On August 6, Genpact put two growth rates beside each other. Advanced Technology Solutions revenue had increased 24.1% from a year earlier. Core Business Services had increased 1.9%.

Advanced Technology Solutions reached $363 million in the second quarter. Core Business Services reached $980 million. Genpact expected at least 25% growth from the faster line in the following quarter while core services would be flat to slightly down.

The figures came from an earnings release, not a product launch. They record sales inside one company, not pure AI revenue. Genpact includes data, technology, and AI work in the faster segment; the release does not isolate what an agentic system earned.

A week earlier, TP had offered a different measure. More than 1,100 AI projects were running on its platform at June 30, 2.3 times the year-earlier count. Concentrix had used another unit in its June results: iX Suite deal count had increased 400% from a year earlier.

Revenue, projects, and deals can all rise during the same market transition. They cannot answer for one another. Small discovery work and a multiyear transformation each become a deal. Some projects never leave a pilot. Recognized revenue can come from implementation, data work, conventional operations, or software wrapped around human delivery. None of the three numbers tells a buyer how much accepted work reached production, how much the full service cost changed, or what happened to the people doing it.

Outsourcing makes those units unusually consequential. The industry transfers operating work, then staffs it at scale. TP says it has nearly 490,000 employees. Genpact’s careers site says its workforce exceeds 140,000. A provider can win technology work while a labor-intensive core grows slowly. A client can save touches in one process while paying for new integration, review, and exception handling elsewhere.

Inside the contract, the financial measure meets the workforce measure. If an accounts-payable agent handles routine matches, the service buyer needs to know whether a signed project became a live workflow, whether accepted invoices increased, whether error and escalation rates held, how the invoice changed, and whether processors moved into paid exception work. Calling all of that AI growth leaves too much room between the sales deck and the shift schedule.

BK Kalra separates the business

Genpact chief executive Balkrishan “BK” Kalra did not report one blended AI percentage on August 6. The company separated a faster technology portfolio from a much larger operating-services base.

Total second-quarter revenue was $1.343 billion, up 7.1% from a year earlier. Advanced Technology Solutions accounted for 27% of that total. Core Business Services accounted for 73%. The first segment grew more than twelve times as fast as the second, but the second remained almost three times its size.

Two reporting segments reveal more than a general claim that AI demand is strong. The split shows where growth entered Genpact’s accounts and how much of the company still sat outside that line. It also prevents a common reading error. The 24.1% rate applies to a defined segment, not to all Genpact revenue or an industry-wide AI market.

Below revenue, the pattern becomes less tidy. Gross margin was 36.5%, and adjusted operating margin was 17.4%. Cash from operations was $72 million, down from $177 million. Segment growth, consolidated margin, and cash conversion describe different parts of the business. No single AI return can be inferred from them.

Suppose a buyer has moved invoice intake, matching, and supplier-email drafting to a Genpact-led program. Some of the provider’s fee may sit in Advanced Technology Solutions during design and implementation. Recurring delivery may sit in core services. A model or workflow platform may be licensed separately. Human review may be bundled into a transaction price. The public segment split cannot reconstruct that contract.

It can change the questions at renewal. Finance can ask which paid work belongs to the new technology line and which remains conventional delivery. Operations can ask which workflows are live rather than configured. Procurement can ask whether a reduction in routine effort lowered the unit price, raised volume capacity, funded more specialist work, or widened the provider’s margin.

A divergent forecast makes those questions more urgent. Genpact expected Advanced Technology Solutions to grow at least 25% in the third quarter while core services were flat to slightly down. Its full-year outlook also called for at least 25% growth in the faster segment. A revenue mix can move because one line expands, because another stalls, or because work migrates between them.

Migration has two meanings here: accounting decides where the provider classifies revenue, while operations decides where the task goes. A buyer needs evidence for both. Otherwise a manual service can be relabeled as AI-enabled before its workload, price, or staffing has materially changed.

