The memo lands on a Tuesday. Service desk moving to a partner, transition next quarter, “no impact to service levels anticipated.” Half the floor already knows how this goes, or thinks they do, and the thread about it runs five hundred comments deep: scripted agents reading from a card, password resets that take three days, tickets closed at 4:59pm so the clock stops.
Here is what those threads usually miss. Outsourced service desks fail at very different rates, and the variable that predicts it is not the country on the org chart. It is the contract. An outsourced service desk succeeds or fails on three decisions made before a single agent is hired: where Level 1 scope ends, who owns the knowledge base, and whether the agreement pays the vendor for closing tickets or for resolving them. Location is a minor variable next to those.
So the useful question is not “does outsourcing IT service desk work.” It is “which of those three did my company already get wrong, and can I still see the evidence.” You usually can, months before any SLA report admits it.
“Closed” is not “resolved,” and your SLA probably cannot tell the difference
Almost every reopened-ticket horror story traces back to one design flaw: the contract measures ticket closure and time-to-close, and nothing else with teeth.
Consider what that incentive does to a rational agent graded on average handle time and tickets per shift. A ticket they cannot fix damages both numbers. So it gets closed with “user unreachable,” or closed with a link to a KB article, or closed and immediately reopened under a new number, which resets the clock and books a second closure. Nobody involved is lazy or dishonest. They are responding correctly to what you chose to pay for.
The fix is boring and it works: measure resolution, not disposition. Pair First Contact Resolution with a reopen rate and a ticket-recurrence rate on the same user or asset within 14 days. MetricNet, which has benchmarked service desks for decades, defines first contact resolution as the percentage of contacts resolved by the service desk on the first interaction with the customer, counting an emailed or web-logged ticket only if it is resolved within one business hour. Anything escalated to another group is disqualified.
Two details get lost in procurement. First, there is gross FCR and net FCR. Gross counts every contact. Net excludes contacts that could never have been fixed at Level 1 anyway, like a failed SAN controller. Net is the one to contract on, because gross FCR punishes a desk for tickets you never gave it the rights to fix. Second, MetricNet’s benchmark library uses a different headline metric for in-house desks than for outsourced ones: “Level 1 Resolution Rate” for in-house, “Net Level 1 Resolution Rate” for outsourced. The benchmarking firm assumes an outsourced desk has a narrower permitted scope. Your contract should say out loud what that scope is.
For a sense of scale on where FCR lands: writing for HDI in 2011, MetricNet’s Jeff Rumburg reported a worldwide average net FCR of roughly 74%, with a range across benchmarked desks running from about 41% at the low end to about 94% at the high end. That spread is the point. A greater than twofold difference between desks doing nominally the same job is not explained by geography. It is explained by scope, tooling, and tenure.
Where Level 1 ends is the whole contract
The single most common failure I see is an L1 scope defined by ticket category rather than by system access.
Category-based scope sounds sensible in the statement of work. “L1 handles account lockouts, password resets, MFA enrolment, software installs from the approved catalogue, printer issues, VPN connectivity, mobile device enrolment.” Then go check what the vendor’s agents are actually granted in AD, Intune, your MDM, your identity provider and your PSA. If they can read but not write, every one of those categories is a ticket they will open, triage, document beautifully, and hand to your remaining internal team.
That is the expensive ticket router, and it is why internal engineers end up hating the arrangement. Volume did not fall. It rose, because every ticket now carries an extra hop and a handoff document, while the cost model assumed L1 would absorb most of it.
The money argument for a well-scoped desk is real, and the numbers point one way. Using 2010 North American benchmark data published by HDI, Rumburg put the average cost per ticket at desktop support, the tier you escalate into, at $62, with a minimum of $27 and a maximum of $490. By type, the average cost per incident was $48 and per service request $113. Those figures are old and directional only, but the ratio is durable: tickets get dramatically more expensive as they travel up a tier, and service requests, the routine “please give me a thing” work, are the expensive ones to escalate. The economic case for an outsourced L1 rests entirely on it closing that routine work. If it cannot write to the systems, it cannot close the work, and the case collapses.
Practical test, and you can run it this week: take the ten highest-volume ticket categories from the last quarter. For each one, write down the exact permission the fix requires. Then ask whether the vendor’s agents will hold it on day one, on day 90, or never. Any category in the “never” column is a category where your ticket count is about to go up.
Deflection is a real number and also a lie
Deflection rate gets quoted in every transformation deck: portal sessions or chatbot conversations that ended without a ticket being created, expressed as a percentage. It is a legitimate operational measure. It is also the easiest number in the building to fake, because “no ticket was created” and “the problem went away” are not the same event.
