The AI Tax: What Context Switching Between Clients Costs Freelancers

You finish the deliverable for the fintech client, close that thread, and open a new one for the lifestyle brand. First order of business: paste the brand voice doc, the product glossary, and the note about what the client decided last month on the naming question. The assistant reads it. Now it is useful again.

You did that yesterday too. You will do it tomorrow. That routine has a name: context switching. For freelancers using AI across multiple clients, it also has a price.

Fifty-four percent of freelancers now report advanced AI proficiency, compared to 38% of full-time employees. Freelancers are roughly 2.2 times more likely to use AI tools than traditional workers, according to Upwork's research on the AI-human work dynamic [4]. That figure gets cited as evidence of freelancer adaptability. It is that. But there is another way to read it: the cohort that adopted AI fastest is also the first to absorb a cost that nobody has put a name on yet.

The AI assistant is not slow or confused when you switch clients. It just does not know who that client is. Brand voice, technical stack, decisions already made, constraints that came out of last month's call: none of that travels with you when you open a new thread. You are the only bridge. And filling that gap is a manual task you perform multiple times a day, in small increments, without ever seeing it on an invoice.

That is the AI tax. This piece tries to name it properly.

The research on context switching already had a name for this

The productivity cost of switching between tasks is not new territory. Researchers were measuring it before AI assistants existed, and the numbers have been consistent enough that they show up in every workplace study worth citing.

Gloria Mark and colleagues at UC Irvine found that interrupted work gets completed faster, but at a measurable cost: more stress, more frustration, more effort [1]. The widely-reported figure of 23 minutes to fully resume an interrupted task traces to interviews with Mark rather than the paper text itself, so treat it as a reasonable estimate rather than a clinical finding. The underlying finding is still well-documented: switching interrupts not just the task but the cognitive state required to do it well.

A 2021 survey by Qatalog and the Ellis Idea Lab at Cornell found that on average, workers take about 9.5 minutes to get back into a productive workflow after switching between digital apps. Context switching hampered productivity for 45% of respondents [3].

A 2022 Harvard Business Review study put a different number on the same problem. Researchers tracked 20 teams across three Fortune 500 companies and found that workers toggled between applications roughly 1,200 times per day, spending just under four hours per week reorienting after those switches. That is about 9% of a working week, gone to the gap between one thing and the next [2].

Knowledge workers spend an average of 9.5 minutes reorienting after each app switch, adding up to roughly 4 hours per week (Qatalog/Cornell 2021; HBR 2022).

All of this research is about toggling between apps and meetings. It was written before AI assistants became a standard part of knowledge work. The switch this piece is naming sits a layer deeper. It happens inside a single tool, before the productive part of the session begins.

The new switch: re-establishing context inside the tool

An AI assistant has no durable memory between threads. When you start a new conversation, the assistant starts cold. Nothing from the previous session carries over. The client's brand guidelines, the decisions you made last week, the technical constraint that keeps coming up, the tone they prefer, none of that is there until you put it there.

Anthropic's own guidance on context engineering is direct on this point: context is a finite resource with diminishing returns. As token counts rise, recall becomes less reliable. The recommendation is to keep structured notes and memory outside the context window rather than expecting the window itself to preserve relevant state across sessions [5].

That guidance is written for developers building agent systems. But the principle lands equally for freelancers using AI day-to-day. The assistant does not hold onto your client context because the architecture does not work that way. The context window is not a vault. It is more like a whiteboard that gets erased when you walk out.

So when you move from one client to the next, you are not just switching tabs. You are rebuilding the entire semantic world the assistant needs to operate in. Different vocabulary. Different constraints. Different history. The fintech client and the lifestyle brand are not just different in tone. They are different worlds, and the assistant starts from the same blank slate each time.

This is not a failure in the tool. It is the structural cost of stateless AI sessions meeting multi-client work.

Freelancers carry the whole load

A salaried employee working on one product team holds one client context. They enter the same project world every morning. The assistant, if they use one, is at least operating in a roughly consistent environment.

A freelancer holds three to eight completely unrelated client contexts simultaneously. Different brands, different codebases or creative frameworks, different communication styles, different approval chains. The switching is not incidental to the job. It is the job.

Upwork's research puts 54% of freelancers at advanced AI proficiency, the highest adoption rate of any worker segment [4]. That means freelancers are already doing the most AI work and simultaneously carrying the most client contexts. The re-priming tax scales directly with client count. For freelancers, there is no way to reduce it by taking fewer clients without taking less work.

Consider a concrete scenario: a freelancer runs client calls for a fintech startup in the morning and works on brand copy for a consumer wellness company in the afternoon. Opening a new AI thread for each is not just a mental context switch. It means re-pasting the fintech client's regulatory constraints and terminology, then doing the same for the wellness brand's voice guide and product terminology, before the assistant becomes useful in either case. Neither thread knows what the other knows. Both start from zero.

Every new client thread starts cold. The assistant has no memory of previous sessions. The freelancer is the only bridge.

The friction is not in the AI. It is in the architecture of how context is stored, which right now means it lives in the freelancer's head and gets re-typed into a prompt box.

The quiet invoice

The switching-cost research gives enough of a baseline to put a rough number on this.