Genpact’s 27% technology share also limits an easy displacement story. Nearly three quarters of quarterly revenue still came from core services. The release did not publish segment headcount, hours, or job movement. It is therefore impossible to calculate how the two growth rates changed employment from the earnings figures alone.

Missing employment detail is not proof that nothing changed. It marks a reporting boundary. Investor accounts show how a company earned revenue. A workforce transition requires different records: roles before and after deployment, paid learning time, internal moves, vacancies removed, new review work, and exits. The contract sits close enough to both records that a customer can request them without pretending either is the other.

TP counts projects while Concentrix counts deals

TP’s first-half release arrived on July 30 with a less comfortable top line. Group revenue was EUR4.883 billion. It declined 1.7% on a like-for-like basis and 4.5% as reported. Core Services declined 1.3% like for like to EUR4.220 billion, while the company said AI-powered solutions grew at a strong double-digit rate.

Then came the 1,100. TP said more than 1,100 AI projects were running on its platform at the end of June, 2.3 times the number a year earlier. “Running” is stronger language than an announced partnership and weaker than a disclosed financial result. The release did not attach project revenue, production volume, customer savings, or employment effects to the count.

One project could cover a language model used by thousands of agents while another is a narrow proof of concept for one queue. Separate phases of the same client program might also be counted separately. Without a denominator for value or workload, adding projects measures activity, not scale.

TP gave one additional commercial clue. Outcome-based work in sales, collections, retention, and back-office solutions grew at a mid-single-digit rate and represented more than 15% of group revenue. The company plans to use AI within those services. That does not reveal what share of the 15% was AI-enabled, but it shows why pricing matters.

Another line in the release complicates the assumption that AI growth means fewer human services. TP said Data Services for AI grew at a strong double-digit rate, driven by demand for human-in-the-loop data work used in model development. People can lose routine process work while other people label, test, and maintain the systems taking it over. TP did not disclose that line’s revenue base, headcount, contract terms, or job quality.

Traditional outsourcing often bills for people, seats, hours, or staffed capacity. Automation can reduce the billable input that supported the old contract. Transaction or outcome pricing offers a different bargain: the buyer pays for a collected account, retained customer, processed claim, or resolved case. The provider can keep part of a productivity gain if quality holds and the outcome is real.

Outcome pricing creates a measurement burden of its own. A collections result can improve because of customer mix, timing, incentives, policy, or human negotiation as well as AI. A retention program can shift difficult cases to senior workers while a system handles easy ones. If the contract pays only for the final count, both parties still need to see quality, customer harm, escalation, and the work pushed outside the measured path.

TP describes itself on its investor site as a company of nearly 490,000 employees working across almost 100 countries. Project growth inside a workforce of that size can mean new engineering and integration work, different tasks for customer-service employees, lower demand for some queues, or all three at once. The public project count does not choose among those paths.

Concentrix disclosed yet another signal. Its second-quarter results put revenue at $2.463 billion, up 1.9% as reported and 0.6% in constant currency. Operating margin was 3.9%, compared with 6.1% a year earlier. Non-GAAP operating margin moved from 12.6% to 11.9%.

Concentrix also said iX Suite deal count grew 400% year over year. That percentage sounds larger than Genpact’s 24.1% and TP’s 2.3 times. It is not a larger version of the same fact. Concentrix counted deals. Genpact measured segment revenue. TP counted projects.

Deal count needs at least three qualifiers before it enters a budget decision. The base count matters: growth from five deals to 25 and growth from 50 to 250 both equal 400%. Deal size matters: a short assessment and a multiyear operating contract each add one. Timing matters: a signed deal can contribute little recognized revenue during the current quarter.

Concentrix did not disclose the count’s base, aggregate contract value, production use, or contribution to margin. Its lower consolidated margin also cannot be assigned to AI from the release. Acquisition effects, mix, investment, pricing, restructuring, currency, and delivery costs can move a group margin. The right conclusion is narrower: commercial interest in iX Suite rose sharply on the company’s chosen count while overall revenue grew modestly.