A user who searches the portal, finds nothing, and walks to a colleague’s desk counts as deflected. So does one who abandons a chatbot halfway. So does one who simply cannot find the “raise a ticket” button.
So if you are handed a deflection figure, ask for it alongside three things: total contact volume across all channels including walk-ups and direct-to-engineer messages, the self-service completion rate (sessions that reached a confirmed successful action, like an actual password reset executed in the portal), and repeat-contact rate within 48 hours. Deflection that is real shows up as flat-or-falling total contact volume. Deflection that is fake shows up as falling ticket volume and rising everything else.
The knowledge base is the actual deliverable
Here is the uncomfortable part for the internal team. Most in-house desks run on undocumented tribal knowledge, which is fine when the same six people have been there four years and stops being fine the moment the work moves to anyone else, including a new internal hire.
An outsourced desk is a knowledge base with humans attached. If the KB is thin, the agents will be exactly as good as the KB and no better, because they have no other source. The transition period decides this, and it is almost always underfunded.
What a serious knowledge transfer looks like, in rough order:
- Shadowing, 2 to 4 weeks. Vendor leads and senior agents listen to live tickets without touching them. They are writing, not resolving.
- Reverse shadowing, 2 to 4 weeks. Vendor agents take tickets with your staff watching and correcting in real time. Every correction becomes a KB edit.
- Assisted operation, 4 to 8 weeks. Vendor owns the queue; your team is on call and reviews a daily sample of closed tickets.
- Steady state with a KB ownership clause. Somebody is contractually responsible for article accuracy, article aging, and new-article creation from tickets that had no article.
The clause most contracts forget is the last one, and specifically the “no article existed” path. If an agent resolves a ticket that had no runbook, the resolution needs to become an article before the ticket can be closed, and that article needs review. Without it, knowledge stays in individual agents’ heads. Which brings us to the thing that then destroys it.
Attrition sets the ceiling on everything else
Rumburg’s FCR analysis identified agent training hours as the primary driver of first contact resolution. Training hours compound with tenure, and so does familiarity with your weird legacy line-of-business app and the permissions quirk on the finance team’s shared mailbox that nobody documented. Attrition burns all of it.
Contact centre attrition is genuinely high, and worth being precise about. Verint’s “State of Agent Experience 2026,” published April 2026, surveyed 1,000 agents at contact centres with 300 or more seats across insurance, telecom, banking, retail and BPOs. It found 31% said they were likely to leave within six months, with a stark age split: 46% of agents aged 18 to 34 versus 8% of those 45 and over. That is self-reported intent, not measured turnover, so treat it as a temperature reading. For an actual rate, a ContactBabel analyst quoted in coverage of the survey put the industry average near 30% annually.
It is worth separating two numbers that get conflated in these arguments. The large Indian IT services firms report company-wide attrition in the low-to-mid teens: in Q1 FY26 results reported in July 2025, TCS reported 13.8%, Infosys 14.4%, HCL Tech 12.8% and Wipro 15.1%. Those are enterprise-wide IT services figures covering engineers, consultants and delivery staff. They are not the attrition rate on a voice-based L1 tower, which typically runs well above the company average. If a vendor quotes you their corporate attrition number during a service desk bid, that is not an answer to the question you asked.
Ask instead for attrition on your account specifically, monthly, split between agents who left the company and agents who moved to another account inside the vendor. The second number predicts your pain, because internal poaching onto a shinier account damages tenure exactly as much as resignation does. Then add a minimum-tenure clause and a notice requirement before a named lead rotates off.
MetricNet lists turnover, alongside handle time, agent utilisation and absenteeism, as a direct driver of cost per ticket. It is not a soft HR number. It is the input that quietly sets your FCR ceiling.
Shadow IT is the failure signal that never shows up in a report
When users lose faith in a desk, they do not complain. They route around it. This is the most reliable early indicator of trouble, and it is invisible in every dashboard the vendor produces, because a ticket that is never raised cannot be counted by the people counting tickets.
The pattern is consistent. Department heads start DMing the one internal engineer whose number they still have. A team buys its own SaaS tool on a departmental card because “IT takes three weeks.” Someone in ops writes a script that touches production. A local “IT person” emerges in a regional office, usually the most technical person in a non-IT role, doing unlogged, unpatched, well-intentioned work.
