Help Me Think, Not Decide for Me

Overview

The business wanted to know where AI could help people who are shopping for a wireless provider, and where it would get in the way. I led the prospect half of a two-part study: eleven discovery interviews with the people who actually sign for business wireless, from a one-owner photography studio to an enterprise procurement team and a county elections office. What came back was a clear and stable line. People want AI to help them think. They do not want it to decide. And the assistant is not competing with other chatbots. It is competing with the AI tools these buyers already trust and use every day.

The request, and how we split it

The request that came in covered a lot of ground: how business customers across small business, enterprise, and strategic segments would want to use AI-powered search, information, and chat. It named two very different places for that to happen. The public site, where someone is evaluating a provider they have no relationship with. And the logged-in account portal, where someone who already pays you is trying to get something done.

Those are not the same person with the same stakes, so we ran them as two connected studies rather than one blended one. I own research for the ecommerce site, so I took the prospect side. A colleague took the customer side and the portal. Same request, same screener, two discussion guides, and a readout that could finally show the whole landscape instead of half of it.

That split turned out to matter for the findings. A customer stuck in a bad chat has a reason to stay. A prospect does not. That single difference explains most of what I found.

My role

I led the prospect study end to end. I shaped the research questions, co-built the shared screener, wrote the prospect discussion guide, ran every interview, did the analysis and synthesis, and presented the readout to product, design, and marketing partners.

Recruiting: the decisions before the first interview

Recruiting was where most of the quality of this study was decided. One screener had to serve both halves of the program, so it had to find people who genuinely make these decisions without narrowing the sample to one comfortable type of business. Fielded on dscout. Here is what I screened on and, just as importantly, what I refused to screen on.

• Responsibility, not job title. People qualified if they were the primary person managing their company’s wireless plans and devices, or helped manage them. Titles vary wildly across an eleven-person church and a five-thousand-person energy company. Responsibility does not.
• Recent behavior, not claimed interest. They had to have actually used a carrier portal or app within the last few months. Someone who touches this once a year cannot walk me through a process they barely remember.
• Company size left wide open. Every band from one person to five thousand and up qualified, so size became a lens for analysis rather than a limit on who got in. That is why the final sample runs from a solo owner to enterprise procurement.
• Industry left open, and quotas on current provider. No industry was excluded, which is how eleven interviews ended up spanning photography, wholesale, faith-based nonprofit, IT services, philanthropy, automotive, energy, tax software, consulting, K-12 education, and elections administration. Provider quotas made sure the pool held both existing customers and people on other carriers, so the prospect half had real prospects in it.
• I asked about AI attitude, and never used it to disqualify. The screener asked whether people seek out new AI tools, wait until they are proven, use them only when required, or avoid them. That gave me spread. It never removed anyone. A study about trust in AI that only recruits people who like AI has answered its own question before it starts.
• Two behavioral tasks before the interview. A one-minute video walking through how they manage their account, and a photo of the login page they actually use. I wanted to see the real starting point rather than the remembered one, and it gave every session a concrete opening.
• A privacy instruction written into the task, not left to chance. The photo request said in bold to send the public login screen only and not to upload anything visible after logging in. If you ask people for screenshots of a business account, you own that risk. You do not put it on the participant to guess.

Who I talked to

Eleven one-on-one remote interviews, about sixty minutes each, semi-structured, conducted on dscout.

The discussion guide, and why it was built this way

Sixty minutes, moving from their world to their current behavior to AI, in that order. The sequencing was the argument. If you open with AI, you get opinions about AI. So I started somewhere else: their role, their company, what they are actually responsible for, and the last time they evaluated or switched a provider and what triggered it. Then how they shop today and where they get stuck, still without AI in the room. By the time it came up, everything they said about an assistant was anchored to a real process they had been through.

From there the guide covered:

• What they already do with AI. What they use in their work and in their life, including AI chat assistants, and how much of it is their own choice versus what their company hands them.
• What makes an AI experience good, and what makes one bad. Not in the abstract. The specific things a chatbot has done that made them keep going, and the specific things that made them give up. This is where the standard they were going to judge us against came from, so I spent real time here.
• What AI could help with while shopping on a site. Open first, so their own priorities came out in their own order, then concrete tasks to react to once they had said their piece.
• Trust, in both directions. What would make them trust an AI answer more, and what would make them hesitate or stop trusting it, including what happens when it is wrong about something that matters.
• What they would be willing to hand over. Which details about their business they would share to get a better answer, at what point in the conversation, and what they expect back in return for it.
• Helpful versus salesy. How they tell the difference, what tips a chatbot from one into the other, and what a genuinely helpful assistant on a company’s website looks like to them.
• When they want a person instead. Where the handoff belongs, what it needs to carry with it, how fast it has to happen, and who they expect to be on the other end.
• Being offered help before asking for it. Whether they would want the site to step in based on what they are looking at, or whether that reads as intrusive.

I closed by asking them to design the thing themselves. By minute fifty, people have thought harder about this than they ever have, and asking them to build it gives you their priorities in their own order rather than their reactions to yours.

