cleaner scores two proposals
Here's an illustrative comparison. A dry cleaner with two counters and a collection van wants to automate "ready for collection" texts (about 600 a month), draft alteration quotes from photos and notes (about 80 a month), and chase garments left uncollected after 30 days.
Proposal X promises "at least 25 hours saved every week", builds everything in the consultant's own Make account, uses a custom GPT for alteration quotes, prices the build but not the running costs, and includes "ongoing management" at a monthly fee with no handover section.
Proposal Y starts with a paid two-week discovery, builds in a Zapier account in the cleaner's name with the consultant added as a user, drafts quotes through an AI step that a staff member approves before sending, lists running costs, sets acceptance tests and ends with a handover list.
| Red flag | Proposal X | Proposal Y |
|---|---|---|
| 1. Savings promised before discovery | Yes ("25 hours a week") | No; estimate after discovery |
| 5. Built in the consultant's accounts | Yes | No |
| 8. Running costs missing | Yes | No |
| 9. No acceptance criteria | Yes | No; 50 past orders as the test set |
| 10. Retired or retiring products | Yes (custom GPT, stops 11 December 2026) | No |
| 12. No handover plan | Yes | No |
| Flags raised | 6 | 0 (discovery fee is on the "usually fine" list) |
The running-cost line in Proposal Y is worth checking yourself, because it shows the kind of sum a good proposal contains. Each ready-for-collection text is a Zap with two action steps (send the text, log it), so 600 texts use about 1,200 Zapier tasks a month, more than the 750 in Zapier's entry Professional tier. Add the quote drafts: 80 a month through an AI step at 1, 3 or 5 tasks per run adds 80-400 tasks. The total of roughly 1,280-1,600 tasks sits around the 1,500-task Professional tier ($58.50 a month billed monthly, $39 billed annually) or the 2,000-task tier ($73.50 or $49), plus text-message charges from the SMS service. If usage runs past the tier, Zapier's pay-per-task billing, on by default for accounts opened since January 2024, charges the extra tasks at 1.25 times the base rate on annual plans and 2.5 times on monthly ones, and stops Zaps at three times the plan limit; with pay-per-task switched off, Zaps pause at the task limit instead. The ten-hour saving estimate is also believable: 600 texts at about a minute each.
Six flags in Proposal X is an easy decision. Real choices are usually closer: one or two flags on each side, with answers that differ in quality. In that case, weigh the replies you got when you asked, not just the count, and put the answers you were given into the contract.
Further reads
- Should You Pay an AI Consultant on Results? Success Fees Explained — When paying on results makes sense, and how to define the result.
- AI Consulting Contracts: 9 Clauses to Check Before You Sign — The contract clauses that turn these checks into obligations.
- How to Scope an AI Project: Deliverables and Acceptance Criteria — Write deliverables and acceptance criteria before anyone builds.
- What an AI Consultant Can't Do for You, and What You Must Own — What stays your job however good the consultant is.
- Where to Find a Good AI Consultant for a Small Business — Where to look for candidates worth putting through these checks.
- Red Flags When Buying AI Marketing Software or Services — The equivalent warning signs for AI marketing software and services.
- Is It Safe to Put Customer Data Into ChatGPT? — The data questions behind red flag six, in more depth.
- What Does an AI Implementation Consultant Actually Do? — The six stages of an AI implementation consultant's work, the evidence each stage should leave behind, and how the role differs from strategy or IT help.
- What an AI Consultant Does for a Law Firm, and What It Costs — The five phases of a law-firm AI engagement, what each should hand over, how fees are built from consultant days, and proposal red flags.
- AI Consultant or DIY: Which Suits a Small Professional Firm? — When a small professional firm can set up AI itself, when outside help earns its fee, and the split approach many firms settle on.
- When Should a Trades Business Bring In an AI Consultant? — The signals that a trades firm needs outside AI help, a ten-minute scoring sheet, what to gather first and what a good engagement leaves behind.
- Should You Hire an Automation Consultant or Build It Yourself? — A five-question test, what DIY really costs in hours, one build that worked and one that didn't, and the hybrid route most small firms should take.
- AI Consultant vs Agency vs Freelancer: Which Should You Hire? — Who diagnoses, who builds and who supports: a decision table for consultants, agencies and freelancers, with a coffee roaster's three jobs sorted.
- AI Automation Agency Pricing: What You Pay For and What's Extra — How AI automation agencies price their work, what the build fee should include, what's extra, and a furniture maker's agency quote decoded line by line.
- Free AI Consultations: What You Get and How to Use Them Well — What fits into a free AI consultation, why advisers offer one, and how to prepare, question and compare calls so you leave with a real diagnosis.
- Scope Creep in AI Projects: How to Handle Change Requests — A testable baseline, a one-page change request, sizing rules and a worked change log for keeping an AI project to its agreed scope.
- How Much Does an AI Readiness Assessment Cost? — Free, vendor-funded and paid readiness assessments compared, where the days go, your staff's share of the cost, and a way to compare three quotes fairly.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Make affiliate programme page; HubSpot Solutions Partner Program policies; Microsoft CSP partner information; Zapier pricing and pay-per-task billing pages; Zapier pricing page and help articles on copying assets between accounts; Make's help pages on scenario blueprints; OpenAI help pages on custom GPT retirement and ChatGPT Work; Microsoft support notices on Copilot Pro and the COPILOT function; OpenAI notices on the Sora discontinuation and the Assistants API removal; EU AI Act (Articles 5 and 50).