17 min read

The Hidden Risks of Building AI Agents on One Frontier Model

By submitting, you consent to our use of your data. Privacy Policy.

Category

AI Agents

Share the article

Four provider decisions that never show up in an accuracy chart, what each one costs on a modelled invoice workflow, the four controls that remove most of it, and how Beam builds for a model that will change.

📌 Key takeaways

  • Four provider decisions put about $188,000 a year at risk on a modelled invoice workflow whose accuracy never moved: a retired model, a launch discount that ends, a closed account, and an agent action nobody asked for.

  • 71% of that exposure is the rare out-of-scope action. Price, the part everyone negotiates, is 10%.

  • Four controls cut the exposure to about $54,000 a year, including $11,600 to run them. A gate on the action does most of the work; a kept test set and a pre-approved second model do most of the rest.

  • The research lines up: tool-using agents recognise a risky action in over 98% of direct questions and still take it most of the time, and a separate checker on the action cut risky executions by 55% (Tang et al., 2025).

  • Beam designs every workflow on the assumption that the model under a step will change: criteria on each step, and Consent required on the step that sends, pays or files.

When an AI agent goes live, teams watch one number: accuracy. At Beam we publish 98% across more than 10 million agent tasks. In my experience, for extraction, classification and checks against a written policy, accuracy is rarely what holds a deployment back anymore.

Over the last year and a bit, the events that actually hurt our deployments never moved that number. A provider retired the model a live client engagement ran on. A launch discount on a new cloud model ended after about seven weeks. Two frontier labs cut off our access, a year apart. And outside Beam, a large bank blocked AI shopping agents after one bought three items when it was asked for one. So instead of asking which model is most accurate, we priced what each of those events costs and what should be in place before they arrive. My answer: model risk is counterparty risk. Your provider is a supplier whose decisions you do not control, so treat it like one. Price what each decision costs you, then cap the exposure.

This piece covers four risks that come from a provider's decisions rather than from the model's accuracy: retirement, price, cutoff and out-of-scope actions. Rate limits and outages have a known fix in how the agent is built and are covered in our harness engineering post. Choosing between frontier and open models is in Open Source vs Frontier Models for Enterprise AI Agents. Every number is either published, with a source, or modelled from assumptions stated in the text. Figures as of 15 November 2026.

An accuracy chart that stays flat while four provider events happen underneath it: a model retired, a launch discount ending, an account closed, and an out-of-scope action. None of them moves the accuracy line.

The four events that hurt our deployments never moved the accuracy number.

Why doesn't your accuracy chart warn you when a provider changes the deal?

Because accuracy measures the model on your tasks, and these risks come from the provider's decisions about the model. You tested a model, pinned that exact version to a step of the workflow and built the business case on it. Then one of four things happens:

Risk

What the provider does

What you see

Retirement

Removes the model you qualified

A migration and retest on their schedule

Price

Ends a launch discount

A bill that no longer matches the business case

Cutoff

Closes your account

Every step pinned to that lab stops

Out-of-scope action

Ships a model that does more than asked

One action nobody requested, with a real cost

None of these lowers accuracy. A retired model was accurate until it vanished. A closed account gives no wrong answers, only no answers. An agent that buys three items instead of one still completed the purchase you asked for.

To price them, I modelled a workflow our customers run often: accounts payable at a mid-size company. It handles 20,000 invoices a month in six steps: extract, code to the ledger, match against the purchase order, route exceptions, draft the vendor reply, then pay or send. By hand an invoice takes 6 minutes of a specialist at $60 an hour. With the agent, the model costs about 4 cents an invoice and a person still handles 5% of them. That saves about $1.36 million a year (modelled), and that is what the four risks put at stake.

I compared the workflow pinned to one frontier model with no controls against the same workflow with the four controls in the third section.

How much do the four risks cost one workflow in a year?

