Which AI Model Should Your Alberta Team Use for Which Task?

Last updated: September 2026

Alberta teams routinely default every task to their platform's top-tier model, from drafting a two-line Slack reply to reviewing a 40-page contract, and that single habit is the biggest driver of AI token burn we see in client usage audits. Splitting complex work into a planning phase on the strongest model and an execution phase on a lighter one has cut token usage by roughly half on the workflows Solway has tested it against — including one accounting reconciliation task that dropped from hours to about 40 minutes. A drafting task and a planning task don't need the same amount of reasoning, and matching model choice, and reasoning effort, to what the task actually requires can cut usage substantially without touching output quality. Here's the framework Solway uses with Calgary and Edmonton clients to make that decision automatic instead of ad hoc.

Why Model Choice Matters More Than Most Teams Realize

Every major AI platform now offers a range of models, from fast and inexpensive to slow and highly capable, plus a separate "reasoning effort" dial that controls how much a model deliberates before answering. Most individual users never touch either setting. They stick with whatever model loads by default and leave reasoning effort wherever the platform sets it. For a single person doing occasional work, that's a minor inefficiency. For a team of twenty running that pattern eight hours a day, it's the difference between a usage allowance that lasts the week and one that runs out by Thursday, forcing a scramble for extra capacity mid-project.

Anthropic's own guidance for admins managing team-wide Claude usage names this directly: model choice has a direct and significant impact on spend, and effort level is the second-biggest lever after it. The fix isn't buying more capacity. It's giving your team a simple rule for which model fits which task.

This isn't unique to Claude. Every major platform bundles a range of models and effort settings, and the pattern of leaving both on their loosest default while quietly hitting weekly caps shows up across ChatGPT, Gemini, and Copilot deployments too. The fix is the same everywhere: someone has to decide, once, which task types warrant the expensive setting and which don't — and then make that decision the default, not a judgment call left to each employee in the moment.

Solway's Plan-Then-Execute Model Split

The single highest-leverage pattern we've implemented with clients is splitting complex work into two phases and assigning a different model to each:

• Planning phase: Use the platform's most capable model to think through structure, catch edge cases, and produce a clear plan or outline. This is where deep reasoning earns its cost.

• Execution phase: Hand the actual drafting, formatting, or step-by-step output to a faster, lighter model working from that plan.

Shaheer Tariq, Co-founder of Solway, used exactly this pattern to run the firm's own monthly accounting reconciliation: feeding bank statements, proposals, and invoices into Claude Cowork, using the platform's strongest model for the first two planning steps, then switching to a lighter model to execute the remaining steps. The full reconciliation, including a finished financial dashboard, took about 40 minutes for a task that previously took hours of manual work.

We've seen the same split solve a real usage problem for a Calgary property development team. Their heaviest users process blueprints, floor plans, and site photos daily and were defaulting to the top-tier model for every step of that review. Splitting the work so a stronger model plans the review approach and a lighter model executes the page-by-page pass cut their usage on that workflow substantially, without a drop in review quality. The pattern generalizes into a simple rule of thumb, laid out below.

A Simple Task-to-Model Guide

That last row matters as much as the first three. If a task is the same every time, the answer isn't picking the right model. It's not asking a model to reason through it at all — a saved Skill or a small script produces the same output every time without spending a single token on reasoning.

Can You Enforce This Across a Team, or Does It Rely on Individual Habits?

Per Anthropic's own plan documentation, team admins on Claude's Team and Enterprise plans can set the default model that new conversations start with for the whole organization, removing the guesswork for anyone who wouldn't otherwise think to choose. That single admin setting, combined with a short workshop on the plan-then-execute split, does more to control usage than any seat upgrade.

The same logic applies on ChatGPT, where OpenAI's own product documentation separates model selection and reasoning effort as distinct, user-facing controls, and ChatGPT Enterprise admins can set group-level and individual usage limits to enforce similar discipline. The platforms differ in interface, not in the underlying economics: a heavier model or higher effort level costs more compute regardless of which company built it.

Is a Smarter Model Always the Better Choice?

No, and this is where most teams overcorrect once they learn about the cost difference. A top-tier model with max reasoning effort will out-think a lighter model on a genuinely hard problem. It will not out-draft it on a routine task, and the token cost difference between the two can be substantial for high-volume routine work. The goal isn't to avoid the capable model. It's to reserve it for the fraction of tasks that actually need it — the planning step, the edge case, the genuinely ambiguous request — and let a lighter model carry everything else.

This Is a Training Problem, Not a Technology Problem

None of this requires new software or a bigger contract. It requires a team that knows the difference between a planning task and an execution task, and a five-minute habit of choosing the right tool for each. That's the gap a half-day Solway Workshop closes, and in Alberta, it's a gap CAPG will help fund: per the Government of Alberta's current CAPG program guidelines, training on model selection and workflow efficiency falls under the Digital and Technological skills category, reimbursed at 50% up to $5,000 per employee per year.

FAQ

Which model should my team default to for everyday work?

For routine drafting, summarizing, and quick questions, the fastest available model at low-to-medium reasoning effort is usually sufficient. Reserve the top-tier model and higher effort settings for genuinely complex, multi-step work.

What does "reasoning effort" mean, and why should Alberta businesses care about it?

It's a setting that controls how much a model deliberates before responding — low, medium, high, or max. Higher effort produces more thorough answers but consumes meaningfully more of your team's usage allowance, so matching it to task complexity is one of the fastest ways to control cost.

Is a smarter or more expensive model always the better choice?

No. It's the better choice for hard, ambiguous, or high-stakes tasks. For routine work, it produces the same practical output as a lighter model at a higher token cost.

Can I set a default model for my whole team instead of relying on individual habits?

Yes. Team and Enterprise plan admins can set the model that new conversations start with for everyone in the organization, which removes the decision for anyone who wouldn't otherwise think to make it.

Does this same logic apply to ChatGPT, or is model routing a Claude-specific idea?

It applies across platforms. ChatGPT also separates model choice and reasoning depth as user-facing settings; the interface differs, but a heavier model or higher effort level costs more regardless of provider.

How much can model routing actually reduce usage?

It varies by workflow, but in our client work, splitting a complex task into a planning phase (strong model) and an execution phase (lighter model) has been the single highest-leverage change we've implemented, often doing more than any other single adjustment — cutting usage on token-heavy workflows by roughly half in the cases we've measured, without a corresponding drop in output quality.

Is training on model selection worth it, or can employees figure this out themselves?

Left alone, most employees default to whatever loads first and never revisit it. A short workshop that gives the team a simple rule of thumb, plus an admin-level default, closes that gap in a single session rather than relying on habits forming on their own.

Does CAPG cover training on model selection and AI workflow efficiency?

Yes. It falls under CAPG's Digital and Technological skills category, reimbursed at 50% of eligible training costs up to $5,000 per employee per year, with no stated minimum-hours requirement under the current program guidelines, and 75% (up to $10,000 per trainee) if the training is for a newly hired, previously unemployed Albertan.