Put the three disclosures on one page and the industry looks busy. Leave their labels attached and the uncertainty becomes visible. A buyer can see technology revenue growing at one provider, platform projects multiplying at another, and deal flow accelerating at a third. The next line in the comparison should not be a synthetic AI growth rate. It should be the evidence each provider owes before activity becomes value.

A project count is not a revenue line

An outsourced process moves through several states before anyone can claim a durable result. Sales signs a deal. A delivery team opens a project. Engineers connect systems and prepare data. A workflow enters production. Employees and agents complete work. The buyer accepts the output. Finance recognizes revenue and cost. Only after a stable comparison can either side estimate savings or return.

Public disclosures often stop at different points in that sequence. Concentrix’s deal count sits near the commercial start. TP’s running projects sit somewhere between implementation and operation, with no published distribution. Genpact’s segment revenue sits in the accounts, but covers a broader portfolio than AI and says nothing about the customer’s accepted output.

Moving from one state to the next can fail. A signed project can wait for data access. A live workflow can route most cases back to people. Faster handling can produce more complaints. A lower provider invoice can be offset by internal review and integration costs. A successful pilot can remain too small to change a budget.

Worker evidence reaches a similar boundary. The International Labour Organization’s June review of empirical GenAI evidence found real but uneven productivity gains, many of them still unverified at scale. Worker-reported time savings of a few percent of hours had not yet become clear gains in measured output, earnings, or employment across the studies it reviewed.

An outsourcing contract can close part of that gap because the buyer and provider already exchange operational data. Start with the actual unit of work: an accurate invoice processed by deadline, a customer issue resolved without repeat contact, a valid claim completed, or a receivable collected within policy. Count attempts separately from accepted outcomes.

Then preserve the baseline. The month before launch may have a different volume, customer mix, backlog, or error rate from the month after. A provider should show the comparison period, case complexity, excluded work, and any policy changes. A percentage without that context can make a quiet queue look like an automated one.

Full cost belongs beside output. Model fees, platform charges, integration, data preparation, security review, employee time, retries, quality checks, escalation, and repair all enter the service economics. If the provider absorbs some costs, its margin may move even when the buyer’s invoice does not. If the buyer absorbs them, a lower contract price may still produce a higher total cost.

Human work is easy to lose in this accounting. An agent may save three minutes drafting a response and spend two minutes correcting unsupported claims. A team leader may review ten more exceptions per hour. A buyer’s employee may reconcile cases the provider marked complete. Those minutes belong to the workflow, even when they sit outside the platform log.

Manager incentives can hide the minutes twice. A delivery manager measured on touchless rate may discourage agents from reopening weak cases. A client manager measured on savings may treat internal repair as business as usual. Reviewers then learn which failures are safe to report and which will make the program look expensive. Reliable measurement needs an error channel that does not punish the person who found the error.

Pricing decides who receives the gain. Under a staffed-capacity contract, fewer people can reduce provider revenue unless the parties renegotiate. Under transaction pricing, the provider may earn the same fee with less labor. Under gain sharing, both parties need an agreed baseline and an auditable definition of savings. Under outcome pricing, quality and customer protection need guardrails so the provider cannot optimize the paid result by discarding hard cases.

No model is automatically fair. A per-person contract can protect employment while rewarding unnecessary effort. A transaction price can fund investment or conceal work intensification. An outcome fee can encourage better service or narrow the definition of success. The measurement has to expose the behavior the price creates.

For a renewal, finance should be able to trace one number through the sequence. Take 100,000 monthly invoices. How many entered the agentic path? How many required a human touch? How many passed the buyer’s accuracy and timeliness standard? How many were reopened? What did the provider and buyer spend in total? Which staffing, overtime, vacancy, or specialist-budget decision followed?

Following the trace stops revenue, project, and deal claims from drifting into one another. Deals show demand; running projects show activity; revenue records a sale. Accepted output at full cost is operating evidence. A sustained change in price, margin, pay, hours, or staffing comes later.

