Last updated: October 31, 2026
All Concepts

The Last Mile

Getting an agent was never the hard part. The hard part is that almost nobody has a mental model of what the thing they are asking actually does.

Alex Myers, The Armchair Futurist

Alex Myers. I build these systems with individuals and small teams, and this page is what I keep having to explain.

97%

of workers use AI poorly or not at all (Section, Jan 2026)

2%

of 4,500 logged work use cases were advanced (Section, Jan 2026)

33%

lose at least half of the time AI saves them to rework (Workday, Oct 2026)

32,813x

spread between the dearest and cheapest model per million tokens (BenchLM, Oct 2026)

I work with individuals and small teams, and the pattern is the same at every skill level. The tools are not the problem, and access is not the problem. What is missing is a mental model of what the machine in front of you actually does.

The gap is mental models

In January 2026 Section assessed working AI use and found 97% of people using it poorly or not at all. Of 4,500 work use cases they logged, 59% were basic and 2% advanced. A quarter of respondents had no work use case at all. Over the same period access to tools rose 16 to 17 percent, and training rose 2 percent.

Google and Ipsos measured the same gap from the other side a month later: 40% of employees use AI at work, 5% are fluent, and 14% were offered training. Workday found in October that only 38% of workers can explain what an AI agent is, and that a third of users lose at least half of the time AI saves them to rework. Rework is what a missing mental model costs, measured in hours.

This is not a prompting-skill problem. Researchers reviewed 102 commercial agents and ran a 31-person study of Operator and Manus, and reported capabilities misaligned with user mental models. You cannot delegate work to a system you cannot picture.

Six assumptions I keep correcting

Almost every failure I see in client work starts as one of these. None of them shows up in the interface, which is why rephrasing a request never fixes them.

Six common assumptions about AI tools, what is true in 2026, and what to do about each
What people assumeWhat is true in 2026What to do
The chat remembers everythingContext is finite and degrades as it fills. Across 18 models, accuracy fell as inputs grew, and needle-in-haystack tests overstate what long context really does.Restate your constraints near the end, and split work that will not fit.
One model for every jobPrices per million tokens span 32,813x. Frontier models rose 61% year on year while the mid-tier fell 41%.Default to a cheap model. Escalate only where the output must be right.
More thinking is always better16,000 reasoning tokens cost 32x more than 500. Stopping near 6,000 cut compute by half for about 6% accuracy loss.Match reasoning effort to the task. Sizing that is a skill, not a setting.
More agents, more throughputMulti-agent beat single-agent by 90.2% on research and used about 15x the tokens. Parallel writers make conflicting implicit choices.Parallelise research. Keep anything that writes in one thread.
Desktop and web chat are the same toolLocal steps, the built-in browser and the Agent SDK are desktop or CLI only. Web chat cannot see your files.Do the work in the harness that can touch the work. Web chat is for questions.
Guardrails are an enterprise concernOne GraphQL mutation deleted a production database and every backup in nine seconds. Only 5% of tech leaders are confident they can see what runs in production.Set permissions and sandboxing before you grant autonomy, not after.

What changes when the models are right

A consultant I work with spent three hours a month assembling board packs from all four of her spreadsheets. She asked an agent to summarise the sales data and flag risks, and got a tidy page that missed the two accounts which had gone quiet for six weeks. Her request described a task; her work needed a decision. When she added a line asking it to flag accounts whose order volume fell more than 30 percent against their own twelve-month average, ranked by revenue at risk, the same agent put both on top.

That exchange is the whole problem in miniature. She had to know what the output was for, and no model could supply that for her. The six assumptions above are the same thing at a different scale: each one is a fact about the tool you have to hold in your head before you can ask it for anything well.

For a one-person shop or a small team this is cheap to fix and expensive to skip. Literacy is not a training day, it is something you keep paying attention to, since the tools move under you. The people I see get value are not the ones with the best access. They are the ones who can say what the model is doing, roughly what it costs, and what it is allowed to touch.

