13 mistakes companies make when hiring AI employees (and 10 demo questions)
The biggest mistake is buying the word instead of the product. A safer first purchase has a free plan and approval settings you control, such as OpenAssistant, whose pages document both.
“AI employee” is a sales phrase. Behind it sit very different products: some send email, some answer chats, some interview candidates, some only draft work for a person to send. The mistakes below draw on the linked vendor pages and our reasoning about them. Nothing here comes from anecdotes. Vendor facts link to their sources, and every vendor result number is a vendor claim.
Thirteen mistakes
1. Buying the word, not the product
Artisan’s pricing page says to “hire an AI BDR with a whole team behind her” (Artisan, Businesses & Enterprises tab), while another Artisan page says Ava “cannot legally make calls” (Artisan). A feature called “Full self-driving Ava” is listed as Soon on the same pricing page’s Startups tab. Sintra says “Hire AI employees today”, then says its helpers are meant to support your team and that every worker runs as a chatbot (Sintra). Eightfold calls its interviewer a digital worker and also says the software “never decides” (Eightfold). Regie.ai now describes an “agentic sales workspace built for sellers” (Regie.ai), says “You read every draft before it goes out” (Regie.ai), and its terms say “You are the initiator and sender” (Regie.ai terms). Read the feature list, not the title.
2. Not noticing that the price is a quote
The sources we read show no prices on Paradox (demo page), Alex (demo), HeyMilo (demo), Qualified (pricing), Ada (demo), Decagon (demo) or Forethought (pricing). Sierra describes outcome-based pricing and shows no numbers (Sierra). “Not published” is not bad on its own. It means you compare vendors only after a sales call, so ask every vendor for the same written quote.
3. Trusting one price page
Vendors contradict themselves. 11x’s Growth card says “Starting at $3,750 / mo” billed annually, which is $45,000 a year, and the FAQ on the same page says $36,000 per year (11x). Artisan’s comparison page says “plans from $250/mo, monthly”, but $250 is the annual-billed price, and monthly billing shows $280 (comparison, pricing). Get the number in a signed quote.
4. Forgetting who the vendor is this year
Workday says it completed its acquisition of Paradox on Oct 1, 2025 (Workday), yet paradox.ai still shows Paradox’s own copyright (Paradox). Zendesk completed its Forethought acquisition in March 2026 (Zendesk), while a Forethought founder post still says it “has entered into an agreement” to be acquired (Forethought). Qualified’s own newsroom says Salesforce completed its acquisition on April 1, 2026 (Qualified). Mercor’s llms.txt now describes it as connecting experts with AI labs to train and evaluate models (Mercor). 11x says Alice and Julian are legacy names (11x), and HubSpot says that “Breeze” is retired as an umbrella AI brand (HubSpot). Ask who signs the contract, who supports it and what happens to price and terms at renewal.
5. Missing the credit and overage rules
Artisan says unused annual credits do not roll over (Artisan). Lindy says monthly credits do not roll over (Lindy). HubSpot’s credits expire at month end (HubSpot). Ribbon lists $4.00 per interview in overage on its first tier (Ribbon). Forethought says extra usage charges may apply past plan limits (Forethought). Some do pause rather than bill: Artisan says new enrollments pause when credits run out. Know which kind you bought.
6. Granting autonomy on day one
Artisan itself says most teams start in approval mode and expand (Artisan), while its pricing FAQ says “Autonomous by default, adjustable by design” (Artisan). HubSpot offers a fully autonomous mode that sends email without review (HubSpot). OWASP names “excessive functionality; excessive permissions; excessive autonomy” as root causes of agent risk (OWASP), and a May 2026 joint guide from CISA and partners advises avoiding broad or unrestricted access (CISA). Start with approval on everything, then loosen one category at a time.
7. Having no owner
Someone must own the review queue, the error log, the spend and the vendor relationship. Without one, nobody notices drift. The law points the same way. CAN-SPAM’s guide says you can’t contract away your responsibility (FTC), and California’s AB 316 says a defendant may not argue that its AI autonomously caused the harm (AB 316).