From transaction processor to exception manager

Genpact’s product material describes a specific destination for human work. On its Agentic Operations page, agents execute high-volume, data-intensive tasks while people supervise results and handle exceptions. The company says transaction processors can move toward work as supervisors, exception managers, decision makers, and value creators.

Genpact has described a role design, not an employment result. The page also says operations can scale without headcount growing at the same rate. For a buyer, the promise is attractive: more volume, more continuous operation, and less dependence on matching every increase with another hire. For an employee, the missing detail is whether the new role exists, how many seats it has, and how a person reaches it.

Exception work is not simply routine work with a better title. The remaining cases are often ambiguous, urgent, emotionally difficult, or financially risky. A processor who once handled a broad mix may spend a future shift on disputes, missing data, policy conflicts, and agent failures. Average handling time can rise even as total touches fall.

That rise can be a sign of successful routing rather than poorer performance. Once simple cases leave the queue, the remaining average describes harder work. Keeping the old handling-time target would push exception managers to clear cases faster just as each case demands more judgment. Managers need a new workload baseline, a severity mix, and a limit on concurrent reviews.

Supervision also concentrates accountability. An employee may review more machine-generated actions than one person previously produced. The job needs authority to stop a workflow, access to the evidence behind a recommendation, and enough time to investigate. If those conditions are absent, “exception manager” can mean carrying the liability for a system someone else configured.

Genpact has published training and mobility signals. During its 2025 investor day, the company said employees completed 11 million learning hours in 2024, an average of 82 hours per person. It divided its workforce strategy into AI builders and AI practitioners and said 60% of open roles in Advanced Technology Solutions were filled internally.

Those are substantial company claims, but their denominators differ. Learning hours cover an entire workforce and many subjects. The internal-fill rate covers open roles in one segment, not every employee whose task changed. Neither number reveals how many transaction processors entered exception roles, what they earned afterward, or whether learning happened during paid time.

A sharper internal example came from the same investor day. Genpact described an HR organization with 20 AI builders and 170 AI practitioners. It said HR headcount had fallen 10% and set a target of more than 25% by the end of 2026. This was a company-reported case and a forward target from June 2025, not proof that the target was reached.

Taken together, the examples resist a comforting story. Genpact can invest millions of learning hours, fill technology roles from within, redesign a function, and pursue lower headcount at the same time. Whether training worked depends on who reached a paid destination, not on whether headcount rose or fell in aggregate.

For an employee, that path starts before the old work disappears. The company identifies which routine tasks will shrink, which exception tasks will grow, and how many roles will be available. It gives practice on live but controlled cases. A manager observes the work. The employee has a paid chance to qualify. Pay, schedule, location, and employment status are attached to the destination.

Workers also need a route to challenge the map. A processor may know that a supposedly routine match involves a supplier exception the project data missed. A reviewer may see that one queue now carries the emotional load of angry customers. Bringing those observations into design changes the workload estimate before staffing and price harden into a multiyear contract.

Aggregate learning hours can hide failure at each step. The course may arrive after a role is removed, or its certificate may not satisfy the hiring manager. An available job could require a different city or night work. A better-paid specialist opening might exist for one person out of twenty. Internal mobility needs a cohort denominator: of the workers materially affected by the workflow, how many trained, qualified, applied, moved, stayed, or exited?

Junior development creates another problem. Routine processing supplied repetition, error feedback, and customer context. If an agent takes the straightforward cases, a new employee may meet only the hardest ones. The service provider needs a practice environment where a person can review normal cases, compare a decision with policy, and earn responsibility before carrying production exceptions.

Service buyers have leverage over this design. A contract specifies volumes, response times, controls, and location. It can also require a role-transition report for the work it changes. Such a request does not give the client control over every employment decision. It makes the labor consequence of its purchased operating model visible.

