Win the room before booking

You leave the review with a page of notes. Risk wants to know what happens if the agent says something it should not. Compliance asks whether callers will be told they are talking to AI. IT asks where the recordings sit and which models touch them. Your deck answered none of these, because it was written to sell the idea, not to clear it.

Approval for an AI voice agent is not one pitch. It is three or four separate yeses, and each team gives its yes only on its own evidence. In regulated sectors, banks especially, risk is the usual blocker. The fear is an agent that "goes rogue" or makes a public mistake, and those buyers ask for proof from real deployments, not a better demo. You can walk into each team's room knowing the questions it will ask and the proof to bring.

Four teams, four first questions

Use this table as your map. Each row is a team you will have to satisfy, the question it will open with, the worry behind that question and the proof to bring.

Team The first question What they are really worried about Evidence to bring
Risk "What happens if it goes rogue?" A public mistake that nobody can stop quickly Named failure modes, a narrow script, stop and escalation rules, and examples from other regulated deployments
Compliance "Does the caller know it is AI, and did they consent?" Breaking disclosure and consent rules that differ by region A region by region map, a disclosure and consent script, a do-not-contact process, a transcript sample
IT and security "Where do the data and recordings live, and which models touch them?" Customer data leaving their control through an integration A data-flow one-pager, a security report, a subprocessor list, an integration design
Business sponsor "What lift will it deliver, and what does it cost to find out?" Paying for an experiment that cannot prove its worth A small pilot, a baseline set before launch, payment tied to measured lift

These first questions are the ones regulated buyers raised on our calls. Risk fears the rogue agent, and the buyers who raised it asked for risk-mitigation examples from other regulated institutions to build their internal case. Data location, models and ownership come up early, and some buyers ask for a SOC 2 report. Financial services buyers also ask what breach liability and compliance guarantees they should require before handing over customer data.

Compliance questions change with geography. Colorado, for example, will require public-facing conversational AI to disclose that it is AI from 1 January 2027. The compliance step later on sets more regions side by side.

The sponsor's question has the same logic. Banks do not want to pay to find out whether it works. They want a short pilot on a small sample, with payment after a measured lift.

Bring each team the answer in its own language, and never send the same deck to all four. Risk wants limits, compliance wants a map, IT wants a data flow and the sponsor wants a number. A single deck that tries to serve all four ends up answering none of them.

Give risk a failure budget

Do not try to convince risk that the agent will never go wrong. Agree in advance how it is allowed to go wrong, and what happens each time. Buyers in banking put the fear in their own words: the AI "can go rogue", and the business wants to know how it would handle that if it did. Risk will not accept an answer that says it won't.

Write down the failure modes before the meeting. Four cover most of it: a wrong answer, an off-script promise, a mishandled dispute and a silent line. For each one, name the action: escalate, stop or log. Then narrow the scope. Start with one outbound use case and a fixed script, as we advise. Outbound is simpler to bound, because you choose who is called and what is said.

Expect four questions:

  • What can the agent say?
  • What does it do when it does not know the answer?
  • Who reads the transcripts?
  • How fast can we switch it off?

Treat the human handover as a design decision, because buyers raise it unprompted. They ask whether the handover is polite, what happens with a sensitive issue or a disputed amount, and whether anyone chases a handover that nobody picked up. Route disputes and sensitive topics to a person by rule, and give an owner to every handover that goes unanswered. Customers expect this too. A Gartner survey found that 87% of 3,566 customers say a company using GenAI for service must offer a way to reach a human.

One commercial bank we spoke with chose to hold off on customer-facing AI until its internal governance framework was finalised. Do not wait for that framework to arrive. Offer to help draft it, with your failure modes, stop rules and handover design as the first inputs.

A framework you helped write is a deliverable. A framework you waited for is a delay.

Give compliance a disclosure map

Compliance does not need reassurance. It needs a map: for each region you call into, what must be disclosed, what consent is required and what proof you keep. Build one row per region and have counsel confirm it before the meeting. Four examples show how much the rows differ:

  • EU: Article 50 of the AI Act requires any AI that talks to people, automated phone systems included, to say it is AI from the first interaction. It has applied since 2 August 2026, and the July omnibus delay did not move it.
  • US federal: AI-voice calls already need consent under the TCPA, following the FCC's 2024 ruling. The FCC's proposed rule to announce AI at the start of a call was still not final in mid-2026.
  • Colorado: public-facing conversational AI must disclose it is AI from 1 January 2027, and disclosure rules are spreading state by state.
  • Nigeria: the central bank's March 2026 AML guidelines accept AI in compliance work but expect transparency, governance and independent model validation at least yearly. That is a model governance bar, not a caller disclosure rule.

Whatever the local minimum, set the agent to answer truthfully whenever a caller asks whether they are speaking to an AI. Whether it announces itself first is a business choice. Lying when asked is never one.

Expect these questions:

  • Where is the consent record, and who can produce it?
  • Which calling hours and retry limits apply?
  • How does a do-not-contact request reach the dialer?
  • How long are recordings kept?
  • Who validates the model, and how often?

Compliance will also ask what callers say once they are told it is AI. Buyers worry that disclosure invites tangents from irritated contacts. We have no data on how callers react once told, so plan to measure it in the pilot.

Bring four artefacts:

  1. A consent and disclosure script.
  2. Calling-hour and retry settings.
  3. A written do-not-contact process.
  4. A sample of real transcripts.

Check that your tooling can enforce the second and third.