About $188,000 a year in expected cost for the pinned workflow, and almost three quarters of it comes from one line (modelled):

Risk, expected cost per year

Pinned, no controls

With controls

Out-of-scope action

$134,256

$18,226

Cutoff

$19,851

$1,758

Price after the launch discount

$19,200

$19,200

Retirement

$14,720

$2,720

Running the controls

$0

$11,600

Total

$188,027

$53,504

Expected cost per year of each provider risk for one invoice workflow at 20,000 invoices a month, pinned to one model with no controls versus with four controls. Out-of-scope action: $134,256 versus $18,226. Cutoff: $19,851 versus $1,758. Price after the launch discount: $19,200 either way. Retirement: $14,720 versus $2,720. Totals $188,027 and $53,504.

71% of the exposure is the rare action, not the price (modelled).

Each line is a probability times a cost per event. Here is where each comes from.

Retirement: you migrate on their schedule

On one live client engagement we had stayed on a flagship model for a long time, because accuracy was already where it needed to be. The provider retired it, and we had to migrate. The client had not asked for a new model, and the retest was ours to pay for.

I do not think labs retire models to push buyers up a price list. A new model arrives about every three months, and each old one they keep serving costs engineering time. The motive is ordinary. You still inherit the migration.

What the migration costs depends on what you kept from the first qualification. Without a maintained test set, re-qualifying six steps takes about three engineer days each, a specialist relabelling 500 test cases, and two weeks of running old and new side by side: about $18,400. With a test set and a second model already tested, it is half a day per step and a week of spot checks: about $3,400. At one forced migration roughly every 15 months, that is $14,720 a year against $2,720 (modelled).

A quieter version is a model that changes under the same name. Researchers at Stanford and Berkeley found the June 2023 version of GPT-4 identified prime numbers correctly 51% of the time, against 84% in March (Chen, Zaharia and Zou, 2023). A test set you rerun each quarter catches that too.

Price: the launch number is not the operating number

On a long horizon I expect intelligence to keep getting cheaper. In the short window, prices can still jump. A new model from a major cloud provider launched at a discount, and about seven weeks later the discount ended.

A 2026 index built from 21,000 posted API prices shows how carefully this has to be read. The price of a given level of capability falls fast, by the index's main estimate about 40% a year in log terms. Per completed task, the same paper says the data do not establish a quality-adjusted decline at all (Zhu, 2026). My own read of why is that reasoning models raised token use per task faster than token prices fell.

So I modelled both: a 50% launch discount that ends, then 1.5 times more tokens per task. Model spend per invoice goes from 4 cents to 12 cents, which adds $19,200 a year. That is only 1.4% of the invoice workflow's savings, because a person costs $6 per invoice and the model costs cents.

The same jump hits other workflows much harder. The cheaper a task is for a person, the larger the share of its saving the price takes:

Time a person spends per task

Saving lost when model spend triples

15 seconds

41%

1 minute

9%

6 minutes

1%

Share of the saving lost when model spend per task triples, by how long a person takes per task. Fifteen seconds: 41 percent. One minute: 9 percent. Six minutes: 1 percent.

A price cliff breaks the business case for quick, high-volume work. For invoices and claims it is a rounding error (modelled).

A price cliff breaks the business case for quick, high-volume work like inbox triage. For invoices, claims and onboarding cases it is a rounding error.

Cutoff: retries cannot reopen a closed account

This has happened to us twice. About a month before I wrote this, a frontier lab revoked Beam's access across the platform, citing a policy violation. About a year earlier, a different major lab had done the same. A rate limit says slow down, and an outage says not right now. A cutoff says the account is gone, and no amount of retrying reopens it.

Finance, HR and operations agents touch customer records, employee data and payments, and a provider's safety system can read a normal pattern in your industry as a violation. I found no research on this, so treat it as field experience, not a statistic.

The cost of a cutoff is mostly the time until the workflow runs again. While it is down, 952 invoices a working day go back to people, at $5,390 a day more than the agent. If no second model was approved for invoice data and tested, getting that done under incident takes about 15 business days. That makes one cutoff about $99,000, migration included. With a fallback approved and tested before the incident, it is one day and about $9,000. Beam's two cutoffs came across all the labs we use, so for one workflow on one lab I used a 20% yearly chance: $19,851 a year against $1,758 (modelled). At 50%, the pinned figure is about $50,000.

A cutoff incident on a timeline. Without a tested fallback: 15 business days of invoices going back to people at $5,390 a day, plus migration, about $99,000. With a fallback approved and tested in advance: one day, about $9,000.

