AI at Work: Building Toward 5× Output with Frontier AI

AI can help people spend less time preparing work and more time applying judgment. Used well, it can draft a first version, organize information, explain unfamiliar concepts, and turn repetitive steps into a repeatable process. Frontier AI—the most capable general-purpose models available for a task—can be especially useful when work involves several connected steps, complex instructions, or a large amount of context.

The opportunity is substantial, but a fivefold improvement is a goal to measure, not a promise about every job. Research on customer-support work found an average productivity improvement of about 14% with AI assistance. A Harvard Business School account of an experiment with consultants reported faster completion and better results for tasks within the model’s capabilities. Neither finding establishes that everyone will produce five times as much work.

Where AI helps

Start with work that has a clear input and an output you can check. AI can turn meeting notes into a draft action list, convert a rough outline into a document, compare options against stated criteria, or help identify patterns in a spreadsheet. It can also explain a technical idea at different levels so colleagues can make a decision with less back-and-forth.

The benefit grows when you connect these activities. A well-defined process can move from approved source material to a draft, a review checklist, a revised version, and several useful formats. Human review remains part of the process, particularly for factual claims, calculations, confidential information, and decisions that affect people.

What 5× could look like

Consider an illustrative weekly briefing that normally takes five hours. Suppose collecting approved inputs takes one hour, drafting takes two hours, formatting takes one hour, and review takes one hour. If a redesigned AI-assisted workflow completes the same acceptable deliverable in one hour, including review and corrections, that task has achieved a fivefold increase in output per hour: five hours divided by one hour equals five.

This is an example, not a measured result or a forecast. If checking and corrections take two additional hours, the improvement is much smaller. Producing five times as many unchecked drafts is not the same as producing five times as much useful work. Measure finished, accurate deliverables against the same quality standard.

Why model capability matters

A more capable model may follow complex requirements more reliably and help with difficult drafting, reasoning, or coding tasks. That can reduce the time spent breaking a task into tiny instructions or repairing an unusable first attempt. Capability alone does not guarantee a better result: the right information, clear constraints, appropriate tools, and review all matter. Use simpler tools when they do the job well, and evaluate more capable models on your actual work.

A practical starting point

Choose one recurring task and record its current time and quality. Give the AI the purpose, intended audience, approved source material, required format, and acceptance criteria. Review its output against those criteria. Record total time, including prompting, checking, and rework. Repeat the experiment before expanding it to other work.

Keep confidential material within tools your organization has approved. Ask for sources when facts matter, verify those sources, and check important calculations independently. The person responsible for the work should retain the final decision.

AI creates value when it helps people deliver more useful work with less avoidable effort. Aim high—including a 5× target for suitable tasks—but let measured quality and completed outcomes establish the result.

Sources

National Bureau of Economic Research, Generative AI at Work: https://www.nber.org/papers/w31161

Harvard Business School AI Institute, Navigating the Jagged Technological Frontier: https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/