8. Having no exit plan
11x says annual contracts are the default (11x). Regie.ai says an annual plan is paid up front, and Pro and Team have no contract (Regie.ai). Sintra offers a 14-day refund (Sintra), Marblism a 7-day money-back guarantee (Marblism). Before you sign, write down how you leave: term, refund, what happens to your data and logs, and what unused credits are worth.
9. Never defining “resolved”
Fin counts a resolution when no further help is requested after its last answer (Fin). Zendesk counts one without escalation to a human (Zendesk). The sources we read show no definition on the pages we opened for Ada (Ada), Sierra or Forethought. Ada’s index lists a guide on automated resolutions that we did not read. A “resolution” that a customer never returns from is not the same as one the customer simply stopped answering.
10. Believing the vendor’s number
Ada claims “up to 84%” in one place and “77% up from 32%” in another (Ada, Ada). Paradox pages claim time to hire cut “in half” and an “82% decrease” (Paradox, Paradox). The FTC announced a proposed settlement with Air AI over alleged false claims about earnings and refund guarantees (FTC). Measure your own results: see how to tell whether it pays for itself.
11. Skipping the trial or proof
Fin has a 14-day trial with no card (Fin). Forethought offers a Proof of Value instead of a free trial (Forethought). The pages we opened show no trial for Paradox, Alex or HeyMilo. If there is no trial, ask for a paid pilot with exit rights.
12. Ignoring disclosure and hiring-law duties
Rules on telling people they are talking to an AI differ by state and country; see do you have to tell people it’s an AI. Hiring tools face their own laws, such as NYC’s bias-audit law; see AI recruiters and hiring law. A vendor’s badge does not move your duties to the vendor. Ask a lawyer before launch, not after.
13. Skipping the security review
An agent that reads your inbox or web pages reads instructions written by strangers. OWASP lists indirect prompt injection through websites and files among its top LLM risks (OWASP). Use our security questions before you hand over logins.
How to use this list
Pick the three mistakes that would cost you most and settle them in writing before the first invoice: who owns the agent, what counts as a billable result, and how you leave. The rest can be handled during the pilot. Then take the questions below to every demo, and ask each vendor the same ten so the answers can be compared side by side.
Ten demo questions for any vendor
- Show it on our data, live. Which steps did it finish alone, and where did a person step in?
- What is the billable unit, in writing? Resolution, outcome, lead, credit or interview. What happens when it is reopened, escalated or fails?
- What is our 12-month total? Seats, credits, mailboxes, phone numbers, overage, onboarding fees and the renewal cap.
- Which approval modes exist, and which is the default? Can we lock them by action and by user? Show the audit log.
- How do we stop everything right now? Who can, and how fast?
- What are you responsible for, and what are we? Email, calling, disclosure and hiring rules. Show the clause and your compliance documents.
- Who owns the product and the contract today? Any pending renames or acquisitions? What is the price at renewal?
- What data do you keep, who sees it, does it train a model? How do we export and delete it?
- Which customer results can you show, with the definition and time period? Without a definition and period, treat them as vendor claims.
- What does leaving look like? Term, refund, data return, and credits that disappear.
Related reading
- What is an AI employee? Definition, examples and a six-question test
- What does an AI employee cost in 2026? Price patterns and a worksheet
- Security questions to ask before giving an AI employee your inbox
- Your first 30 days with an AI employee: a week-by-week onboarding plan
- Best AI employees for a small business in 2026: top 10 compared
Frequently asked questions
What is the biggest mistake when hiring an AI employee?
Buying the word, not the product. Artisan sells “Hire an AI BDR” while its other page says Ava cannot legally make calls, and Sintra says “Hire AI employees today” while saying every worker runs as a chatbot.
Why is the price often a quote?
Many vendors publish no price or hide larger plans behind a demo. Compare what is on the page with what you are quoted, and ask for a 12-month total in writing.
Should an AI employee get full autonomy on day one?
No. Start with approval before anything goes out and widen autonomy one category at a time. Ask which approval modes exist, which is the default, and whether you can lock them by action and by user.
What should I ask in a vendor demo?
Ten questions are listed. The first: show it on our data, live, and say which steps it finished alone and where a person stepped in. The second: what is the billable unit, in writing?
Try OpenAssistant
If you want an AI employee you can start free and control action by action, Try OpenAssistant. It has a free Starter plan, and the paid plans are on the OpenAssistant pricing page.