If Everyone Works Faster with AI, Why Hasn’t the Organisation Become Faster?

Geoffrey Chen

After AI Enters the Workflow · Season Two: “From Personal Tool to Organisational Capability” · Article 9

After a team adopts AI, email drafting, meeting summaries and first versions of reports become visibly faster. Employees feel that they complete more small tasks each day, and management can count an increase in generated output. Customer waiting time does not fall. Projects do not finish earlier. The approval queue grows.

Organisational speed is not the sum of individual speed. Work passes through handovers, waiting, decisions and rework. If AI accelerates one stage while the next stage retains fixed capacity, the result is faster delivery of work into a bottleneck.

Local productivity and organisational throughput are different measures

An employee who prepares a draft in two hours instead of four has doubled local efficiency. If the document waits three days for one specialist to approve it, the customer experiences almost no improvement.

The organisation normally cares about an end-to-end result: how long from incoming need to an accepted outcome; how much work remains in progress; how much returns because of error; and how much eventually creates value for a customer or the institution.

Words, prompts and first drafts are easy to count. They may be inventory rather than completion. The more unverified and undecided material produced, the larger the queue.

AI moves bottlenecks from production towards coordination

When drafting becomes cheap, people request extra analysis and several alternative versions. Professional review, legal checking, management approval and release become scarce.

Coordination cost rises as well. More drafts require comparison, naming, version management and meetings. AI reduces the time to create a first option and may increase the time needed to decide among options.

A field experiment across 66 firms and 7,137 knowledge workers found that users of a generative AI tool later spent about two fewer hours each week on email and worked less outside ordinary hours. Yet individual provision did not produce a detected change in task quantity or composition. NBER, “Shifting Work Patterns with Generative AI”

Individual time gains can be real without automatically reorganising the firm.

Existing metrics absorb the saved time

If performance continues to be measured through message volume, report volume and response speed, employees will use AI to improve those indicators. They may receive no space to reduce meetings, improve source records or repair the process.

Saved time can also be absorbed by new demands. An NBER study of AI exposure and work time found, under its data and modelling, an association between higher exposure, longer working days and reduced leisure. The authors considered how gains are distributed through organisations and markets. NBER, “AI and the Extended Workday”

This is not a law that every AI user will work longer. It shows that technology does not decide who receives the saving. The organisation must state whether the benefit will reduce burden, improve quality, expand output or create a new service.

General technologies need complementary intangible investment

Research on the “Productivity J-Curve” argues that general-purpose technologies require co-invention in processes, products, business models and human capital. Those intangible investments may obscure early productivity gains before their later benefits appear. NBER, “The Productivity J-Curve”

For AI, complementary investment includes knowledge curation, evaluation, role redesign, approval reform, training and new measures. A licence can be bought quickly; these structures change slowly.

If the organisation will not bear that cost, it can only place AI on top of the old workflow. Employees produce faster while the old institution continues to wait.

Map the flow to locate actual waiting time

Choose one use case and record every stage from request to completion: active work, queues, handovers, rework and approval. A common discovery is that production of prose occupies only a small part of total elapsed time.

Then decide how the saving should change the system. Remove handovers, combine duplicate approvals, create a fast path for low-risk cases, automate objective checks and prevent extra output with no defined purpose.

Do not merely give more AI to the bottleneck role. If the bottleneck comes from concentrated authority, conflicting objectives or untrustworthy information, faster recommendations will not solve it.

OECD analysis of generative AI productivity similarly emphasises the need for firms to adapt organisations, processes and strategies in order to obtain the potential gain. OECD, “The effects of generative AI on productivity, innovation and entrepreneurship”

Limit work in progress instead of rewarding unlimited generation

When review and approval capacity is finite, teams should limit the number of AI-produced items in progress. Complete and adopt existing work before generating more candidates.

Entry rules can help. Every analysis must identify a decision-maker and deadline. Every draft must state its sources and intended use. When review capacity is exhausted, low-priority bulk generation should wait.

Limiting output sounds contrary to AI efficiency. It protects end-to-end throughput. The organisation does not need to maximise generation; it needs to maximise outcomes that pass every necessary stage.

Redesigning decision rights may matter more than faster generation

Some waiting comes not from insufficient capability but from poorly designed authority. If one manager must approve every low-risk item, that manager becomes a permanent bottleneck. AI can prepare material faster but cannot by itself alter delegation.

The organisation should distinguish decisions that must remain central from those that can be delegated under explicit conditions. Low-value, reversible and rule-bound work may use pre-authorisation and later sampling; exceptions and consequential work retain escalation. Management must accept that effective control can come from boundaries and monitoring rather than a signature on every item.

It should also ask whether AI has created new “decision requests”. If the system continuously generates options and every option requires a managerial choice, small matters once resolved by staff become a new approval load. Good design reduces unnecessary choices as well as automating their production.

End-to-end service measures should include elapsed time from request to adoption, the number waiting beyond a threshold, reasons for rework and customer outcomes. Local savings become organisational performance only when these measures improve.

The organisation should also trace where the saved time goes. It may be reinvested in stronger review, earlier customer response and exceptions that genuinely require judgement. It may instead become more meetings, more drafts and a higher expectation of immediate availability. The first pattern increases institutional capacity; the second merely increases work density. An AI benefit review should therefore record more than estimated minutes saved per task. It should examine adoption, waiting, rework, extension of the working day, and whether employees can convert the released time into work of greater value. Otherwise, a local efficiency claim may conceal a system that is busier but delivers no earlier result.

Conclusion: redesign flow, not only individual actions

Individual speed matters. It may reduce overtime and improve experience. For the organisation to become faster, however, it must change how work moves among roles.

The final rule is:

AI creates organisational productivity only when it shortens end-to-end completion, reduces rework or increases accepted outcomes. If it only increases intermediate artefacts, speed becomes backlog.

Management should look beyond how much employees generate and ask what customers and the institution receive earlier. The strongest organisational use of AI may not make every stage faster. It may remove unnecessary stages and prevent the decisions that matter from spending most of their life in a queue.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.