The 1.8 million workers beyond one company

In the Philippines, a company-level role claim becomes an economic question. The OECD’s 2026 country survey estimates that information technology and business process management employs about 1.8 million people. That equals 3.7% of national employment. Sector revenue amounts to roughly 8% of gross domestic product.

About 72% of IT-BPM employment sits in business process outsourcing, including contact centers and back-office operations. The sector has also moved into global capability centers, health information management, software, and other higher-value services. Both facts matter. Routine work remains large, and upgrading has already begun.

OECD analysts call AI the industry’s most urgent structural and technological risk. They also warn that the Philippines could lose an early advantage if workforce capability does not move quickly enough toward AI-intensive delivery. That national warning does not predict a layoff total. It identifies a competitive squeeze: global buyers want lower cost and more advanced work while other countries compete for the same contracts.

An ILO study published in February puts a boundary around displacement claims. More than one quarter of Philippine employment, about 12.7 million jobs, has some exposure to generative AI under its occupation-based index. Only 3.6% of jobs fall in the highest exposure category with elevated displacement risk.

Exposure is potential task contact, not observed job loss. The ILO expects transformation to matter more than outright automation across the labor market. Its distribution findings are less reassuring. Women face about twice the exposure rate of men because of occupational concentration. Young people are more likely to work in service, sales, and clerical-support occupations that fall into higher-risk categories.

Aggregate employment should not erase those group differences. A provider can report a stable total while entry hiring falls, one city loses routine work, or women remain concentrated in roles with shrinking task volume. Cohort reporting by location, job family, level, and gender can reveal the change without publishing individual records. Worker representatives need the same figures before expansion, not months after a sourcing decision.

This distribution meets the service industry’s promotion ladder. A young worker may enter through a voice or transaction role, learn customer context, and move into quality, training, workforce management, analytics, or operations leadership. If entry tasks contract before an alternative practice route exists, the industry can preserve senior expertise for a time while weakening its future supply.

A national reskilling announcement cannot complete that route. The OECD lists programs from the Technical Education and Skills Development Authority, the Department of Information and Communications Technology, and industry partnerships. Short courses and credentials can expand access. They do not identify which employer will recognize a credential, release an employee to study, or offer a paid destination.

Service buyers influence the demand side from thousands of miles away. When a bank changes a contract from staffed capacity to automated transactions, it may remove positions, create exception work, or shift volume to a more capable center. When a retailer accepts a lower price without asking about training or quality load, the provider decides how to absorb the cut. Procurement becomes part of labor-market policy even when nobody in the sourcing meeting uses that phrase.

Across countries, the June ILO evidence review cautions against treating saved time as an automatic social dividend. Time savings had not consistently appeared as greater measured output, earnings, or employment. The review also found risks around autonomy, coordination, job quality, inequality, and opportunities for younger workers.

Outsourcing makes those effects easier to displace from view. The client sees a service level and an invoice. The provider sees staffing, occupancy, attrition, and margin. The worker sees a queue. A national agency sees sector employment months later. Each record is valid, but the connection can disappear across organizational and national borders.

A role-transition clause will not solve the Philippines’ skills challenge. It can make one buyer accountable for one changed workflow. The provider reports the affected cohort, paid training, assessment, new roles, pay bands, moves, vacancies, and exits. The buyer reviews the report beside quality and savings before expanding the model to another queue.

The ILO study calls for social dialogue alongside targeted regional and sector policy. In a contract, that can be modest and concrete: the provider explains a planned task change to affected teams, records objections about workload and customer risk, and shows the buyer how those objections changed the deployment. Consultation is evidence only when it can alter a decision.

Scale makes small rates matter. If a program changes work for 500 employees, reporting only the 30 people who entered new specialist roles creates a success story with no denominator. Report all 500. Some may keep redesigned jobs, some may transfer, some may leave voluntarily, and some positions may end. Naming each path is more useful than describing the workforce as broadly upskilled.