Where to look for it, concretely:
- Expense reports for software and cloud charges under the approval threshold
- New OAuth app grants and third-party integrations in your identity provider
- Direct messages to internal engineers as a share of their week (just ask them to log it for two weeks)
- Local admin grants and unmanaged devices appearing on the network
- The gap between ticket volume and your total headcount trend, which should track loosely and will diverge when people give up
By the time this is visible, CSAT may still look fine. That is not a contradiction, it is a sampling problem: only people who still use the desk get surveyed, and the ones who gave up hold the worst opinion. CSAT is lagging and self-selecting. Net FCR and total contact volume lead it. If you have room for one number on the exec dashboard, do not make it CSAT.
Predict your own outcome
Take these as signals rather than rules. They call it right far more often than the general argument about offshoring does.
Strong signals it will work: the contract names net FCR and reopen rate, not just closure and time-to-close. Agents get write access to identity, MDM and the software catalogue by a defined date. A knowledge transfer period of at least eight to twelve weeks is funded and staffed by your people, and those people are not simultaneously being made redundant. Somebody on your side owns the relationship full time, with a title and a calendar, not as a side duty. Attrition is reported at account level monthly. The vendor asked hard questions about your environment during the bid, including ones you could not answer.
Strong signals it will not: the business case is expressed entirely as headcount cost per seat. The transition is four weeks. The internal team doing the knowledge transfer has already received notice, which means the knowledge transfer will be exactly as good as their goodwill. Nobody can tell you what the current FCR or reopen rate is, so there is no baseline and therefore no way to prove degradation later. The SLA has a ticket-volume-based pricing term, which pays the vendor more when more tickets exist. And the one that predicts everything: the scope document lists ticket categories but not system permissions.
What good looks like once it is running: total contact volume flat or down while ticket volume drops; net FCR trending up quarter on quarter; reopen rate visible and controlled; a knowledge base growing from resolved tickets rather than a one-time import; the same named lead a year later; and your internal engineers doing project work instead of password resets, which was the entire point. That outcome is achievable. It is just not the default, and none of it happens by accident.
Common follow-up questions
How long before I can fairly judge the new desk? Expect a dip through transition and for roughly a quarter after go-live. Judge at month six and again at month twelve. Month nine is often the real test, because that is when initial-cohort attrition starts eroding the tenure built during knowledge transfer.
Named agents or a pooled model? Under about 2,000 users, a dedicated pod is usually worth the premium, because environment familiarity is most of the value. Above that, a pooled model with a dedicated lead and a minimum-tenure clause is more realistic. Either way, insist on notice before rotation.
Is AI-driven deflection changing this? It shifts the mix rather than removing the problem. Automation genuinely closes the highest-volume scripted work, which makes everything left in the human queue harder. Handle time rises and FCR falls even when the desk is improving, so reset your baselines. If the SLA is not adjusted for a harder residual mix, you will penalise the vendor for your own automation.
Can I get the baseline if we never measured it? Partially, and it is worth the effort before transition. Pull reopen counts, ticket-to-ticket recurrence on the same user within 14 days, and category volumes from your existing ITSM tool. Even a rough baseline stops the argument in month eight from being about whose memory is right.
What if the decision is already made and I think it is wrong? Put the three questions in writing to whoever owns the contract: what permissions L1 will hold, how many weeks of knowledge transfer are funded, and whether the SLA measures resolution or closure. A written question creates a record, and it occasionally changes the answer, because these terms are more often unexamined than deliberately bad.
If you are standing up a support function, taking one back in-house, or auditing one that is not delivering, AB7 Solutions works on both sides of that problem: BPO and KPO delivery teams, IT support staffing and staff augmentation, and the technology services work behind them. Happy to look at your scope document and metric set and tell you plainly where the risk sits, including when the honest answer is that you should keep it internal. Call +1 321 341 7733, email ab@ab7solutions.com or director@ab7solutions.com, or read more at www.ab7solutions.com.
Sources: Jeff Rumburg, “Metric of the Month: First Contact Resolution,” HDI (thinkhdi.com); Jeff Rumburg, “Metric of the Month: Cost per Ticket,” HDI (thinkhdi.com); MetricNet, “First Contact Resolution Rate” and “Service Desk Cost per Ticket” (metricnet.com); Verint, “The State of Agent Experience 2026,” April 2026, as reported by No Jitter (nojitter.com); Q1 FY26 attrition figures reported by TCS, Infosys, HCL Technologies and Wipro, July 2025.
Related reading
Related reading
Related AB7 service: BPO and back-office outsourcing.