Why AI Assistant Costs Run Away, and Five Practices That Control Them

A pattern repeats in organisations six months into AI adoption. Teams that were productive with an AI assistant start hitting usage limits mid-task. Work stalls. Someone proposes upgrading the plan, and the same thing happens again a few weeks later on the larger allowance.

The instinct is that the team is using the tool too much. Usually they are using it inefficiently, which is a different problem with a much cheaper fix.

You are paying for context, not for questions

This is the mechanic almost nobody explains, and it changes how you think about everything else.

When you send a message to an AI assistant, the model does not receive only that message. It receives the entire accumulated conversation, plus any instruction files, plus the definitions of every tool it has access to. That whole payload is reprocessed on every single turn.

Ask ten short questions in one long session and you are not paying for ten short questions. You are paying for the first question, then the first two, then the first three, and so on. Consumption compounds with conversation length regardless of how brief your individual messages are.

Once that is clear, the practices below stop looking like arbitrary settings and start looking like the obvious consequences.

Five practices that control it

1. Audit what is connected

Every integration an assistant can reach carries a description of what it does and how to call it. Historically all of those descriptions loaded into context at the start of every session, whether or not the tools were used. Connect a dozen systems and a meaningful share of the working context was consumed before anyone typed anything.

This has improved. Modern implementations defer most of that detail and load it only when a tool is actually needed. But the principle still holds: connections are not free, and most people accumulate them without ever removing one.

The practical version is a periodic audit. List what is connected, note when each was last genuinely used, and disconnect anything that is not earning its place on the current project. This takes minutes and is the only item on this list that requires no behaviour change afterwards.

2. Keep the instruction file short

Most teams give their assistant a standing instructions file covering conventions, context and preferences. That file is read on every message, so its length is a recurring cost rather than a one-off.

The common failure is treating it as documentation. It fills with background, examples and explanation, and eventually it is long enough to measurably reduce the working space available for the actual task.

Write it as an index, not a manual. Point to where information lives rather than reproducing it. A reference to a file costs a line; pasting the file costs its full length on every turn for the rest of the project.

A useful discipline is a hard ceiling, somewhere around 200 lines. Anything that does not survive that limit probably belongs in a linked document.

3. Agree the task before work begins

This is the largest source of waste and the least discussed, because it does not look like a settings problem.

The expensive failure is not a long conversation. It is a long conversation spent confidently solving a problem nobody asked about. Every wrong assumption produces work that is then discarded, and the discarded work stays in context, still being reprocessed on every subsequent turn. You pay for it twice: once to produce it, then repeatedly to carry it.

The fix is a standing instruction that the assistant should confirm its understanding before acting on anything ambiguous, and ask rather than assume. Something like: do not make changes until you are confident about the task; ask follow-up questions whenever more clarity would help.

This feels slower. It is substantially cheaper, and it produces fewer revisions.

4. Match the model to the task

Model families differ in capability and cost by a wide margin, and the most capable option is not the right default for everything.

The rough allocation that works: the strongest reasoning models for planning, architecture and genuinely hard problems; mid-tier models for the bulk of execution work; the lightest models for narrow, well-defined, repetitive subtasks. Running everything on the most capable tier is the equivalent of sending your most senior person to every meeting.

Worth noting that model names and tiers change frequently. The principle is durable; the specific mapping needs revisiting every few months.

5. Reset context deliberately

Because consumption compounds with conversation length, a long session eventually spends most of its budget reprocessing history that is no longer relevant.

Most tools offer a way to compact or summarise a conversation, condensing what has happened into a shorter summary and continuing from there. The mistake is waiting until you are forced to do it, by which point you have already paid for a great deal of redundant reprocessing and quality has usually started to drift.

Check consumption periodically rather than reactively, and reset at around two thirds full rather than at the ceiling. Starting a genuinely fresh session when the topic changes is even better, and most people resist it out of a vague sense that continuity is valuable when it usually is not.

What this means at an organisational level

Individually these are habits. At organisational scale they are governance, and the distinction matters for anyone responsible for an AI adoption programme.

If a team is hitting limits and the response is to buy a larger allowance, the underlying inefficiency is now funded rather than fixed, and it will scale with headcount. The same team on the same practices will hit the larger ceiling too.

The useful diagnostic question is not how much the tools cost. It is whether anyone has ever examined how they are being used. In most organisations the answer is no, because AI tool usage sits outside the normal procurement and review cycles that would otherwise catch this.

This is more acute in the UAE than in most markets. According to Microsoft’s 2025 Work Trend Index, 70.1% of UAE workplaces reported active AI tool use by early 2026, against a global average of 17.8%. High adoption is a genuine advantage, but it also means more teams here are further along the consumption curve, and more of them are hitting limits before anyone has established how the tools should be used. The organisations that adopted earliest are the ones most likely to be funding inefficiency at scale right now, and they are also the ones with the most to gain from a fortnight of attention to it.

This is the same pattern that shows up everywhere else in AI adoption. The technology performs as specified. The returns disappoint because the workflow around it was never designed, and adding budget to an undesigned workflow produces more of the same output rather than better output.

Frequently asked questions

Why does a short question still consume a lot of usage?
Because the model reprocesses the entire conversation history, instruction files and available tool definitions on every turn, not just your latest message. Consumption therefore compounds with session length rather than tracking the size of individual messages.

Does disconnecting integrations reduce consumption?
It helps, though less than it used to. Tool definitions historically loaded in full at the start of every session; most modern implementations now defer that detail until a tool is actually needed. Auditing connections is still worthwhile, because they accumulate and are rarely reviewed.

What is the single highest-impact change?
Agreeing the task before work starts. Effort spent solving the wrong problem is wasted twice: once producing output nobody wanted, and again reprocessing that output in context for the remainder of the session. It is also the change that most improves quality rather than only cost.

Should teams just buy a larger plan instead?
Only after examining usage. Increasing the allowance for an inefficient workflow funds the inefficiency and scales it with headcount. The practices above cost nothing and usually resolve the problem without a change in spend.

Further reading

Related: AI Adoption in the UAE covers the organisational framework these practices sit inside. Enterprise AI Adoption covers governance at scale. Why Your Marketing Isn’t Working covers why automation amplifies an undesigned process rather than fixing it. AI Adoption for UAE SMEs covers the smaller-organisation starting point.

Martin Alva

Brand, Digital & AI Transformation Leader, Dubai UAE

Senior Manager at Space42 (A G42 & Mubadala Company). 20+ years of brand strategy, digital transformation, and AI adoption across MENA and Europe. 5 MENA Effie Awards. 500+ campaigns across the region.

Connect on LinkedIn →