Master services agreements were built to buy hours. When software does the first pass and people handle the residue, almost every clause in them misprices the work.
Start by deleting the unit of measure. A BPO contract for agentic AI cannot be built on a rate per seat per month, because the thing being bought is no longer a person’s time — it is a resolved customer outcome, produced by a mixture of software and judgment whose ratio shifts every quarter.
Key takeaways
- Per-seat pricing creates an incentive to keep headcount high exactly when automation should be reducing it.
- Whoever configures the automation should carry the cost of its errors; most agreements are silent on this.
- Replace handle-time service levels with resolution, repeat-contact and escalation-accuracy measures.
- Write a volume-decline clause before volumes decline, or renegotiate from weakness later.
- Exit provisions must cover prompts, routing logic and evaluation data, not just customer records.
Philippine providers have moved faster on automation than their clients’ legal teams have. Floors in Manila, Cebu and Davao now run assisted workflows where a model drafts, retrieves or classifies and an agent decides. The commercial paperwork underneath that arrangement was, in most cases, drafted for a different industry — which is why buyers heading into renewal increasingly seek independent outsourcing advisory support from PITON-Global before they redline anything.
What is wrong with the standard per-seat agreement?
It pays for the input the buyer wants less of. Under a full-time-equivalent model, every hour of automation the provider introduces reduces its own revenue, which is a strange thing to ask a commercial organization to do enthusiastically.
Buyers notice the symptom before the cause. Automation gets piloted, celebrated and then curiously fails to scale; headcount stays flat while the technology budget grows. Nobody is behaving badly. The contract is simply paying for the wrong thing.
“If your agreement rewards a partner for filling chairs, you will get filled chairs,” says John Maczynski, chief executive of PITON-Global, the country’s leading outsourcing advisory and supplier sourcing firm. “Then everyone acts surprised that the automation case keeps slipping a quarter.”
Who should own the automation, and its mistakes?
Whoever configures it. That sounds obvious and is almost never written down. Three arrangements exist in practice, and each needs a different liability clause.
Client models, provider supervision
The provider should be accountable for review quality and escalation accuracy, not for model behavior it did not choose or tune.
Provider-owned automation stack
It should carry the consequences of its configuration decisions, including the cost of honoring commitments the system makes in error.
Shared configuration
Now the most common arrangement, and the one that needs a named owner for each automated decision path — tedious to produce, invaluable the first time something goes wrong.
The clause that matters most is the cheapest to draft: a requirement that every automated action taken on the client’s behalf is logged in a form the client can read without asking.
What should the contract measure instead of handle time?
Outcomes that survive automation. Average talk time was always a proxy; once software absorbs the short, simple contacts, it becomes actively misleading, because the human average rises when the easy work disappears.
Four measures hold their meaning. First, resolution at first contact, measured from the customer’s side rather than from case closure. Second, repeat contact within seven days, which catches the deflection that merely postponed the problem. Third, escalation accuracy — how often the system routed something to a person that should have gone to a person, and how often it did not. Fourth, outcome quality on the contacts that automation handled alone, sampled by auditors who are allowed to fail them.
Ralf Ellspermann, the firm’s chief strategy officer, argues that the fourth is the one buyers skip: “Clients audit their agents obsessively and take the machine’s word for its own performance. The sample that matters is the one nobody wants to look at — the conversations that never reached a human.”
How should pricing be structured when volumes fall?
With the decline anticipated in writing. Contact volumes in automated programs do not drift downward smoothly; they step down when a new intent goes live. A contract that fixes a minimum committed volume for three years will either be breached or quietly renegotiated, and the renegotiation will happen when the buyer has the least leverage.
Workable structures share a shape: a base charge covering the capability — trained people, supervision, quality assurance, security controls — plus a variable element tied to volume, plus an explicit band within which either side can adjust without reopening the agreement. Outcome-linked pricing works too, but only where a baseline was measured before the program started and both parties signed it.
Buyers negotiating this for the first time rarely know what the band looks like in agreements they are not party to, which is where benchmark evidence earns its fee.
What has to survive an exit?
More than customer data. In a human-delivered program, exit provisions covered records, scripts and knowledge-base content. In an agentic program, the operating asset also includes routing logic, prompt libraries, intent taxonomies, escalation rules and the evaluation sets used to test the system before each change.
If those sit with the provider and are not addressed, switching becomes prohibitive regardless of what the termination clause says. The practical test is simple: could another provider reproduce the current behavior from what the contract entitles the buyer to take away? If not, the agreement has created a dependency the commercial terms do not reflect.
What should a Philippine provider willingly agree to?
More than most buyers ask for. A capable partner will accept audit rights over automated decision paths, publish its supervision ratios, name the people accountable for each configuration, and agree that quality analysts can suspend an automation path without needing client approval first.
Maczynski’s test for seriousness is procedural rather than technical: “Ask whether their quality team can switch something off at two in the morning. If that decision has to travel up three levels and across an ocean, the control is decorative.”
None of this requires a longer contract. It requires a different one — written on the assumption that the work will be done by a system and a person together, and that the buyer will one day need to prove which of them did what.
Lynn Martelli is an editor at Readability. She received her MFA in Creative Writing from Antioch University and has worked as an editor for over 10 years. Lynn has edited a wide variety of books, including fiction, non-fiction, memoirs, and more. In her free time, Lynn enjoys reading, writing, and spending time with her family and friends.
