Everyday AI Tools Guide — Part 2: Building a Reliable Workflow

Everyday AI Tools Guide — Series Navigation

Part 1: Choosing the Right Tool
Part 2: Building a Reliable Workflow (You are here)
Part 3: Practical Prompts and Safe Automation

Professional using an AI tool on a laptop

Photo: Matheus Bertelli / Pexels. Used under the Pexels License.

Updated September 25, 2026 · Part 2 of our beginner-friendly AI tools series. Every limit, retry count, notice period and retention window below was read on that date from the vendor’s own documentation, or from the standards body or regulator named. No subscription or usage prices appear anywhere in this article.

Part 1 compared what these tools can hold. This part is about what they do when something goes wrong — which is the question that decides whether a workflow survives contact with a real week of work.

The number that stopped being published

Start with a finding that is itself the news. Of the three major providers, only one still publishes concrete per-tier throughput numbers. Anthropic’s documentation lists them by tier and model: on its lowest paid tier, 1,000 requests per minute, 2,000,000 input tokens per minute and 400,000 output tokens per minute for most current models, with smaller ceilings on specific ones.

OpenAI’s rate-limit page names the dimensions — requests per minute, requests per day, tokens per minute, tokens per day, images per minute — and then directs you to your own account settings for the figures. Google’s Gemini page names requests per minute, input tokens per minute and requests per day, and points to AI Studio. Neither publishes a per-tier table any more.

The practical consequence is small but real: you cannot size a job against a published number for two of three vendors, so any plan that depends on throughput has to begin with a look at your own dashboard on the day you build it. Anyone quoting you a Tier 1 figure for those two from memory is quoting something that is no longer on the page.

What a rejected request actually costs, in seconds

All three return HTTP 429 when you exceed a limit — Google phrases it as 429 RESOURCE_EXHAUSTED — and all three publish a header telling you when to try again. Anthropic’s documentation is blunt about it: the retry-after value is the number of seconds to wait, and earlier retries will fail.

Where they diverge is how many times their own software libraries retry before giving up, and that difference turns into wall-clock time on your desk.

Default automatic retries	Count	Added delay, 1s doubling	Added delay if the server says wait 60s
Anthropic official libraries	2	about 3 seconds	120 seconds
OpenAI official libraries	2	about 3 seconds	120 seconds
Google Python library	4	about 15 seconds	240 seconds

Converted to one ordinary job: 200 requests, of which 5 percent hit a rate limit carrying a 60-second retry-after, with two retries each. That is 1,200 seconds — 20 minutes of a workday spent entirely on waiting, from a failure rate most people would call negligible. The delay columns above are this article’s arithmetic on each vendor’s published retry count and its published backoff description, not figures the vendors print.

Two of the three explicitly recommend adding random jitter so that retries do not all land at the same instant; Anthropic’s pages describe exponential backoff and honoring the header but do not mention jitter. One more caveat worth knowing before you rely on retries at all: Anthropic notes that a 429 caused by a spending cap carries no retry-after header and will keep failing. Retrying is not a universal remedy.

The 200-document job, run two ways

This is the single conversion in this article that changes how a week actually goes. Suppose 200 documents need the same treatment — summarize, extract four fields, flag anything unusual.

Done interactively, at a generous two minutes each including paste, read and correction, that is 400 minutes: about 6.7 hours, a full working day, and a day during which the person cannot do anything else.

Done through the batch interface, the same 200 items are 0.2 percent of a single Anthropic batch (cap 100,000 requests) or 0.4 percent of an OpenAI batch (cap 50,000). You submit once and collect later.

Published batch limit	Anthropic	OpenAI	Google
Turnaround statement	most complete in under 1 hour; expire at 24 hours	complete within 24 hours, often quicker	target 24 hours, usually much quicker
Maximum requests per batch	100,000	50,000	not published
Maximum input file size	256 MB	200 MB	2 GB (inline requests under 20 MB)
How long results stay available	29 days from creation	30 days after completion	6 weeks

The same 24 hours, meaning two different things

All three publish the number 24, and it does not mean the same thing in all three places. For OpenAI and Google it is a completion target — the job is expected to finish inside it. For Anthropic it is an expiry: a batch that has not completed at 24 hours expires. Same figure on the page, opposite consequence for an unattended overnight job. If your process assumes results will simply be there in the morning, that distinction is the one that bites.