Back at the finance operation

Return to a service buyer preparing its 2027 renewal. The operation processes one million supplier invoices a month across several countries. Three hundred and twenty provider employees handle intake, matching, exceptions, supplier messages, and close support. The numbers are hypothetical, but the meeting is ordinary.

Its provider arrives with 14 AI projects, a growing suite of signed deals, and a proposal to move more work into an agentic operation. It forecasts a lower cost per invoice and says the team can absorb volume growth without proportional hiring. Finance likes the unit-price curve. Procurement wants a multiyear commitment. Operations remembers that the last pilot routed unusual tax cases back to people.

Comparing public companies will not settle the renewal. The provider has to follow this contract from sale to work.

Contract measurement sheet

RecordMinimum evidence in the renewal
Commercial startSignature date, contract value, term, and the work included in the AI-enabled scope
DeploymentProduction date, systems connected, monthly eligible volume, and the share actually routed through the workflow
Accepted workCompleted units that met accuracy, timeliness, policy, and customer standards, with reopen and repair rates
Full costProvider fee, software, integration, data, security, buyer labor, human review, retries, escalation, and remediation
Price and valueBaseline unit cost, current unit cost, recognized savings, volume effects, and who receives any gain
Human workloadRoutine touches removed, review and exception minutes added, overtime, queue intensity, and decision authority
Worker transitionAffected cohort, paid learning hours, assessment, roles opened, internal moves, pay change, vacancies removed, and exits

Three substitutions would invalidate the sheet. Signed demand does not equal production. Production does not equal accepted value. A training total does not equal a completed worker transition.

For the million-invoice operation, the first renewal month might show that 760,000 invoices were eligible for the new path and 610,000 entered it. Of those, 515,000 could complete without a manual touch, 83,000 could receive a review, and 12,000 could return to conventional processing. The buyer would still need accuracy, late-payment, duplicate-payment, supplier-contact, and reopen results before accepting the work.

These example counts illustrate a measurement design, not a performance forecast. Their purpose is to keep the denominator visible. “Eighty-four percent touchless” means little if the program silently excluded the hardest quarter of the workload or if buyers’ employees repaired completed cases outside the provider system.

Cost follows the same discipline. Compare the old provider invoice with the new one, then add the buyer’s integration and oversight cost. Separate one-time implementation from recurring operation. Record model and platform changes that alter price during the term. If the parties claim savings, specify whether they are cash released, avoided hiring, reduced overtime, or capacity used for more volume.

Then open the staffing decision. Perhaps 70 routine roles are expected to shrink while 24 exception and control roles grow. The proposal should say when, where, and at what pay. It should show how many affected workers can enter training, how long practice takes, what happens if they do not qualify, and whether other openings exist. A plan for 24 destinations is not a transition plan for 70 people.

No client needs employee names or control over personnel decisions. Cohort data can protect privacy while revealing outcomes. The contract can require quarterly counts and a review before another workflow expands. It can also protect paid practice time, require a manageable reviewer ratio, and prohibit savings calculations that omit the buyer’s new human work.

Price can reinforce the same design. A base transaction fee can cover accepted routine work. A separate fee can recognize complex exceptions instead of encouraging the provider to avoid them. Gain sharing can use verified net savings after full cost. Quality holdbacks can cover serious errors and customer harm. A transition budget can be released against completed practice and role moves rather than course attendance alone.

None of these terms guarantees a good result. They make a disputed result inspectable. The provider can show that its investment produced reliable work. The buyer can verify the bill. Employees and workforce planners can see whether new roles matched the work removed. Both sides can correct a program before a project count becomes a larger rollout.

Genpact’s 24.1% segment growth, TP’s 1,100 running projects, and Concentrix’s 400% deal-count increase each describe real company activity within the limits of the disclosure. They should remain three measures.

At renewal, one sequence should control the decision: a signed scope became a live workflow, the workflow produced accepted work at full cost, the commercial gain reached an invoice or operating budget, and the people affected had a recorded outcome. That is AI growth an operator can buy. A missing line is a specific piece of the contract to fix before the next queue moves.