Umar M. Sharif.
Services2026Research synthesis + strategyProposal
An HR outsourcing and staffing group · Leadership briefing

Where the AI gain lands: exposure, evidence and pricing for a services firm

A research synthesis and strategy proposal on how much AI helps, which roles it touches, and who keeps the saving inside headcount-priced contracts.

At a glance
−19% to +35%
measured AI effect range across 7 studies
16%
of HR functions have an ROI metric for AI
26% → 41%
tool adoption after a 9½-minute training (randomised, 164 participants)
13
role families mapped for exposure
01 · The problem

Automation removes billable hours before it removes cost.

The leadership question was where AI value would land inside the business. The evidence on AI at work is wide and not like-for-like, measurement inside HR functions is thin, and part of the workforce sits on client sites under headcount pricing. The briefing synthesises 7 studies, maps exposure across 13 role families and proposes how to price and sequence the work.

Thesis

Automation removes billable hours before it removes cost. Inside headcount-priced contracts, the saving lands with the client.

02 · The analysis

The range is the finding: licences alone can land anywhere on this line.

← Slower / worse Faster / better →
−19% slowerRCT, 16 developers, 246 tasks, tool access only
−19%
−19% less likely correctfield experiment, 758 consultants, tasks outside the model's strengths
−19%
~0 no task-mix shiftRCT, 7,137 workers, 66 firms, 6 months
~0
+12.2% more tasksfield experiment, 758 consultants
+12.2%
+13.8% issues resolved per hourstudy of ~5,000 support agents, in-workflow assist
+13.8%
+26.1% more tasksRCTs across 4,867 developers at three large firms
+26.1%
+35% issues resolved per hoursame support deployment, least experienced agents
+35%
−10% 0 +10% +20% +30%
Source: 7 peer-reviewed or preregistered studies, 2023–2026; not like-for-like, so no median is shown.

Training lifted adoption from 26% to 41%; access without training showed no significant improvement.

50% 25%
26%
41%
Before training
After 9½-minute training
+0.27 grade improvement vs untrained access Access without training: no significant improvement
Source: Randomised study, 164 participants, anonymised.

4 in 10 HR functions use AI, yet only 16% have an ROI metric for it.

of HR functions use AI39%
do not formally measure AI investment success56%
have an AI ROI metric16%
Source: Survey of HR functions, anonymised.
03 · The framework

Role exposure inventory: who works where, and who keeps the gain.

Role familyWhere they workExposureWho keeps the gain
Own payrollGain kept by: the firm
Payroll officerOwn payrollHIGHThe firm
HR operations administratorOwn payrollHIGHThe firm
Billing and invoicing analystOwn payrollHIGHThe firm
Recruitment coordinatorOwn payrollHIGHThe firm
Recruitment consultantOwn payrollPARTIALThe firm
Compliance and statutory officerOwn payrollPARTIALThe firm
HR consultantOwn payrollPARTIALThe firm
Product and platform engineerOwn payrollPARTIALThe firm
Executive search consultantOwn payrollLOWThe firm
Client relationship managerOwn payrollLOWThe firm
Client sitesGain kept by: the client
Back-office processing associateClient sitesHIGHThe client (under headcount pricing)
Site supervisor and coordinatorClient sitesPARTIALThe client (under headcount pricing)
Field, facilities and support staffClient sitesLOWThe client (under headcount pricing)
Exposure ratings are analyst judgement, to be validated with the firm's own role and volume data.

AI work comes in 3 tiers, from assisted work up to agentic systems.

T1Assisted work
T2Building
T3Agentic systems
Pricing recommendation

Price AI work as a hybrid: fixed fee up front, a recurring charge, and a performance component only where it can be measured.

01 · FIXEDDiagnostic + implementation fee
02 · RECURRINGPlatform / usage charge
03 · CONDITIONALPerformance componentOnly where baselines, data access and attribution are agreed in advance.
Churn-model design · proposal

Build the churn models in order of value at risk.

1Deployed-associate attrition30/60/90-day exit risk; survival model + gradient-boosted classifier.
2Client contract non-renewalScored monthly, ranked by probability × contract value.
3Placement early exitMatch-quality scoring at offer stage.
4Internal delivery-staff churnDeferred: label volume is the constraint.
Design rules
ASet the lead time first.
BUse PR-AUC, not accuracy.
CExplain scores per person.
99%
Where 1% churn, predicting nobody churns scores 99% accuracy.
04 · What it proposes

A 60-day design phase with a gate before further spend.

A fixed-scope design phase produces an AI strategy and a 12-month plan. The gate sits before any further spend, and the unit measured at the gate is a workflow changed and measured, with no credit for people trained.

DECISION 1

Run a 60-day fixed-scope design phase that ends in an AI strategy and a 12-month plan.

DECISION 2

Gate further spend on one unit: a workflow changed and measured.

DECISION 3

Price with the hybrid model, with a performance component only where baselines, data access and attribution are agreed first.

DECISION 4

Validate the exposure ratings with the firm's own role and volume data, then build churn models in order of value at risk.

This page describes a proposal. No outcome is claimed for it.
05 · Method notes
01The seven studies are third-party research; the synthesis and framing are mine.
02The studies differ in design, population and task, so they are not like-for-like and no median is shown.
03Exposure ratings are analyst judgement, to be validated with the firm's own role and volume data.
04Ratings, rankings and the build order are proposals, not findings.
05Public research is described by study type only.
Next case
The activation bottleneck →
Finding the bottleneck in a subscription funnel.
Book a 30-min AI Audit Back to all work