Answer IT before IT asks

IT will not judge the idea. It will judge the data flow, so bring a one-page diagram before anyone requests it. Buyers ask where their data and call recordings live, which AI models touch them and who owns them. Answer those three questions in writing, then add the rest of the one-pager:

  • What goes in: the minimum fields the agent needs per contact, such as a name, a number and an amount due. Do not send the whole customer record.
  • What is stored: recordings, transcripts and outcomes, and where each sits.
  • For how long: a retention period for each item, and how deletion happens.
  • Who can read it: roles, not names, and how access is logged.
  • Whether it trains models: a plain yes or no, in the contract as well as the sales deck.

Ask the vendor for a SOC 2 report and a named list of subprocessors. A SOC 2 report tests a provider's controls against the AICPA trust criteria, and security is always in scope. Ask for a Type 2 report if you can get one. Type 1 only shows that controls were well designed on one day, while Type 2 tests that they worked over a period of months. Read the system description for the subprocessors and data flows, because that is where the answers sit.

Be straight about hosting. Residency rules are real and they are spreading. Nigeria's central bank told banks and payment firms to keep payment transaction data on servers in Nigeria by 1 January 2027. Other regions have their own residency rules, so ask the vendor early where data is hosted and whether that location can be fixed.

Then cover integration. A telecom buyer expected a CRM link to need an API integration, and said data protection over that link would be the real question. Another asked whether the tool replaces the contact centre system or builds around it. Prepare answers on:

  • restricted API access and least privilege
  • what the agent does when the CRM or the vendor is down
  • whether a spreadsheet upload is enough for the pilot

Security teams may also start asking for third-party red-team certification of AI agents. One AI agent vendor now sells such a certification, which tests for manipulation and data leaks, to security and risk teams. It is not yet a standard demand, but expect the question.

Make the pilot the evidence pack

One small pilot can answer all four teams, if you design it as evidence rather than as a launch. Keep it to a single use case and a small sample, a few hundred contacts at most. Between 100 and 200 numbers is small enough for a person to read every transcript. Put a human on the dashboard while calls run, and write the stop rules into the plan before the first dial. Risk gets its failure budget, compliance gets its transcripts, IT gets a bounded data flow, and the sponsor gets a number.

Agree the success measure with every approver up front, and set the baseline before launch. The sponsor's condition from earlier, payment only after a measured lift, depends on it. Measurement is the usual gap. Gartner looked at 432 AI customer-service use cases and found that 42% could not be measured at all, and only a quarter showed positive ROI. If you cannot state today's contact rate or cost per outcome, you cannot show a lift later.

Test before any real contact is called. Talk to the agent yourself, then run a test call to your own phone, then review the transcripts after each call. Ask risk and compliance to do the same, because a team that has heard the agent handle a dispute trusts it more than a team that has read about it.

Insight7 Voice campaigns (beta) let you test the agent in the browser with no phone number, and a reviewer can read each transcript as soon as a call ends.

Order the meetings on purpose

Go to the sponsor first, then risk, then compliance, then IT. The sponsor sets the scope and the return you will be judged on. Risk and compliance then set the limits, and IT has to build inside those limits. If IT is the first to see the project, it is looking at something nobody has defined yet.

Check who your sponsor is before you book anything. Deals stall when the person driving them is neither the user nor the budget holder. In Nigerian enterprise deals they move only when a director or C-level sponsor is engaged early, and we would expect the same elsewhere. Name an executive owner before the formal reviews begin. When several departments could use the tool and none has committed, the bank needs one of them to adopt it first, so that its result can carry the case to the others.

When a team says no, ask what evidence would change the answer. Write that evidence into the pilot as a test.

Give your champion a financial case they can carry upstairs. Buyers weigh the cost of the voice agent against what it costs them to acquire or serve a customer, and they ask for those numbers to make the internal argument.

Related reading: AI Voice Agent Business Case Template for Presenting to Leadership, How AI Voice Agents Redact Sensitive Data From Call Recordings and Transcripts and Contact Center AI Risk Assessment Checklist for Customer-Facing AI Agents.

Questions teams ask about approval

Who should own the project once it is approved?

Give it to the business owner who holds the number, not to IT or risk. Risk, compliance and IT stay as named reviewers. Make one person accountable for the day-to-day controls: who reads transcripts, who can stop a campaign, and who chases an unanswered handover. If that person is not named before launch, each team will assume another team has it.

Can we start with inbound instead of outbound?

You can, if the scope is tight. A narrow line, such as after-hours calls or a short list of FAQ topics, with a rule-based handover to a person, is the version easiest to approve. Make disputes, sensitive topics and anything off the FAQ list automatic handovers. Then count how often callers hit those handovers in the first weeks, because that count tells risk whether the scope can widen.

Do we need to tell callers it is AI if the law does not require it yet?

In a regulated business, default to yes. The safer position, if anyone challenges the calls later, is to have disclosed. Disclosure rules are spreading, so review the script whenever the rules change. Treat the wording as a pilot variable. Test a short, plain disclosure against your measures and keep the one that holds up.

What if our data is in silos, our customer base is small or the budget is closed?

"Not now" is a legitimate answer, and several buyers give it for exactly these reasons. Check the pilot's real requirements before you accept it. A narrow outbound pilot needs a list of contacts and a handful of fields, not a unified customer record. If the base is too small to show a lift, or the budget really is shut until next year, use the wait. Set your baseline, write the failure modes and draw the data-flow page. When the budget reopens, you will be asking for a decision on evidence, not starting a review from scratch. Many firms plan budgets late in the year, and Nigerian firms do so in Q4, so the groundwork is most useful just before it.

Approvals are earned by narrowing the ask, not by widening the pitch. Regulated buyers tell us the blocker is risk, and risk relaxes when the scope is small enough to be bounded. Each team will say yes to one use case, one script and one set of stop rules long before it says yes to "AI voice". If you want to talk through your approval plan, book a demo with Insight7.

What is the one failure this team would need to see handled before they said yes?