Going into 2027, fewer than the marketing suggests — and the gap is not small. A provider can demonstrate genuine autonomous resolution on a handful of well-defined intents, cite containment of around 80% on those paths, and allow a buyer to conclude that the same machinery runs the rest of the operation. It does not, and the contract covers all of it — a mismatch traced in detail in reporting on agentic AI in Philippine BPOs.
This is the uncomfortable version, the one that stays out of capability decks: most buyers running 2027 selection cycles cannot distinguish a working automation stack from a fluent description of one. A meaningful number of vendors have worked that out. Anyone drawing up a shortlist of providers running call centres across Manila and Cebu needs an independent technical read before signing, because a sales process is not an operation.
Key takeaways
- Containment near 80% describes selected, grounded intents, not a whole queue.
- Four unprepared questions separate operators from resellers in under ten minutes.
- Ungrounded systems produce 10% to 25% error rates, delivered fluently and confidently.
- Enterprise buyers can ground a system in weeks; mid-market buyers usually have a documentation problem first.
- Ask for last month’s evidence, not next year’s roadmap.
What does the 80% figure actually describe?
It describes performance on intents that were selected, scripted, grounded and tested for exactly that purpose — password resets, order status, balance checks, appointment changes. It is a real number on a narrow population. Extended to the whole queue in a buyer’s mind, it becomes a forecast nobody made and nobody will honour.
The slippage happens in conversation rather than in writing. Nobody states that containment applies across all traffic. The demo runs, the figure gets mentioned, and the buyer performs the extrapolation unprompted. By the time the discrepancy surfaces, it surfaces as a service level report rather than a disagreement about scope.
Ralf Ellspermann, founder and chief strategy officer at PITON-Global, sees the consequence land on the buyer every time: “You do not want to be the test subject for a vendor’s R&D.”
How do you tell a real stack from a described one?
Ask four questions that cannot be answered by a sales engineer. Which model is running, at which version, and who tuned it. Where the system retrieves its answers from, and how often that source refreshes. What it is permitted to do without a person. And who holds the authority to disable a path overnight without convening anybody.
The answers separate operators from resellers in under ten minutes. What makes the exercise reliable is that none of the four can be improvised convincingly, and all four are things an organisation genuinely running the technology would know without preparation.
| What the deck says | What to ask for | What a hollow answer sounds like |
| “AI-powered resolution” | Model, version, and the person who tuned it | “We partner with a leading platform” |
| “Grounded in your data” | The retrieval source and its refresh cadence | “It learns from your ticket history” |
| “Human in the loop” | Who can halt a path, and how quickly | “That goes to our governance forum” |
| “Continuously improving” | The test library replayed before shipping | “We watch the accuracy dashboard” |
The right-hand column is the tell. None of those replies is dishonest. Every one of them describes a company that does not control the thing it is selling, which is a different problem and in some ways a worse one.
Why does grounding decide whether any of this is safe?
Because it governs how the system fails. A model answering from its own training memory, without a retrieval layer attached to current policy, produces error rates in the 10% to 25% range — and the errors arrive fluent, confident and indistinguishable in tone from correct answers. That is not a quality gap. It is a liability generator that happens to be fast.
Buyers tend to interrogate speed and containment, which are visible, and skip the retrieval architecture, which is not. The result is a procurement process that optimises the metric a vendor volunteers and ignores the one that determines whether a regulator, a customer or a journalist ever gets involved.
The practical consequence shows up in the complaint file rather than the dashboard. A fluent wrong answer does not register as a system failure anywhere in the stack. It registers as a customer who was told something untrue by a company, acted on it, and found out later. No containment report captures that event.
Is there a real divide between large and small buyers?
There is, and it is widening. Enterprises arrive with consolidated, governed, reasonably current documentation and can ground a system in weeks. Mid-market buyers arrive with policy spread across spreadsheets, inboxes and institutional memory, and discover that automation does not resolve an information problem. It industrialises one.
