AI migration in BPO in 2026 — Why the 20% rule beats going all-in

At a glance
- AI migration in BPO is the phased shift of contact-centre or back-office workflows from humans to AI.
- Most migrations fail because of overreach: real-world AI accuracy sits at 60% to 70%, which breaks at scale without humans in the loop.
- The safer playbook is to start at 20%, measure against humans, refine, then expand only when accuracy is proven.
- BPOs and clients that move gradually win long-term; those who go 100% on vendor promises usually end up reversing.
What is AI migration in BPO?
AI migration in BPO is the structured, phased shift of contact-centre, back-office, or support work from human-only delivery to AI-assisted or AI-led delivery.
- It moves work in stages, not all at once.
- AI runs alongside humans and is measured against a human baseline.
- Work only expands when accuracy is proven, not when a vendor says so.
Unlike a one-off automation project, migration is phased. First, AI is put on a defined slice of work. Next, it runs beside humans and is measured against their results. Then it is refined, and only then is it expanded.
The distinction matters because BPO work has very low tolerance for error. Every customer chat is brand exposure. Every misrouted refund is a cleanup cost. Every wrong answer in a regulated industry is a compliance risk. So an accuracy gap that is fine in an internal tool can be serious in a customer-facing workflow.
In 2026, the AI-in-BPO conversation is dominated by all-or-nothing pitches. Vendors promise full agent replacement, public BPO share prices wobble on the thesis, and executives chase the cost savings. However, the data is starting to push back. McKinsey’s State of AI in 2025 reports that 88% of organisations now use AI in at least one function. Still, only around a third are scaling it enterprise-wide, and 51% have reported negative impacts from AI use.
RAND Corporation has separately found that more than 80% of enterprise AI projects fail, roughly twice the rate of non-AI IT projects. Meanwhile, MIT’s NANDA Initiative reports that only 5% of generative AI pilots produce measurable return.
Operators actually running these migrations in production report something more grounded. They see partial gains, real accuracy gaps, and timelines longer than the slide deck suggests. In the 591st episode of the Outsource Accelerator Podcast, Michael Bian, CEO of SixEleven BPO, a 20-year-old, 8,000-agent BPO based in Davao, has settled on a deliberate 20% framework that keeps his clients’ rollouts intact.
Why most AI migrations in BPO fail
To begin, the headline failure rates are bad enough on their own. Customer service is especially exposed. For example, the failure mode is visible, public, and brand-damaging: bad answers to live customers. For the same reason, many leaders still ask whether AI will replace call centre agents at all.
McKinsey’s 2026 AI Trust research puts it bluntly. As AI shifts from advising to acting on its own, the cost of getting it wrong grows a lot.