Sources

  1. 1.The AI Proficiency Report, January 2026Section. 97% of workers use AI poorly or not at all; 25% have no work use case; of 4,500 use cases, 59% basic and 2% advanced; training received rose 2% while tool access rose 16-17%.
  2. 2.The Path to AI Fluency: AI Works for AmericaGoogle/Ipsos, February 2026, n=4,464 employees and 1,511 managers. 40% use AI at work, 5% are AI fluent, 14% were offered training.
  3. 3.AI@Work, October 2026Workday, about 6,000 workers, fieldwork September 2026. 38% can explain what an AI agent is; 33% lose at least half their time savings to rework; 41% conceal their AI use.
  4. 4.Context Rot: How Increasing Input Tokens Impacts LLM PerformanceChroma Research, July 2025. Eighteen models degrade as input length grows, non-uniformly; distractors matter and needle-in-haystack tests overstate long-context ability.
  5. 5.LLM pricing statisticsBenchLM, as of 29 October 2026. Median $0.75 input and $3.00 output per million tokens across 209 models; dearest API about 32,813x the cheapest.
  6. 6.Why Johnny Can’t Use AgentsACM Conference on AI and Agentic Systems, September 2025. Review of 102 commercial agents plus an N=31 study of Operator and Manus: agent capabilities were misaligned with user mental models.
  7. 7.Efficient reasoning and the cost of thinking longerApril 2026. 16,000 reasoning tokens cost 32x more than 500; stopping near 6,000 cuts compute about 50% for roughly 6% accuracy loss.
  8. 8.How we built our multi-agent research systemAnthropic Engineering, June 2025. Multi-agent beat single-agent by 90.2% on research tasks; token use explained 80% of the variance; about 15x the tokens of a chat.
  9. 9.Do not build multi-agentsCognition, June 2025. Parallel agents make implicit conflicting choices; context engineering beats prompt engineering, and writes stay single-threaded.
  10. 10.AI agent deleted a production database in 9 secondsZenity, April 2026. One GraphQL mutation wiped the volume and every backup; the most recent recoverable backup was three months old.
  11. 11.AI governance report 2026Retool, 307 CTOs, CIOs and CISOs. Only 5% are confident in their visibility into production AI tools; 93% worry about vibe-coded tools reaching production.
  12. 12.State of AI Architecture 2026Stacklok, 520 leaders surveyed September 2026. 81% agree frontier labs and hyperscalers make moving between vendors difficult; 84% say avoiding lock-in matters.
  13. 13.Permissions and sandboxingAnthropic. The documented permission and sandboxing model for a coding agent, which is the guardrail layer most individual users never configure.

Stuck at the last mile?

The gap closes faster with someone who has walked it. Book a free 15-minute call and we will find where you are stuck.

Book Free Strategy Call

Frequently asked questions

What is the last mile in AI adoption?

The last mile is the distance between an agent that can do almost anything and a person who has a mental model of what it actually does. Access stopped being the barrier in 2026. Section found 97% of workers still use AI poorly or not at all, and only 2% of 4,500 logged use cases were advanced. The gap is models of the tool, not access to it.

Why do capable people get so little from AI tools?

Because the interface is easy, people assume the work is easy. They do not know that context is finite, that models differ in price by more than 30,000 times, that reasoning effort has a cost curve, or that a desktop harness can touch files a web chat cannot see. So they rewrite the same request, get plausible but wrong output, and conclude the tool does not work. Workday found a third of users lose at least half their time savings to rework.

What mental models do I need to use agents well?

Six cover most of it: context is finite and degrades as it fills; models are priced and capable across a huge range, so routing matters; reasoning effort costs, and sizing it is a skill; parallel agents multiply tokens and conflict, so writes stay single-threaded; desktop harnesses do things web chat cannot; and permissions come before autonomy. None of these are programming skills. They are literacy, and they are trainable in weeks, not years.