What should I ask a customer service outsourcing provider before I sign?

Published 23 September 2026

Every shortlisted provider should be asked the 18 questions on this page before you sign. They are not a generic RFP template. Each one exists because one of the 12 criteria we score against could not be answered from public sources alone, and the gap it leaves is exactly what a procurement conversation should close.

Our main ranking of 29 customer service outsourcing providers prints these 18 questions as a plain list. This page is the working behind them: for each question, which part of the scoring it comes from, what a top answer actually has to show, and where this research already found (or could not find) a public answer.

Evidence you can check yourself

Which of your current clients look like us in size, industry and language mix? May we speak to one?

A case study about "a leading airline" tells you nothing you can check. One that names the client, dates the work and puts a number on the outcome tells you everything. Named references are also what this industry publishes least. (How we judge client evidence)

≥2 case studies naming the client, dating the engagement and quantifying an outcome; strongest when at least one is corroborated outside the provider's own site.

What certifications do you hold, with certificate numbers, scopes and dates?

Naming a standard is not the same as letting you check it. A certificate number plus the issuing body can be confirmed in a public registry in about two minutes; a badge on a web page cannot. One of the 29 providers we looked at publishes one: Sutherland. (How we judge certifications)

One provider in 29 (Sutherland) publishes a certificate identifier confirmed in the certifier's own registry. The rest name standards without a number, body, validity date or scope, including some inside regulatory filings that establish the standard is held but not what it covers.

For PCI DSS: what level, when was the current AoC issued, and by which QSA?

There is no such thing as a "PCI certificate", so anyone waving one is describing something that does not exist. The real evidence is a current Attestation of Compliance, or a listing on a card network's own register of compliant providers. In marketing copy "certified" and "compliant" mean the same thing: nothing, until one of those appears.

A stated service-provider level plus a current AoC reference or registry listing where payment flows are actually served, not just the phrase "PCI DSS compliant."

People and delivery stability

What is your agent attrition, annualized, on accounts like ours, and how is it calculated?

Almost no provider publishes attrition, so from outside we read it indirectly through employee reviews, weighing the spread and how recent they are rather than the headline score. Ask anyway: a provider that can tell you how it calculates its own number gives you something no review site can. (How we read employee signals)

A rating ≥4.0 across at least 25 reviews with recent activity and no contradiction between what employees and what clients describe.

Which languages on your list are delivered by native or C1+ speakers today, from which sites?

The number of languages a provider markets and the number it staffs are different figures, and the gap can be enormous. We count a language as confirmed only when the claim lines up with real hiring at a named site, or a client who names it. (What happens when you check language counts)

Naming the hub each language is delivered from, not just the language itself: a claim tied to a location is a claim that can be checked against that hub's hiring activity.

Walk us through your last seasonal ramp: numbers, timeline, service level held.

"Fast ramp" with no numbers attached is a sales line. A ramp described with the agent count, the timeframe and the service level they held while doing it is evidence somebody has actually done this before. (How we judge ramp evidence)

A named case: N agents added in W weeks, service level stated before and during the ramp.

Show us your QA methodology: sampling rate, rubric, calibration cadence. When was the last calibration session?

"We have a quality team" is the answer that scores zero. What separates a real quality system from the word used as decoration is a stated sampling rate, a scoring scheme, a regular calibration session, and a route back to the agent who needs the feedback. (How we judge quality systems)

A stated sampling rate, a named rubric or scoring framework, and a calibration cadence, ideally with a recent date, since a QA system that hasn't calibrated recently is not demonstrating the practice it claims.

Who exactly will manage our account, and how many other accounts do they carry?

Account-team load is invisible from the outside. Nothing a provider publishes tells you how thinly a delivery manager is spread, and no amount of scoring public evidence can reach it, which is exactly why this one has to be asked in the room rather than looked up.

A named manager, their current account count, and what happens to your account when one of the others has a crisis. "You will have a dedicated team" without a number is the answer to be wary of.

Money and exit terms

What is included in the rate, and what is billed separately?