For example, the most widely cited cautionary tale is Klarna. In 2023, the Swedish fintech replaced about 700 customer service agents with an OpenAI-built chatbot. At first, it claimed the AI handled two-thirds of all customer queries. However, by May 2025, CEO Sebastian Siemiatkowski reversed course publicly. He admitted the all-AI approach had produced “lower quality” support. Customer satisfaction had reportedly dropped by around 22%.
Klarna is now rebuilding human capacity in a hybrid model. In that setup, AI handles routine queries and humans handle escalations and complex cases. The same split now shapes how most teams approach AI in customer service.
The pattern recurs because the math underneath is unforgiving. As Michael put it:
“AI is only sixty to seventy percent accurate. That’s something I try to explain to my clients, that AI will not solve all your problems, but it’s a really good tool to augment and help productivity. A human in the loop is still needed to counter-check its work.”
A 60% to 70% accuracy rate sounds promising in isolation. However, applied to a contact centre running thousands of interactions a day, it means an unmanageable volume of wrong outcomes. Without humans in the loop to catch and correct them, the errors compound. That is also why barriers to AI adoption show up fastest in live, customer-facing work.
The 20% rule: a framework for AI migration in BPO
In contrast, Michael’s framework is simple and tested in production:
“I always tell clients not to do a hundred percent migration, just do a twenty percent migration. Let’s compare apples to apples, the result of the human versus the AI.
Let’s refine the AI, then let’s gradually adjust our staffing once this new form of technology is ready and we are sure that it is accurate.”
The 20% figure is not arbitrary. It is a slice large enough for real comparison data, small enough to contain errors, and set up to measure AI against a clear human baseline. Still, the other 80% stays human-led. As a result, it provides both fallback capacity and the comparison set.
Importantly, the migration only expands when the AI’s accuracy has been proven against humans on the same workload. So it does not expand when the vendor claims it is ready. In practice, a typical phased path looks like this:
| Phase | AI Coverage | Typical Duration | Gate to Next Phase |
|---|---|---|---|
| Baseline | 0% | Pre-migration | Establish human KPI baseline: CSAT, resolution rate, escalation rate, action accuracy |
| Pilot (20% rule) | 10% to 20% | 3 to 6 months | AI matches human CSAT and resolution rate on the same workload |
| Expansion | 30% to 50% | 3 to 6 months | Stable accuracy across edge cases; escalation rate at or below baseline |
| Mature hybrid | 60% to 80% | Ongoing | Documented playbooks for failure modes; human QA continuity preserved |
5 best practices for AI migration in BPO in 2026
Next, translating the 20% rule into procurement terms gives BPO clients a clear checklist. In short, five practices separate well-run migrations from the failure statistics.
1. Start at 20%, not 50% or 100%
The urge to go bigger faster is the single biggest predictor of failure. A 20% slice gives you enough data to test real performance. Still, it keeps the bulk of contacts away from error risk. Bigger initial slices look efficient on the slide deck. In practice, they tend to be the first thing reversed.
2. Define accuracy benchmarks before kickoff
First, accuracy must be measured against the same workload the humans handle, on the same metrics: resolution rate, CSAT, escalation rate, and accuracy of the action taken. Still, Klarna’s lesson is that volume metrics can look fine while quality quietly collapses underneath. Deloitte’s 2026 State of AI in the Enterprise report, based on 3,235 leader interviews, names workforce readiness and outcome measurement as the most common gaps in scaled AI programs.
3. Build the knowledge base with the BPO, not in parallel
Again, most migrations fail at the knowledge base, not the model. For example, the BPO running the workflow knows the edge cases, the historical exceptions, and the customer-specific quirks. So excluding the BPO from building that base guarantees gaps that only surface in production. The same care supports well-run AI-augmented BPO services, where the model and the team improve together.
Michael’s onboarding playbook captures the principle directly from the BPO side:
“The first 30 days, 30 to 60 days, it’s an observation period and learning period for us. Before we even try to give any suggestion, we have to have a very good understanding of your product or your service, not only from the big picture perspective, but also from the day-to-day operations.”

4. Keep humans in the loop for QA and refinement
The human-in-the-loop layer is not a temporary artefact. It is the part of the system that spots degradation, catches new failure modes, and feeds corrections back into the model. As a result, removing it too early is the most common path to a Klarna-style reversal. Good AI call centre software is built to pass hard cases to a person, not hide them.
5. Tie staffing changes to proven accuracy, not vendor promises
Finally, the hardest discipline in any migration is keeping headcount decisions downstream of proven AI accuracy on the real workload, not upstream of vendor commitments. Because of this, the cost of reversing a too-fast headcount cut is far higher than the cost of a slightly slower migration.
FAQs
Which BPO workflows should be migrated to AI first?
High-volume, low-complexity contacts come first. For example, password resets, order status, basic FAQ queries, and simple data entry. Complex escalations, regulated interactions, and high-value account work are migrated last, if at all.
What happens to BPO agents whose work AI takes over?
The strongest operators move agents up the complexity ladder. So agents shift into QA, supervision, escalation handling, and AI model evaluation. Michael frames it as a long-running internal mobility model at SixEleven:
“A lot of our management people are actually homegrown. They started as agents but they moved up the ladder. They became our team leaders, our QAs, our operations managers. Ninety percent of our management staff are actually homegrown.”
SixEleven has also grown its back office division by training agents into AI annotation work for client training pipelines.
Is AI creating new BPO work as well as replacing existing work?
Yes. Data annotation for AI training is one of the fastest-growing BPO service lines in 2026. For example, it covers factory machine movement annotation, content moderation at scale, and scientific research data labelling. Michael sees it directly in his own book:
“AI is lessening the headcount on some projects, but the good thing is it also creates jobs in other verticals, and we are able to cater to such.”
How does pricing change when a BPO migrates work to AI?
Per-FTE billing erodes as AI absorbs hours. So the emerging alternative is outcome-based pricing: per resolution, per ticket deflected, or per outcome delivered. Still, most operators are still working out the commercial model.







Independent