The retention spread matters for a different reason. Results live 29 days, 30 days or 6 weeks depending on the vendor — a 13-day difference between the shortest and longest for the same nominal 24-hour job. A monthly reporting cycle that collects results on day 30 is inside one window and outside another.

Guaranteed structure, and the cases the guarantee excludes

Structured output — asking for results as data rather than prose — is what lets a workflow run without a person reading every line. The promises are not equivalent.

OpenAI states that structured outputs ensure responses adhere to a supplied schema, and contrasts that with plain JSON mode: both produce valid JSON, only structured outputs ensure schema adherence. Anthropic makes the same guarantee and then publishes its own exceptions, which is more useful: schema compliance holds “in most cases,” and the documented exceptions are a refusal, hitting the output token limit mid-structure, and enum values coming back with different capitalization. Google promises the least and says so, advising that output is syntactically correct JSON but that values should still be validated in your application.

The published numeric limits fall in an odd place. Anthropic publishes 20 strict tools per request, 24 optional parameters across all schemas, 16 union-typed parameters, a 180-second schema compilation timeout and a 24-hour grammar cache. OpenAI’s current page publishes no numeric schema limit at all — but Microsoft, hosting the same models on Azure, publishes a maximum of 100 object properties and a maximum nesting depth of 5, along with a list of unsupported schema keywords. Those figures belong to Microsoft’s documentation, not OpenAI’s, and a schema built against one page can fail on the other.

There is even a direct contradiction between vendors on individual keywords: Google lists minItems and maxItems among supported schema keywords; Microsoft lists them as unsupported in strict mode. Nobody is lying. The implementations differ, and the only safe assumption is that a schema is not portable.

The two features a careful workflow wants together, and cannot have

Citations are the other half of reliability: a pointer back to the source sentence so a human can check a claim in seconds rather than rereading a document. Anthropic’s citation feature publishes an unusual guarantee — citations are guaranteed to contain valid pointers to the provided documents, because the quoted text is extracted rather than written by the model. Page locations are 1-indexed for PDFs; character locations are 0-indexed.

And there is a trap: on Anthropic, citations and structured outputs are incompatible, returning a 400 error if both are requested on a user document. Verifiable quotes and guaranteed data shape are the two things a reviewable workflow most wants, and they currently do not combine on that platform. OpenAI’s equivalent comes with a contractual condition instead — when web results are displayed to end users, inline citations must be clearly visible and clickable. That is a design requirement, not a suggestion.

Tip: Batch processing only pays off if the inputs arrive as clean files rather than phone photographs — a sheet-fed scanner turns a stack of paper into consistent, text-layer PDFs in one pass. Browsing sheet-fed document scanners shows the category; match the daily page capacity to your actual volume. (These are Amazon Associate links — we may earn a small commission on qualifying purchases.)

Nobody publishes an error rate. A regulator published one.

Searching all three vendors’ documentation, help centres and usage policies for a numeric accuracy, hallucination or error rate turns up nothing. Not a figure with a caveat — no figure at all. What they publish instead is qualitative: Anthropic’s support pages say users should not rely on Claude as a singular source of truth and should scrutinize high-stakes advice. Google’s says its assistants can hallucinate and present inaccurate information as factual. OpenAI’s usage policy carries no verification disclaimer but does prohibit automating high-stakes decisions without human review, naming medical, legal, financial, employment, housing and government services among the sensitive areas.

The only hard accuracy number in this whole subject comes from the Federal Trade Commission. In an order announced April 28, 2025 and finalized in August 2025, the FTC found that a company advertising an AI content detector as 98 percent accurate achieved 53 percent on general-purpose content. Claimed error rate 2 percent; found error rate 47 percent — a factor of roughly 24. The order requires the company to hold competent and reliable evidence for effectiveness claims.

That contrast is the argument of this article in one line. The vendors publish no accuracy number, so any specific one you are quoted came from a marketing claim, and marketing claims in this field have already been fined.

Who can read your drafts, by tier and by day count

Reliability includes knowing what happens to the text after you send it. The published defaults differ most sharply between paid and unpaid tiers of the same product.

