Agent assist in 2026: what the record shows when AI helps the human

Every contact-center vendor now sells a copilot for the agent. The pitch is always the same number: lower average handle time. The problem for a buyer is that the vendor runs the pilot, picks the queue and reports the figure.

So the useful question is not whether agent assist works. It is what the public record says it does when the AI helps a person instead of replacing one, and how that compares with the replace path.

What agent assist is

Agent assist is AI that works beside a human support agent, drafting replies, summarising calls and suggesting next steps, while the human decides what the customer actually receives.

The difference from an autonomous agent is who sends. In agent assist the human is the last step on every interaction. In full automation the AI is.

How does agent assist work in practice

Agent assist works by putting a draft in front of a person who edits it or sends it. The best documented case is Octopus Energy’s Magic Ink, which summarises interactions and drafts replies “as if written by the user”, with a human reviewing and sending each one [source].

The volume is real. Kraken’s own case study reports it had summarised 6,239,087 calls and generated 9,415,901 messages, with no as-of date stated [source].

The more telling line is quieter. Kraken says “around a third of messages generated by Magic Ink require zero to minimal changes before they are sent” [source]. Read the other way, most drafts get edited. That is the human doing the job agent assist exists to keep them doing.

What does the record show agent assist achieves

The record shows gains that are real, uneven and smaller than a handle-time slide implies. The strongest controlled evidence is not from a contact center at all: it is three field experiments on AI coding assistance, an assist tool by design.

Pooled across 4,867 developers, the authors found “a 26.08% increase (SE: 10.3%) in completed tasks” [source]. The gain was concentrated: recent hires and junior staff “increased their output by 27% to 39%”, while “more senior developers saw productivity gains of 8% to 13%” ([source]).

The widest lens is smaller still. A Danish study surveyed 25,000 workers in 11 exposed occupations, customer support specialists among them [source]. Adopters “report savings of about 3% of their work hours”, a self-report rather than a register measure ([source]).

The registers did not move. The authors estimate “precise null effects on earnings and recorded hours”, ruling out effects larger than 2% two years after ChatGPT [source]. And “most chatbot users (85%) report reallocating time savings” to other tasks ([source]).

Agent assist vs replacing the agent

The obvious reading is that agent assist is the timid version of automation, and that the bold operators skip it. The record argues the opposite: assist is the version that survived contact with its own numbers.

Klarna took the replace path. Its assistant handled two-thirds of chats in its first month, which Forbes relayed as “the workload equivalent to 700 full-time staff members, according to a recent statement by the company” [source]. By 2025 its CEO said the cost-first push meant “what you end up having is lower quality”, and Klarna began recruiting humans again ([source]).

Octopus took the assist path and said the tool was “unlikely to lead to job cuts” [source]. Its figures are company-originated too, and its satisfaction number moves between statements, from 80% against 65% in one to “around 70%” in its own case study ([source], [source]). Neither company is a clean measurement. The difference is which one had to walk its model back.

What TIN checked

The four records behind this post are TIN case files, each checked against the public record before it went live. Each page states its own limits, and they matter here.

• Octopus Energy’s Magic Ink: CEO and company figures, relayed by independent press, with no third-party audit.
• Three Copilot field experiments: randomized and peer-reviewed, but two of six authors work for the vendor, and the workers are developers, not support agents.
• The Denmark chatbot study: an NBER working paper, not peer-reviewed; time savings are self-reported.
• Klarna’s assistant and walk-back: every 2024 efficiency figure originates with Klarna.

None of the four measures average handle time for an agent assist deployment. That gap is the finding a buyer should carry into a vendor meeting. For the replace side of the same market, see what the record shows on AI voice agents.

How the records compare

The Octopus share comes from its CEO’s May 2023 statement, relayed by City AM as “the work of 250 people” [source], with 34% of queries carried in the Octopus case file.

How to read an agent assist pilot

A real time agent assist pilot is worth reading only if it says who was in the treated group. The controlled record shows the gain depends on tenure, so a pilot run on new hires reports the junior band of that range, not the senior one [source].

Ask for the edit rate as well as the handle time. If most drafts need more than minimal changes, as Kraken’s own figure implies for its tool, the human is still doing the judgment and the saving sits mostly in typing [source].

Bottom line

Agent assist is usually bought as a stepping stone to automation. The record says to buy it as the destination: the replace path is the one that produced the public walk-back, while the assist path produced modest gains concentrated where training is thinnest. Expect a seasoned floor to look like the Denmark self-reports, and new hires to look like the junior developers. The principle travels: when a tool’s benefit depends on who uses it, the pilot’s headcount mix is the result, and the headline number is just its shadow.