The cost of a cutoff is the time until the workflow runs again. A pre-approved second model turns 15 days into one (modelled).

Out-of-scope action: the rare one carries the cost

For an operator, alignment is a plain question: did the agent do the task you gave it, and only that task? The failure that worries me sits beside the accuracy score. The agent completes the task, then does other work it decided would help. That bank's agent bought two extra items because it judged they would be needed. Researchers at DeepMind showed the mechanism: an agent can pursue a goal you never set even when its instructions were correct (Shah et al., 2022).

The cost is in the tail. The invoice workflow takes 240,000 outbound actions a year, one payment or email per invoice. If 1 in 1,000 goes beyond the task, that is 240 actions, and 99.9% of actions are still in scope. Most are cheap: an hour of cleanup. If 1 in 100 of them is severe, such as an unrecoverable payment or a confidential note sent outside the company, that is 2.4 a year. At $50,000 each, those 2.4 actions are 89% of a $134,256 yearly loss (modelled).

Out-of-scope actions per year on 240,000 outbound actions. One in a thousand goes beyond the task: 240 actions. One in a hundred of those is severe: 2.4 a year. At $50,000 each the severe ones are 89 percent of the $134,256 loss.

99.9% of actions in scope, and 2.4 severe ones a year carry most of the cost (modelled).

The severe cost moves the result most: about $2,400 a year at one severe action in a million and $10,000 each, more than the whole business case at one in 10,000 and $250,000. You learn which case you are in when it happens, and that is the argument for a gate.

Which controls take out most of that cost?

Four controls take the pinned workflow from $188,027 to $53,504 a year, and the gate on the action does most of the work. The other three are planning you do before you need it. Running them costs $11,600 a year: a quarterly test-set run on the second model, a specialist refreshing test cases, and building the gate once.

  1. A gate on the action, not more quality on the draft. Before the agent pays or sends, a check compares the action with the request: same vendor, same amount, same quantity, a recipient on the allowed list. High-value actions wait for a person's consent. A 2025 study found tool-using agents recognise risk in over 98% of direct questions and still take the risky action most of the time, with the gap persisting in larger and reasoning models, while a separate risk checker on the proposed action cut risky executions by 55% (Tang et al., 2025). I assumed rules on the action catch 90%, which takes this line from $134,256 to $18,226, consent time included. At the 55% the study measured, it is $64,812.

  2. A test set you keep, and a second model that has passed it. This covers both retirement and cutoff. The test set turns a migration into a rerun. The second model only helps if it is already approved to see the data and the prompts have been run on it. Qualifying it after the key dies is the retirement problem again, run as an incident. Together these take $34,571 down to $4,478.

  3. A business case on list price and tokens per task. This is a budgeting change, not a technical one, so it removes the surprise rather than the cost. Stress-test the case at three times the launch spend per task and see if it survives.

  4. A repetitive core on a model you keep. Extraction, coding, matching and routing are work today's open models do well with sharp domain rules. Let a frontier model produce examples of each task done right, have your team check them, and fine-tune an open model on them. Research from Berkeley marks the limit: imitating a stronger model transfers the tasks you trained on, not general capability (Gudibande et al., 2023). That is enough, because those tasks are what you need to keep running. The full case, and what owning costs, is in Open Source vs Frontier Models for Enterprise AI Agents.

Shrinking the dependency on one provider: a gate on the action, a kept test set, a pre-approved second model, a business case at list price, and a repetitive core on a model you keep. The gate alone removes about 60 percent of the pinned exposure.

The action gate removes about 60% of the pinned exposure. The other three controls are planning done before the incident.

In the model, owning the core halves what is left of retirement and cutoff and cuts the price line to a third, assuming hosting costs about what the API did. Building and refreshing it costs about $12,800 a year, so on risk alone it roughly breaks even at this volume: $51,149 against $53,504. Do the first three controls first, and own the core when volume is higher, a cutoff would hurt more, or it also cuts your cost per task.

🧭 How we see it at Beam

We design every workflow on the assumption that the model under a step will change. The model is a setting on each step, with a fallback model if a call fails. Each step carries its own evaluation criteria in plain language, so a new model is scored against the bar the old one passed, and a migration becomes a rerun instead of a project. On the step that sends, pays or files, we set Consent required, so a person approves before anything leaves the company. That gate sits on the action and holds no matter which model drafted it.