Anthropic, business and API: retained data is never used for model training without express permission; inputs and outputs are deleted from the backend within 30 days of receipt or generation. Flagged content in abuse monitoring is held up to 2 years, and trust-and-safety classification scores up to 7 years. Consumer tiers train only if the user opts in.
OpenAI, API: data sent to the API has not been used to train models since March 1, 2023 absent an opt-in. Abuse-monitoring retention is 30 days. Zero-retention is available but requires prior approval, and several conversational endpoints are explicitly not eligible — those are retained until deleted.
Google, unpaid tier: the terms state that human reviewers may read, annotate and process API input and output, and that content may be used to improve products including machine learning technologies.
Google, paid tier: prompts and responses are not used to improve products. Logs are kept for a limited period to detect violations; no day count is published for that window. Search-grounded prompts and outputs are stored 30 days on both tiers.

For an office deciding which tier to buy, that free-versus-paid line on one vendor is the single largest difference in the list, and it is a difference in who reads the text rather than in what the tool can do.

Sixty days, one hundred eighty days, or no answer

The quietest way a working process breaks is that the model underneath it is retired. The published notice periods are three very different commitments.

Vendor and model class	Published minimum notice	Business days that buys you
OpenAI, generally available models	at least 6 months	about 128
OpenAI, specialized variants	at least 3 months	about 64
OpenAI, preview models	potentially 2 weeks or less	about 10
Anthropic, publicly released models	at least 60 days	about 42
Google, all classes	no published minimum	unknown in advance

Google’s page says the listed shutdown dates are the earliest possible dates and that the exact date will be communicated with advance notice, without defining the notice. Its enterprise documentation adds one concrete mechanic: one month before retirement, new access is blocked for inference and tuning, and a retirement date is announced once the next stable version ships — notice tied to a successor shipping rather than to a fixed window. The business-day column is this article’s conversion at five days per week; the notice periods are the vendors’.

Anthropic pairs its shorter notice with something the others do not publish: a version header that pins API behaviour, with a stated commitment that within a version, existing input and output parameters are preserved, though new optional inputs and new enum values may be added. Shorter notice, more explicit stability contract. Neither arrangement is strictly better; they trade off differently, and a workflow you cannot re-test in 42 business days should not be sitting on the 60-day commitment.

The standard that tells you to check the citations

If you want a neutral authority rather than a vendor, the National Institute of Standards and Technology published the AI Risk Management Framework, AI 100-1, in January 2023, organized around four functions: GOVERN, MAP, MEASURE and MANAGE. Its Generative AI Profile, AI 600-1, followed in July 2024 and supplies a government term for the failure mode — confabulation, defined as generating and confidently presenting erroneous content.

The most directly usable line for an office workflow is action MS-2.5-003 in the MEASURE function: review and verify sources and citations in generative system outputs during pre-deployment risk measurement. A related action, MP-2.3-001, asks that output accuracy be assessed by comparing it to ground truth data. Neither is binding on a small business, and both are exactly what a sensible reviewer would do anyway — which is the point of citing them: the checking step is not excessive caution, it is the published standard.

A workflow that survives its own vendors

Weighing only what is documented, four decisions follow, and none of them require technical skill.

Pin the model version and write the date next to it. On a 60-day minimum, you have about 42 business days from notice to re-test. Put a recurring reminder at 45 days, not at 60.
Move anything over roughly 50 items to the batch route, and read your vendor’s 24-hour sentence carefully — completion target or expiry deadline. The 200-document example above is a working day recovered either way.
Ask for citations on anything a person will act on, and structured data on anything a spreadsheet will ingest — and confirm first whether your platform allows both at once, because at least one currently does not.
Decide the tier question on who reads the text, not on capability. One vendor’s unpaid terms permit human review of inputs and outputs; its paid terms do not. That is a policy difference, not a feature difference, and no amount of prompt engineering changes it.

The cost of all this is real and worth stating: verification steps add time to every run, and the checking usually cannot be automated away, since the thing being checked is whether the automation was right. What you buy with that time is the ability to answer “where did this number come from” without redoing the work. On the published record, no vendor will answer that question for you, and one regulator has already shown what happens to a company that claims otherwise.

Every documentation page this compares

Reliable AI workflow with prompts, review and privacy checks

Search This Blog

TechDailyBrief