The Qatalog/Cornell study found that workers take about 9.5 minutes to reorient after switching between digital apps [3]. Applied to AI re-priming, that baseline is conservative (re-establishing full client context in a new thread involves more deliberate effort than clicking into a different app), but it is the grounded starting point.

A freelancer with five active clients who opens a new AI thread once per client per working day spends roughly 47 minutes before any productive AI output begins. Across a five-day week, that is about 3.9 hours: close to the four hours per week the HBR study found workers spending on app-toggling reorientation [2], except this cost lands entirely in the setup category rather than inside the work itself.

Two things make this estimate conservative. First, it counts only one re-prime per client per day. In practice, a mid-session context limit, a browser refresh, or a decision to start a clean thread mid-afternoon means additional re-primes. Second, it does not account for the cases where the context paste is incomplete, the assistant produces output that misses a constraint, and the freelancer has to correct and re-orient again.

The cost is invisible because it is spread across the day in small pieces. No single moment feels significant. Ten minutes here, five minutes there. But the aggregate is real, and it falls entirely outside any project tracker or time log because it happens before the billable work starts.

That is exactly the shape of a tax: routine, incremental, easy to miss, and never on any invoice you received.

Context that lives somewhere

The fix for re-explaining client context is not to explain it more efficiently. It is to stop re-explaining it at all.

The two approaches that can help are different in form but similar in structure. The first is building a personal context library: a set of files per client, kept outside any AI thread, that you paste at the start of each session. This is already what many freelancers do informally. The problem is that it is still a manual step, and the quality of the re-prime depends on remembering to do it and on the paste being current.

The second approach is moving client context into durable, re-readable state that an assistant can retrieve directly. When context lives somewhere the assistant can read on demand rather than somewhere you have to type it in, the re-priming step shrinks or disappears. Anthropic's guidance on context engineering describes exactly this: keeping structured notes and memory outside the context window is the recommended architecture for preserving state across sessions [5].

Agiflow is one example of that pattern in practice. It does not run AI agents or host the assistant. What it provides is structured project state (projects, tasks, notes, decisions, and artifacts) that an external assistant reads over MCP rather than being re-told each time [6]. If you want to understand how that pattern works at depth, the MCP project management tools post covers the mechanics.

The broader point applies regardless of specific tooling: context that lives in a board the assistant can re-read is durable in a way that context living in your head is not. The assistant does not have a memory problem you can fix by explaining better. It has a stateless architecture that you work around by giving it state to read.

For the related problem of context decay inside long coding sessions (a different failure mode, one session rather than multiple clients), the AI coding agents losing context post covers why that is also a memory architecture question rather than a model failure.

The tax you were never sent a bill for

There is a version of this that ends with a productivity tip about prompt templates and clipboard shortcuts. That is not the ending.

The more honest version is this: freelancers have adopted AI tools faster than any other segment of the workforce. They are also carrying more client complexity than any other segment. The re-priming cost is a structural consequence of that combination, not a user error.

Naming it matters. A cost you can describe, you can measure. A cost you can measure, you can decide whether to absorb or reduce. The minutes spent re-establishing client context are not wasted because you are bad at prompting. They are wasted because context lives in the wrong place.

That is worth knowing, and worth thinking about the next time you paste the same brand voice doc into a fresh thread for the third time this week.

If you want more writing like this delivered when it comes out, the subscribe link is below. And if this resonated with a freelancer you work with, it costs nothing to forward.

---

References

[1] Mark, Gudith & Klocke, "The Cost of Interrupted Work: More Speed and Stress," CHI 2008, UC Irvine. https://doi.org/10.1145/1357054.1357072. Relevant because it documents that interrupted work is completed faster but at measurably higher stress and effort cost; the widely-reported 23-minute figure traces to interviews with Mark, not the paper text.

[2] Murty, Dadlani & Das, "How Much Time and Energy Do We Waste Toggling Between Applications?" Harvard Business Review, 29 Aug 2022. https://hbr.org/2022/08/how-much-time-and-energy-do-we-waste-toggling-between-applications. Relevant because it quantifies toggling at ~1,200 times per day and nearly four hours per week spent reorienting after switches, across 20 teams at three Fortune 500 companies.

[3] Qatalog + Ellis Idea Lab, Cornell University, "Workgeist" study, 2021. https://www.ciodive.com/news/app-switching-enterprise-productivity-software-qatalog/602082/. Relevant because it finds workers take on average 9.5 minutes to return to productive workflow after switching digital apps, and that context switching hampered productivity for 45% of respondents.

[4] "Upwork Research Reveals New Insights Into the AI-Human Work Dynamic," Upwork, 9 Jul 2025. https://www.upwork.com/press/releases/upwork-research-reveals-new-insights-into-the-ai-human-work-dynamic. Relevant because it establishes that 54% of freelancers report advanced AI proficiency vs 38% of full-time employees, and that freelancers are roughly 2.2x more likely to use AI than traditional workers.

[5] "Effective Context Engineering for AI Agents," Anthropic Engineering. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents. Relevant because it describes context as a finite resource with diminishing returns, notes that recall degrades as token counts rise, and recommends keeping structured notes and memory outside the context window to preserve relevant state across sessions.

[6] Agiflow product-knowledge MCP (mcp-integration + project-management domains), internal. Relevant because it verifies that Agiflow does not run or host AI agents; it exposes projects, work units, tasks, artifacts, vault, and status as structured state that external assistants read and update over MCP.