When should AI escalate? A routing card for cost, risk and quality

Using the most expensive model at every step is not a safety strategy. It usually means nobody has decided when the workflow genuinely needs more capability.

The opposite can be just as expensive. A low-cost default that retries the same task three times, misses exceptions, and leaves a person to clean up the result may cost more than the stronger model would have.

What is missing is AI routing: explicit rules that send each step to a low-cost default, a stronger model, human approval, or a hard stop. The route should follow risk and accepted quality, not whichever model topped this week's leaderboard.

Token price is only part of the bill

OpenAI describes three GPT-5.6 tiers with different capability and price profiles: Luna for cost-sensitive volume, Terra as a balanced tier, and Sol for the hardest work. xAI positions Grok 4.5 for coding and agentic knowledge work. Perplexity has demonstrated a different pattern in its Computer environment: a base model that can hand selected work to a stronger advisor route.

Source: OpenAI – GPT-5.6

Source: SpaceXAI – Introducing Grok 4.5

Source: Perplexity – GLM 5.2 advisor preview

Vendor tests can show what is possible, but they cannot choose your route. Your documents, failure costs, response times, and review habits will not match their benchmark.

A better measure is cost per accepted outcome. Count model and tool usage, attempts, elapsed time, and human review until the output is genuinely usable. OpenAI makes the same basic argument in its guidance on AI investment: evaluate useful work per dollar and measure the full cost of reaching a defined quality bar.

Source: OpenAI – How to manage AI investments in the agentic era

Four routes are enough for a first pilot

A routing policy does not need to begin as a sophisticated platform. Four clear exits are enough.

1. Default route: the cheapest option that passes

Send common, reversible steps to the simplest route that reaches your quality bar. Good candidates include classification, field extraction, formatting, and first drafts.

Default must not mean “let everything through.” Each output should face a small acceptance test, such as:

• all required fields are present
• the total matches the source data
• each claim points to an approved source
• the answer uses the required language and format
• prohibited personal data is absent from the output

If the test passes, the job moves on. If it fails, the workflow should not keep retrying the same instruction.

2. Escalation route: more capability on a named signal

A stronger model should be called by a rule, not a hunch. Useful escalation signals can be observed and recorded:

• two sources disagree
• a required value is missing
• classification confidence falls below the pilot threshold
• the task needs several dependent reasoning steps
• the default route has failed once
• an unusual case matches no approved template

Put a ceiling on this route too. If the stronger model misses the quality bar after one attempt, send the case to a person or stop it. Do not let it enter an expensive loop.

3. Human route: somebody must own the decision

Some steps should not be solved by buying more tokens. A person should approve actions that are hard to reverse, make a promise to a customer, or carry financial, legal, or security consequences.

Examples include:

• sending a contract or a live customer response
• approving a refund
• changing permissions
• publishing externally
• posting an accounting entry or making a payment
• acting on sensitive student, patient, or employee data

Give the reviewer a short decision packet: what the AI proposes, which sources it used, which tests already passed, and what triggered review. A bare “approve” button is not a control.

4. Hard stop: the job must not continue

Stop the workflow when it lacks the conditions for a safe decision. For example:

• a source cannot be reached
• the user or agent lacks permission
• personal data appears where it cannot be processed
• sources conflict on a high-impact question
• the run budget has been exhausted
• the same failure keeps returning
• no reviewer is available before the deadline

Write down what happens next. Who receives the alert? Where does the case land? How does the team complete it manually? “The agent stopped” is a status message, not a fallback.

Fill in the routing card before choosing a model

Take one recurring job and complete these fields. Keep the card to one page.

• Accepted outcome: What must be true before the output may be used?
• Default route: Which low-cost route handles normal cases?
• Escalation signal: Which observable condition calls a stronger model?
• Human gate: Which decisions need a named role?
• Hard stop: Which conditions end the run?
• Prohibited actions: What may the AI never do, regardless of model?
• Evidence: Which sources, test results, and decisions must be saved?
• Budget: How many attempts and escalations may one case consume?
• Fallback: How is the work completed when the AI route is unavailable?

If you cannot write a precise escalation signal, you do not have a routing rule yet. Human review is the more honest default.

Example: a supplier variance report

Suppose finance needs a weekly report on supplier invoices that differ from contract terms or previous levels.

Default route: The AI reads approved invoice and contract fields, calculates the variance, and drafts a report. It cannot change the source system. A deterministic check confirms that the invoice amount, supplier, and contract reference are present.

Escalation route: A stronger model is used when the contract has several price conditions, the invoice lacks a clear reference, or two documents appear to conflict. It gets one attempt and must cite the passages supporting its conclusion.

Human route: A finance owner approves all proposed actions above a set amount and every case where the AI recommends contacting the supplier.

Hard stop: The job ends if the contract is missing, the document contains a prohibited data type, or the amounts cannot be reconciled with the source. The case goes into the normal manual queue.

Now the route can be tested. It does not depend on everyone agreeing that one model is “smart.”

Measure whether routing works

Track a small set of decision-ready measures:

• First-pass acceptance: What share passes without another run?
• Escalation rate: What share of work leaves the default route?
• Useful escalation: How often does the stronger model lift the result to the accepted level?
• Human rework: How many minutes are needed after the AI output?
• Cost per accepted outcome: Model, tools, and review combined.
• False pass: How many bad outputs clear the acceptance test?
• Time to owner: How quickly does a stopped case reach the right person?

A low escalation rate is not automatically good. It may mean the default works well, or that the rules allow too much through. Read it alongside actual quality.

Test historical cases before going live

Start with 20–50 completed cases where you already know the accepted outcome. Run the routing policy without letting it send, publish, post, or change anything.

Review the failures, not just the average. Which cases should have escalated? Where did an expensive route fail to improve the result? Note the stop rules that were missing too.

Change one rule at a time. Once the route survives the historical test, run a time-boxed pilot in shadow mode or with mandatory approval. Set an end date and a maximum budget before the first live run.

That is roughly where model selection becomes interesting. Not before.

If the routing card exposes unclear quality bars, permissions, or ownership, the next step is not another model comparison. Map the workflow in Tool Forge and build a small pilot where every escalation can be explained.