How I analyzed it

1. 1. Capture. Rawest impressions right after each session, before reading the transcript.
2. 2. Per-participant review. Each transcript read in full for behavior and contradiction, not just stated preference.
3. 3. Coding. Tagged against recurring concepts, with new codes added as they emerged.
4. 4. Affinity mapping. Grouped into themes, each pressure-tested for whether it held across the sample.
5. 5. Segment lensing. Re-examined by company size to separate universal patterns from segment-specific ones.
6. 6. Confidence weighting. Unprompted, recurring patterns treated as well-supported. Single voices held as hypotheses and labeled that way in the readout.

What I found

The bar is already set, and it is not set by you

Most participants use mainstream AI assistants in their daily work, most often Gemini and Copilot, followed by ChatGPT and Claude. Several had already used those tools to research carriers, compare rates, or build a shortlist before ever opening a provider site. Daily use has made them fluent in conversational, personalized AI, and that is now the baseline. When an on-site chatbot felt rigid or generic, a few said it was worse than the assistant they already trust. So a provider’s assistant is not competing with rival chatbots. It is competing with an opinion the buyer has already formed somewhere else, using a tool that felt better.

The first answer is the whole audition

Past experience has taught these buyers that a chatbot shows its quality immediately, so several described testing it fast and deciding almost at once whether to keep going. A weak first answer meant calling, going back to browsing alone, or leaving for another provider. They leave rather than persist, because as a prospect they have no relationship holding them there. Getting the first answer right matters more than how much the assistant can eventually do.

The exit and the handoff are two separate promises

A visible, easy path to a person is what makes people willing to try the assistant at all, because many go into a chat already hoping to reach someone and have been trapped in bots that hid the way out. The second promise is what happens after. Having to repeat everything to the next person came up again and again as the thing that makes a handoff pointless. The conversation, the inputs, and the selections have to travel with them. The visible exit earns the trust. The context carrying over keeps it.

Everyone wants a human, but for different reasons by company size

Across every size, people saw the assistant as useful for early research and not for taking a purchase all the way through. Why they wanted a person is where segments split. Smaller businesses wanted reassurance and clarity, and expected the discounts and offers to come through a conversation. Enterprise and government wanted negotiation, validation they could defend to a boss, and accountability, meaning someone answerable for the commitment. Same stated need, three different jobs underneath it.

Selling too early ends the conversation

Asking for contact information before giving a real answer reads as lead capture rather than help. Several said it would make them stop, and one described it plainly as a reason to cancel. The signal is not only the form. It is the intent behind the wording. What is interesting is that the resistance is specific to being asked too early. Most were willing to share location, line count, or current provider once they could see it improving the answer.

For a prospect, accuracy is not enough. It has to be checkable

A prospect is choosing a provider for a whole company, and that choice tends to last for years. It carries more weight than a consumer choosing for themselves and more than a customer managing an account they already committed to. So participants treated the assistant’s output as a draft to check rather than an answer to trust, especially on pricing, coverage, and comparisons. One enterprise buyer tied it straight to his job: an inaccurate number taken to leadership falls back on him. Marketing claims with nothing behind them lowered trust instead of raising it, because there was no way to confirm them.

AI is an addition to a good website, not a replacement for one

When shopping, people go to the plan and pricing pages first. The chatbot is not the first thing they reach for. Many want to compare for themselves and prefer clear plan cards or columns they can weigh on their own, because they want to see the full range and keep control of the decision. The site is also where trust gets validated: the assistant can summarize, but the buyer goes to the page to confirm.

What they actually wanted it to do

Ranked from the interviews, the most-wanted capabilities were concrete and unglamorous:

1. Compare plans, devices, features, coverage, and cost side by side, including against their current provider.
2. Show pricing, current promotions, and a realistic total monthly cost including taxes and fees, without a form first.
3. Check whether service is available at a business address, with local coverage, speeds, and reliability.
4. Explain what switching involves and what it would take.
5. Recommend a tailored package from line count, current contract, required speed, and devices.
6. Provide evidence and source links, like coverage maps and third-party information, so claims can be verified.
7. Connect them to the right person by scheduling a call, arranging a callback, or transferring to an agent.
8. Preserve the conversation and the configuration through the handoff, and email a summary so nothing has to be repeated.

What it changed

The study reframed what the assistant is for. Not a way to complete a sale, and not a replacement for the site, but the part of the journey that helps someone think, ending in a clean handoff to a person. The findings turned into design principles the team could actually hold a decision against: answer before asking for anything, cite and link so claims can be checked, separate verified fact from estimate from sales claim, keep the path to a person visible at all times, carry the context through the handoff, and match how proactive the assistant is to whether someone is exploring or executing.

Read next to the customer study, it also gave the organization something it had not had before: one view of what AI should do for people who have not bought yet and for people who already have, and where those two answers genuinely differ.

Limits, and what is next

Eleven interviews is a sample for depth, not for weight. I treated recurring, unprompted patterns as well-supported and held single voices as hypotheses, and I labeled which was which in the readout so nobody had to guess. The sample is broad across size and industry, including education and government, but each of those is one voice.

The next phase is quantitative. The interviews surfaced the factors that pull a prospect toward a person and the factors that break trust in an AI answer. A follow-up survey built from those factors would test how much each one weighs at scale, and whether the split we saw by company size holds across a larger sample. That study is planned, not yet fielded.