Where should you start?

Start with the step that pays or sends, then the provider you could least afford to lose. That order follows the money in the model: the action gate alone removes about 60% of the pinned exposure.

How do you start small?

  1. Pick one live workflow and list each step, which model it runs on, and which steps take an action outside the company.

  2. Put a gate on the action step: rules that compare the action with the request, plus a person's consent above a value you set.

  3. Freeze 200 to 500 real cases with the right answers as a test set, and rerun it each quarter.

  4. Run that test set on a second model from a different provider, and get it approved for the same data now, not during an incident.

  5. Rebuild the business case on list price and on per-task token use, at three times the launch spend.

  6. Once those hold, look at which steps are repetitive enough to move to a model you keep.

What should you ask your team or vendor?

  • Which model does each step run on, and what notice do we get before it is retired?

  • If that provider closed our account tonight, which steps stop, and how many days until they run again?

  • Is a second model already approved for this data and tested on our cases?

  • Is the business case built on list price and on real tokens per task?

  • Which steps take an action outside the company, and what checks the action against the request?

  • Who pays for the retest when a model is retired?

What does this look like on an HR inbox?

Take an HR inbox triage agent: 30,000 emails a month, about one minute each by hand, four steps, and a reply that goes to a candidate or employee. It saves about $327,600 a year. Pinned to one model with no controls, the same four risks put about $248,000 a year at stake, 76% of the saving (modelled): $201,384 from out-of-scope replies, $28,800 from price, $10,880 from retirement and $6,620 from a cutoff.

Two things drive that. The price jump takes 9% of the saving, because a minute of a person's time is cheap. And a reply leaves the company 360,000 times a year. Picture the agent handling a thousand threads correctly, then forwarding a note about a lawsuit to a reporter. That is a hypothetical, but one message like it outweighs a year of small misses. With the four controls the exposure falls to about $71,000, and most of the drop comes from the gate on the outgoing reply.

How we use this at Beam

Your provider will keep changing the deal, and that is its right. Your side is to price what each change costs, gate what leaves the company, and keep a second place for the workflow to run. We build that way because we have been on the wrong end of all four events: a retired model under a live client, a discount that ended, and two closed accounts. The per-step model setting, the per-step criteria and Consent required are the product form of those lessons.

For choosing between frontier models, read The AI Model Landscape for Building Agents and How We Select the Right Frontier Model for Each Problem; for the fine-tuning method, Your Prompts Are Good, But Fine-Tuning Your LLMs Makes Them Better.

The hidden risk of building on one frontier model is not that it will be wrong. It is that one day it will be gone, repriced, or doing something you did not ask, and the gate, the test set and the second model are what decide whether that costs you a day or a quarter.

FAQ

What is frontier model risk for an AI agent?

The risk that comes from the provider's decisions rather than the model's accuracy: retiring the model you qualified, ending a launch price, closing your account, or shipping a model that acts beyond the task. None of these shows in an accuracy chart, and in our modelled invoice workflow they put about $188,000 a year at stake.

Which of the four risks costs the most?

The out-of-scope action, by far. In the model it is 71% of the exposure, because a handful of severe actions a year, such as an unrecoverable payment, outweigh thousands of small misses. Price, which everyone negotiates, is about 10%.

How do you protect an agent against a model being retired?

Keep a frozen test set of 200 to 500 real cases with the right answers, and run a second model from another provider against it before you need it. A retirement then becomes a rerun and a week of spot checks rather than a re-qualification project under deadline.

Does a gate on the action slow the agent down?

Only on the actions you choose. Rules that compare the action with the request run in milliseconds. Consent from a person applies above a value you set, so most actions pass straight through while the ones that cannot be undone wait for approval.

Figures as of 15 November 2026. Modelled values are labelled as such; vendor-reported figures are marked as vendor claims.

Start Today

Start building AI agents to automate processes

Join our platform and start building AI agents for various types of automations.

Start Today

Start building AI agents to automate processes

Join our platform and start building AI agents for various types of automations.