That produces a two-speed market inside the same country, served by the same providers, using the same vocabulary. Large programmes run genuine autonomous resolution with measurable grounding. Smaller programmes get sold the identical language against source material that cannot support it, then conclude the technology underdelivered when the real constraint was never the model.
The honest advice for much of the mid-market is unwelcome and rarely offered: fix the knowledge base first, automate second. No vendor is paid to say that, which is precisely why it is worth hearing from somebody who is not selling the implementation.
What separates the providers that are genuinely ahead?
Evidence, produced without notice. The operators running real systems can show the last set of regression results, name the intents that changed behaviour after the most recent model version, and produce the cases their reviewers pulled last month alongside what was altered as a result.
John Maczynski, chief executive of PITON-Global, uses a single request to sort a shortlist: “Show me what your reviewers caught last week, and what you changed because of it. The good ones have it open on a screen. Everyone else sends a follow-up email.”
That test works because it inverts the usual dynamic. A capability presentation is prepared, rehearsed and designed to survive scrutiny. An operational record is not. Asking for yesterday’s evidence rather than tomorrow’s roadmap moves the conversation onto ground where only genuine operators are comfortable.
It also surfaces a subtler distinction. Some providers have the technology and no governance around it — the system works, nobody reviews what it did, and the first indication of drift is a complaint trend. Others have governance and modest technology, which is the safer failure mode and the easier gap to close. Buyers should be able to tell which one they are looking at, because the remediation plans are completely different.
What should buyers do differently this quarter?
For the 2027 cycle, stop evaluating capability claims and start evaluating operations. Send a live queue sample and watch it worked. Ask for the regression results from the most recent release. Request three escalated cases and their outcomes. Each takes under an hour, none can be staged, and together they predict performance better than any scorecard.
Procurement teams built the current evaluation for a labour purchase, where the relevant questions were headcount, attrition, language proficiency and price. Those questions remain worth asking and no longer determine the outcome. The variable that determines the outcome is whether the automation layer is controlled, and nothing on a standard scorecard asks about it.
Ellspermann is direct about where the risk has migrated: “The decision that damages a programme is almost never made during delivery. It is made during selection, months before anyone sees a number they dislike.”
That timing is what makes the shortcut so expensive. A weak provider chosen in March does not announce itself in April; it announces itself in September, by which point the integration is built, the staff are trained, the volumes are committed, and the cost of correcting the decision has multiplied several times over. Selection is the cheapest moment to be rigorous and the one most often rushed, because it is the stage where nothing appears to be at stake yet.
So which ones are real?
The ones that answer the four questions from memory, produce evidence without a week’s notice, and can name what their system is not allowed to do. That population is smaller than the marketing implies and considerably larger than the cynical read suggests — the technology works, the operators exist, and the country’s position in this market is not in question.
What is in question is whether a given buyer can identify them. The ones who can will get a genuine capability at a real saving. The ones who cannot will sign with a reseller, run a pilot that disappoints, and conclude that the Philippines oversold agentic AI — when what actually happened is that they bought a description and received one.
Call centres in the Philippines are not uniformly ahead or uniformly behind on this technology. They are spread across a wide range, the range is invisible from a capability deck, and the hour spent finding out where a provider actually sits is the highest-return hour in the entire sourcing process.
Frequently asked questions
Does 80% containment apply to all contacts?
No. That figure describes performance on intents selected, scripted and grounded for automation, such as order status or password resets. Extending it across an entire queue is an inference buyers make and vendors rarely correct.
What is AI-washing in outsourcing?
Describing a whole operation in the language of a narrow pilot. A provider licenses a platform, configures a handful of intents, then markets the capability as general. It is usually overstatement rather than fraud, and the contract still covers everything.
How can a buyer verify a provider’s AI claims quickly?
Ask which model and version runs, who tuned it, where answers are retrieved from, and who can halt a path overnight. Then request last month’s escalated cases and what changed. Neither can be staged.
