InsightsPractice
PracticeP100People & GovernanceBy Fabrice Macarty · Founder, CEO

The return on AI depends on which tasks it is given

The same AI tool can raise quality or lower it, depending on the task it is given. Four moves give that allocation an owner, a measure, a budget and a quarterly review, so the return on AI stops falling between the technology budget and the people budget.

+40%
higher quality for consultants using AI on tasks inside its capability frontier; outside it, they were 19 percentage points less likely to reach a correct solution (Dell'Acqua et al., Organization Science, 2026; 758 BCG consultants).
13%
of AI users say they are rewarded for reinventing work with AI even when the results fall short (Microsoft Work Trend Index 2026, a survey of 20,000 AI users in ten countries).
AI & DigitalStrategy & GovernancePeople & Leadership
The decision stakes

From the tool
to the task.

As AI and agents take on more of the execution, the opportunity is to expand human agency and redesign how work gets done.Satya Nadella, Chairman and CEO, Microsoft, LinkedIn, 5 May 2026

The return on AI is set task by task: in the experiment, the same tool produced better or worse work than the control group, depending on the task. Allocation is a management decision, and it needs what any such decision needs: an owner, a measure, a budget and a review. The four moves below supply them.

Sightis reads allocation as a responsibility before it is a lever: the executive who answers for the combined output of people and AI is also the one who can show a client, an auditor or a regulator why a task went to the machine. That answer, backed by the scorecard and the review, is what can make the gain defensible.

The decision tool

Four moves
to own the allocation.

01

Task Map

For each core process, list the tasks, starting with those that take most of the hours. The experiment gives a starting hypothesis: inside the AI's frontier it found quality gains, outside it more errors (Dell'Acqua et al., 2026). Test where each process's own frontier lies on a sample of real work before scaling.

The map turns a technology choice into decisions a manager can sign and revisit. Without it, cost can decide the allocation alone: Klarna's chief executive said cost had been “a too predominant evaluation factor”.
02

Combined Scorecard

Give each process one line at executive committee level: the output of people and AI together, its quality (error rate, rework, complaints) and the named owner who answers for each AI-assisted result. Report the three together each quarter.

Adoption rates and headcount, reported separately, describe the inputs. The combined line describes what the work produces, and the named owner, with the scorecard in hand, is the person a client or a regulator can question.
03

One Envelope

Put the AI tools budget and the reskilling budget for each process in one envelope, under the same owner, and approve it against the combined scorecard line. Revisit the split between tools and people at each budget review.

When the tools sit in the technology budget and the training in the people budget, each side can report success while the process loses quality. One envelope can give one person the trade-off, if that person holds the budget authority.
04

Allocation Review

Each quarter, ask one question per process: which tasks did the AI improve, and which did it degrade? Move tasks back to people where quality fell most, extend the AI where it rose, and update the task map.

PwC's 2026 AI performance study finds its AI leaders more disciplined than peers about pruning initiatives, yet only 28% of them say they review their AI portfolio to terminate initiatives to a large or very large extent. A review that can take a task back keeps the allocation honest.
The proof

Klarna

Swedish payments and banking company; in February 2024 its AI assistant took over two-thirds of customer service chats in its first month, by the company's own count.

700

full-time agents' worth of work, by Klarna's estimate, done by the AI assistant in its first month, across 2.3 million conversations, with requests resolved in under 2 minutes against 11 before, Klarna stated on 27 February 2024.

Klarna's assistant, launched in February 2024, took on requests such as refunds and returns, its Form F-1 states. Its first-quarter 2025 results, published on 19 May 2025, report customer service cost per transaction down 40% since the first quarter of 2023, from $0.32 to $0.19, “with no drop in consumer satisfaction”, and credit the shift with “consistently high quality”. Eleven days earlier, Bloomberg had reported that Sebastian Siemiatkowski, co-founder and chief executive, was hiring people again for customer service; as CX Today reported on 12 May, he said that “cost unfortunately seems to have been a too predominant evaluation factor” and that “what you end up having is lower quality”. The assistant stayed: in an interview published by CMSWire on 20 May, he put its load at about 1.3 million requests a month, the work of about 800 people.

In the same month, Klarna's results described the quality of its customer service as consistently high, and its chief executive described it as lower. A combined scorecard puts both readings on one line before the allocation is revised.

The full case

PDF · 5 slides