Almost nobody publishes a rate card and we do not expect one: of the 29 providers scored, 24 publish nothing about pricing in any form. What you can reasonably insist on is the logic of the bill: what sits inside the rate, and what puts an extra line on the invoice. (What providers publish about pricing)

Inclusions and exclusions named, even without a number attached: "per-FTE, covers coaching and QA, excludes overtime and off-hours premiums" is a real answer; "competitive pricing" is not.

What are the exit terms: notice, knowledge transfer, data return?

Exit terms are published even more rarely than pricing, which is telling in itself: the cost of leaving is the part of the deal a provider has least interest in making clear before you sign. (How we judge commercial transparency)

Stated notice period, a knowledge-transfer process, and a data-return or deletion commitment on exit: on the record, not "we'll work it out."

What is the smallest team and shortest term you'll start with, and what does a structured pilot look like?

A provider you can only reach through a long enterprise sales process scores at the bottom here no matter how capable it is, because a mid-sized buyer cannot practically get in. This question finds that out in one answer. (How we judge entry terms)

A pilot offering with a stated size (≤10 FTE) or term (≤3 months) and defined success criteria, ideally with a case referencing a pilot that converted to a larger engagement.

AI and what stays human

What runs on AI in your delivery today, and what is the human-in-the-loop control on each AI touchpoint?

"Agent assist" said out loud is not evidence of anything. What counts is which parts of the work run on AI, which tools do it, and who checks the output. Only five of the 32 providers we assessed publish any stated control over what their AI produces.

A named workflow, a named tool, and a specific control, not "we use AI responsibly."

What stays human in your model, deliberately?

This question is scored under the same part of the framework as the previous one. A top answer rewards a provider that states a limit on AI honestly over one that claims AI handles everything: a stated boundary is evidence of a considered position, not a weakness to hide.

A specific category of decision named as staying human, with a reason.

Regulated and sensitive work

Where would our data be processed, and can you guarantee EU-only processing contractually?

Where your data physically gets processed is a contract question, not a marketing one. A provider who can point you at the clause has answered it; one who says "we're GDPR compliant" has not. (How we judge data handling)

Data-residency options stated in public materials (EU processing available on request), not first learned during contract negotiation.

Who are your subprocessors?

A published subprocessor list is the good answer. "Available on request" is the common one, and it means you cannot know who else touches your data until after you are already in conversation. (How we judge data handling)

A public list, or at minimum a named process for how the list is shared and how you're notified when it changes.

List every subprocessor touching our data, by entity and country, and tell us how we are notified when that list changes.

The sharper version of the question above: a provider that can answer with a specific notification mechanism is operating the disclosure as a process, not producing it on demand for this one RFP.

A named change-notification process, not a one-time list.

Which delivery entities and facilities are named in the scope of the certificate you just cited? Send the scope statement, not the badge.

A certificate can be completely genuine and still not cover the company or the building where your work would actually happen. Scope is the detail marketing pages leave out more than any other. (How we judge certifications)

The scope statement itself, naming the specific entities and locations it covers.

For any capability acquired in a merger: which certifications extend to the acquired entities today, and which are still being brought into scope?

A parent company's certification does not automatically stretch to a business it has just bought. TDCX is the worked example: its own published evidence says whether the three sites it gained in the 2025 Open Access BPO acquisition fall inside its existing certificates simply is not stated anywhere. (How we judge certifications)

A named timeline for bringing acquired entities into an existing certification's scope, not a claim that the group certificate already covers everything.

What this page does not do

It does not score your shortlist for you, and it does not say which answer is good enough for your situation. A pilot minimum that excludes a small buyer is a fair answer from a provider built for enterprise scale; the bottom of that measure score there is a description, not a verdict.

What it does is turn the criteria this research applies from public evidence into the same questions asked directly, so a provider's private answer can be checked against the same standard as its public one.

Related: what outsourcing companies publish about pricing and how many publish a control over their AI, two of these eighteen questions in full, on their own pages, with the field-wide counts behind them.